Datasets:

jamescalam

/

llama-2-arxiv-papers-chunked

like 17

2103.15691

9

then rasterises them into 1D tokens zi2Rd. The sequence of tokens input to the following transformer encoder is z= [zcls;Ex1;Ex2;:::;ExN] +p; (1) where the projection by Eis equivalent to a 2D convolution. As shown in Fig. 1, an optional learned classification token zclsis prepended to this sequence, and its representation at the final layer of the encoder serves as the final representation used by the classification layer [17]. In addition, a learned positional embedding, p2RNd, is added to the tokens to retain positional information, as the subsequent self-attention operations in the transformer are permutation invariant. The tokens are then passed through an encoder consisting of a sequence of Ltransformer layers. Each layer `comprises of Multi-Headed Self-Attention [68], layer normalisation (LN) [2], and MLP blocks as follows: y`=MSA (LN(z`)) +z`(2) z`+1=MLP(LN(y`)) +y`: (3) The MLP consists of two linear projections separated by a

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

10

z`+1=MLP(LN(y`)) +y`: (3) The MLP consists of two linear projections separated by a GELU non-linearity [28] and the token-dimensionality, d, remains fixed throughout all layers. Finally, a linear classifier is used to classify the encoded input based on zL cls2Rd, if it was prepended to the input, or a global average pooling of all the tokens, zL, otherwise. As the transformer [68], which forms the basis of ViT [18], is a flexible architecture that can operate on any sequence of input tokens z2RNd, we describe strategies for tokenising videos next. 3.2. Embedding video clips We consider two simple methods for mapping a video V2RTHWCto a sequence of tokens ~ z2 Rntnhnwd. We then add the positional embedding and reshape into RNdto obtain z, the input to the transformer. Uniform frame sampling As illustrated in Fig. 2, a straightforward method of tokenising the input video is to uniformly sample ntframes from the input video clip, embed each 2D frame independently using the same method

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

11

straightforward method of tokenising the input video is to uniformly sample ntframes from the input video clip, embed each 2D frame independently using the same method as ViT [18], and concatenate all these tokens together. Concretely, ifnh
nwnon-overlapping image patches are extracted from each frame, as in [18], then a total of nt
nh
nw tokens will be forwarded through the transformer encoder. Intuitively, this process may be seen as simply constructing a large 2D image to be tokenised following ViT. We note that this is the input embedding method employed by the concurrent work of [4]. #!"Figure 2: Uniform frame sampling: We simply sample ntframes, and embed each 2D frame independently following ViT [18]. !" # Figure 3: Tubelet embedding. We extract and linearly embed nonoverlapping tubelets that span the spatio-temporal input volume. Tubelet embedding An alternate method, as shown in Fig. 3, is to extract non-overlapping, spatio-temporal “tubes” from the input volume, and to linearly project this to Rd. This method is an extension of ViT’s embedding to 3D,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

12

“tubes” from the input volume, and to linearly project this to Rd. This method is an extension of ViT’s embedding to 3D, and corresponds to a 3D convolution. For a tubelet of dimensionthw,nt=bT tc,nh=bH hcandnw=bW wc, tokens are extracted from the temporal, height, and width dimensions respectively. Smaller tubelet dimensions thus result in more tokens which increases the computation. Intuitively, this method fuses spatio-temporal information during tokenisation, in contrast to “Uniform frame sampling” where temporal information from different frames is fused by the transformer. 3.3. Transformer Models for Video As illustrated in Fig. 1, we propose multiple transformerbased architectures. We begin with a straightforward extension of ViT [18] that models pairwise interactions between all spatio-temporal tokens, and then develop more efficient variants which factorise the spatial and temporal dimensions of the input video at various levels of the transformer architecture. Model 1: Spatio-temporal attention This model simply forwards all spatio-temporal tokens extracted from the video, z0, through the transformer encoder. We note that this has also been explored concurrently by [4] in their

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

13

video, z0, through the transformer encoder. We note that this has also been explored concurrently by [4] in their “Joint Space-Time” model. In contrast to CNN architectures, where the receptive field grows linearly with the number of layers, each transformer layer models all pair…1CLSNPositional + Token EmbeddingTemporal + Token Embedding Embed to tokens…1N2…1N…TTemporal Transformer EncoderMLP Head Class…CLS100CLS0CLS0Spatial Transformer EncoderSpatial Transformer EncoderSpatial Transformer EncoderFigure 4: Factorised encoder (Model 2). This model consists of two transformer encoders in series: the first models interactions between tokens extracted from the same temporal index to produce a latent representation per time-index. The second transformer models interactions between time steps. It thus corresponds to a “late fusion” of spatial- and temporal information. wise interactions between all spatio-temporal tokens, and it thus models long-range interactions across the video from the first layer. However, as it models all pairwise interactions, Multi-Headed Self Attention (MSA) [68] has quadratic complexity with respect to the number of tokens.

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

14

quadratic complexity with respect to the number of tokens. This complexity is pertinent for video, as the number of tokens increases linearly with the number of input frames, and motivates the development of more efficient architectures next. Model 2: Factorised encoder As shown in Fig. 4, this model consists of two separate transformer encoders. The first, spatial encoder, only models interactions between tokens extracted from the same temporal index. A representation for each temporal index, hi2Rd, is obtained after Ls layers: This is the encoded classification token, zLs clsif it was prepended to the input (Eq. 1), or a global average pooling from the tokens output by the spatial encoder, zLs, otherwise. The frame-level representations, hi, are concatenated intoH2Rntd, and then forwarded through a temporal encoder consisting of Lttransformer layers to model interactions between tokens from different temporal indices. The output token of this encoder is then finally classified. This architecture corresponds to a “late fusion” [34,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

15

The output token of this encoder is then finally classified. This architecture corresponds to a “late fusion” [34, 56, 72, 46] of temporal information, and the initial spatial encoder is identical to the one used for image classification. It is thus analogous to CNN architectures such as [24, 34, 72, 86] which first extract per-frame features, and then aggregate them into a final representation before classifying them. Although this model has more transformer layers than Model 1 (and thus more parameters), it requires fewer floating point operations (FLOPs), as the two separate transformer blocks have a complexity ofO((nh
nw)2+n2 t)compared toO((nt
nh
nw)2)of Model 1. Transformer Block xLK V Q MLPLayer Norm Layer NormMulti-HeadAttention K V Q Layer NormMulti-HeadAttention Temporal Self-Attention BlockSpatial Self-Attention BlockToken embeddingPositional embeddingFigure 5: Factorised self-attention (Model 3). Within each transformer block, the multi-headed self-attention operation is factorised into two operations (indicated by striped boxes) that first

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

16

only compute self-attention spatially, and then temporally. Model 3: Factorised self-attention This model, in contrast, contains the same number of transformer layers as Model 1. However, instead of computing multi-headed self-attention across all pairs of tokens, z`, at layerl, we factorise the operation to first only compute self-attention spatially (among all tokens extracted from the same temporal index), and then temporally (among all tokens extracted from the same spatial index) as shown in Fig. 5. Each self-attention block in the transformer thus models spatio-temporal interactions, but does so more efficiently than Model 1 by factorising the operation over two smaller sets of elements, thus achieving the same computational complexity as Model 2. We note that factorising attention over input dimensions has also been explored in [29, 78], and concurrently in the context of video by [4] in their “Divided Space-Time” model. This operation can be performed efficiently by reshaping the tokens zfromR1nt
nh
nw
dtoRntnh
nw
d(denoted byzs) to compute spatial self-attention. Similarly, the input

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

17

byzs) to compute spatial self-attention. Similarly, the input to temporal self-attention, ztis reshaped to Rnh
nwnt
d. Here we assume the leading dimension is the “batch dimension”. Our factorised self-attention is defined as y` s= MSA(LN( z` s)) +z` s (4) y` t= MSA(LN( y` s)) +y` s (5) z`+1= MLP(LN( y` t)) +y` t: (6) We observed that the order of spatial-then-temporal selfattention or temporal-then-spatial self-attention does not make a difference, provided that the model parameters are initialised as described in Sec. 3.4. Note that the number of parameters, however, increases compared to Model 1, as there is an additional self-attention layer (cf. Eq. 7). We do not use a classification token in this model, to avoid ambiguities when reshaping the input tokens between spatial and temporal dimensions. Model 4: Factorised dot-product attention Finally, we

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

18

temporal dimensions. Model 4: Factorised dot-product attention Finally, we develop a model which has the same computational complexity as Models 2 and 3, while retaining the same number of parameters as the unfactorised Model 1. The factorisation of spatial- and temporal dimensions is similar in spirit K V Q Self-Attention BlockLayer NormMulti-HeadDot-product AttentionConcatenateLinear K V Q Scaled Dot-Product AttentionLinearLinearLinearSpatial HeadsK V Q Scaled Dot-Product AttentionLinearLinearLinearTemporal HeadsFigure 6: Factorised dot-product attention (Model 4). For half of the heads, we compute dot-product attention over only the spatial axes, and for the other half, over only the temporal axis. to Model 3, but we factorise the multi-head dot-product attention operation instead (Fig. 6). Concretely, we compute attention weights for each token separately over the spatialand temporal-dimensions using different heads. First, we note that the attention operation for each head is defined as Attention( Q;K;V) = SoftmaxQK> pdk V: (7) In self-attention, the queries Q=XW q, keys K=XW k,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

19

pdk V: (7) In self-attention, the queries Q=XW q, keys K=XW k, and values V=XW vare linear projections of the input X withX;Q;K;V2RNd. Note that in the unfactorised case (Model 1), the spatial and temporal dimensions are merged asN=nt
nh
nw. The main idea here is to modify the keys and values for each query to only attend over tokens from the same spatialand temporal index by constructing Ks;Vs2Rnh
nwd andKt;Vt2Rntd, namely the keys and values corresponding to these dimensions. Then, for half of the attention heads, we attend over tokens from the spatial dimension by computing Ys= Attention( Q;Ks;Vs), and for the rest we attend over the temporal dimension by computingYt= Attention( Q;Kt;Vt). Given that we are only changing the attention neighbourhood for each query, the attention operation has the same dimension as in the unfactorised case, namely Ys;Yt2RNd. We then combine the outputs of multiple heads by concatenating them and using a linear projection [68], Y= Concat( Ys;Yt)WO.

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

20

the outputs of multiple heads by concatenating them and using a linear projection [68], Y= Concat( Ys;Yt)WO. 3.4. Initialisation by leveraging pretrained models ViT [18] has been shown to only be effective when trained on large-scale datasets, as transformers lack some of the inductive biases of convolutional networks [18]. However, even the largest video datasets such as Kinetics [35], have several orders of magnitude less labelled examples when compared to their image counterparts [16, 39, 58]. As a result, training large models from scratch to high accuracy is extremely challenging. To sidestep this issue, and enable more efficient training we initialise our video models from pretrained image models. However, this raises several practical questions, specifically on how to initialise parameters not present or incompatible with image models. We nowdiscuss several effective strategies to initialise these largescale video classification models. Positional embeddings A positional embedding pis added to each input token (Eq. 1). However, our video models have nttimes more tokens than the pretrained image model. As a result, we initialise the positional embeddings by “repeating” them temporally from Rnw
nhdto

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

21

Rnt
nh
nwd. Therefore, at initialisation, all tokens with the same spatial index have the same embedding which is then fine-tuned. Embedding weights, EWhen using the “tubelet embedding” tokenisation method (Sec. 3.2), the embedding filter Eis a 3D tensor, compared to the 2D tensor in the pretrained model, Eimage. A common approach for initialising 3D convolutional filters from 2D filters for video classification is to “inflate” them by replicating the filters along the temporal dimension and averaging them [8, 22] as E=1 t[Eimage;:::;Eimage;:::;Eimage]: (8) We consider an additional strategy, which we denote as “central frame initialisation”, where Eis initialised with zeroes along all temporal positions, except at the centre bt 2c, E= [0;:::;Eimage;:::;0]: (9) Therefore, the 3D convolutional filter effectively behaves like “Uniform frame sampling” (Sec. 3.2) at initialisation,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

22

Therefore, the 3D convolutional filter effectively behaves like “Uniform frame sampling” (Sec. 3.2) at initialisation, while also enabling the model to learn to aggregate temporal information from multiple frames as training progresses. Transformer weights for Model 3 The transformer block in Model 3 (Fig. 5) differs from the pretrained ViT model [18], in that it contains two multi-headed self attention (MSA) modules. In this case, we initialise the spatial MSA module from the pretrained module, and initialise all weights of the temporal MSA with zeroes, such that Eq. 5 behaves as a residual connection [27] at initialisation. 4. Empirical evaluation We first present our experimental setup and implementation details in Sec. 4.1, before ablating various components of our model in Sec. 4.2. We then present state-of-the-art results on five datasets in Sec. 4.3. 4.1. Experimental Setup Network architecture and training Our backbone architecture follows that of ViT [18] and BERT [17]. We consider ViT-Base (ViT-B, L=12,NH=12,d=768), ViT-Large

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

23

(ViT-L,L=24,NH=16,d=1024 ), and ViT-Huge (ViT-H, L=32,NH=16,d=1280 ), whereLis the number of transformer layers, each with a self-attention block of NHheads Table 1: Comparison of input encoding methods using ViViT-B and spatio-temporal attention on Kinetics. Further details in text. Top-1 accuracy Uniform frame sampling 78.5 Tubelet embedding Random initialisation [25] 73.2 Filter inflation [8] 77.6 Central frame 79.2 and hidden dimension d. We also apply the same naming scheme to our models (e.g., ViViT-B/16x2 denotes a ViTBase backbone with a tubelet size of hwt= 16162). In all experiments, the tubelet height and width are equal. Note that smaller tubelet sizes correspond to more tokens at the input, and thus more computation. We train our models using synchronous SGD and momentum, a cosine learning rate schedule and TPU-v3 accelerators. We initialise our models from a ViT image

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

24

We train our models using synchronous SGD and momentum, a cosine learning rate schedule and TPU-v3 accelerators. We initialise our models from a ViT image model trained either on ImageNet-21K [16] (unless otherwise specified) or the larger JFT [58] dataset. We implement our method using the Scenic library [15] and have released our code and models. Datasets We evaluate the performance of our proposed models on a diverse set of video classification datasets: Kinetics [35] consists of 10-second videos sampled at 25fps from YouTube. We evaluate on both Kinetics 400 and 600, containing 400 and 600 classes respectively. As these are dynamic datasets (videos may be removed from YouTube), we note our dataset sizes are approximately 267 000 and 446 000 respectively. Epic Kitchens-100 consists of egocentric videos capturing daily kitchen activities spanning 100 hours and 90 000 clips [13]. We report results following the standard “action recognition” protocol. Here, each video is labelled with a “verb” and a “noun” and we therefore predict both categories using a single network with two “heads”. The topscoring verb and action pair predicted by the network form an “action”, and action accuracy is the primary metric.

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

25

an “action”, and action accuracy is the primary metric. Moments in Time [45] consists of 800 000, 3-second YouTube clips that capture the gist of a dynamic scene involving animals, objects, people, or natural phenomena. Something-Something v2 (SSv2) [26] contains 220 000 videos, with durations ranging from 2 to 6 seconds. In contrast to the other datasets, the objects and backgrounds in the videos are consistent across different action classes, and this dataset thus places more emphasis on a model’s ability to recognise fine-grained motion cues. Inference The input to our network is a video clip of 32 frames using a stride of 2, unless otherwise mentioned, similar to [21, 20]. Following common practice, at inference time, we process multiple views of a longer video and aver-Table 2: Comparison of model architectures using ViViT-B as the backbone, and tubelet size of 162. We report Top-1 accuracy on Kinetics 400 (K400) and action accuracy on Epic Kitchens (EK). Runtime is during inference on a TPU-v3. K400 EKFLOPs (109)Params (106)Runtime (ms)

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

26

Runtime is during inference on a TPU-v3. K400 EKFLOPs (109)Params (106)Runtime (ms) Model 1: Spatio-temporal 80.0 43.1 455.2 88.9 58.9 Model 2: Fact. encoder 78.8 43.7 284.4 115.1 17.4 Model 3: Fact. self-attention 77.4 39.1 372.3 117.3 31.7 Model 4: Fact. dot product 76.3 39.5 277.1 88.9 22.9 Model 2: Ave. pool baseline 75.8 38.8 283.9 86.7 17.3 Table 3: The effect of varying the number of temporal transformers,Lt, in the Factorised encoder model (Model 2). We report the Top-1 accuracy on Kinetics 400. Note that Lt= 0corresponds to the “average pooling baseline”. Lt 0 1 4 8 12 Top-1 75.8 78.6 78.8 78.8 78.9 age per-view logits to obtain the final result. Unless otherwise specified, we use a total of 4 views per video (as this

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

27

age per-view logits to obtain the final result. Unless otherwise specified, we use a total of 4 views per video (as this is sufficient to “see” the entire video clip across the various datasets), and ablate these and other design choices next. 4.2. Ablation study Input encoding We first consider the effect of different input encoding methods (Sec. 3.2) using our unfactorised model (Model 1) and ViViT-B on Kinetics 400. As we pass 32-frame inputs to the network, sampling 8 frames and extracting tubelets of length t= 4 correspond to the same number of tokens in both cases. Table 1 shows that tubelet embedding initialised using the “central frame” method (Eq. 9) performs well, outperforming the commonly-used “filter inflation” initialisation method [8, 22] by 1.6%, and “uniform frame sampling” by 0.7%. We therefore use this encoding method for all subsequent experiments. Model variants We compare our proposed model variants (Sec. 3.3) across the Kinetics 400 and Epic Kitchens

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

28

encoding method for all subsequent experiments. Model variants We compare our proposed model variants (Sec. 3.3) across the Kinetics 400 and Epic Kitchens datasets, both in terms of accuracy and efficiency, in Tab. 2. In all cases, we use the “Base” backbone and tubelet size of 162. Model 2 (“Factorised Encoder”) has an additional hyperparameter, the number of temporal transformers, Lt. We setLt= 4 for all experiments and show in Tab. 3 that the model is not sensitive to this choice. The unfactorised model (Model 1) performs the best on Kinetics 400. However, it can also overfit on smaller datasets such as Epic Kitchens, where we find our “Factorised Encoder” (Model 2) to perform the best. We also consider an additional baseline (last row), based on Model 2, where we do not use any temporal transformer, and simply average pool the frame-level representations from the spatial encoder before classifying. This average pooling baseline performs the worst, and has a larger accuracy drop Table 4: The effect of progressively adding regularisation (each row includes all methods above it) on Top-1 action accuracy on

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

29

baseline performs the worst, and has a larger accuracy drop Table 4: The effect of progressively adding regularisation (each row includes all methods above it) on Top-1 action accuracy on Epic Kitchens. We use a Factorised encoder model with tubelet size162. Top-1 accuracy Random crop, flip, colour jitter 38.4 + Kinetics 400 initialisation 39.6 + Stochastic depth [31] 40.2 + Random augment [12] 41.1 + Label smoothing [61] 43.1 + Mixup [82] 43.7 16x8 16x4 16x2 Input tubelet size787980Top-1 Accuracy 16x8 16x4 16x2 Input tubelet size0.51.01.5TFLOPs ViViT-B ViViT-L (a) Accuracy (b) Compute Figure 7: The effect of the backbone architecture on (a) accuracy and (b) computation on Kinetics 400, for the spatio-temporal attention model (Model 1). 16x8 16x4 16x2 Input tubelet size72.575.077.580.0Top-1 Accuracy 16x8 16x4 16x2 Input tubelet size0.20.4TFLOPs

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

30

16x8 16x4 16x2 Input tubelet size0.20.4TFLOPs Spatio-temporal Factorised encoder Factorised self-attention Factorised dot-product (a) Accuracy (b) Compute Figure 8: The effect of varying the number of temporal tokens on (a) accuracy and (b) computation on Kinetics 400, for different variants of our model with a ViViT-B backbone. on Epic Kitchens, suggesting that this dataset requires more detailed modelling of temporal relations. As described in Sec. 3.3, all factorised variants of our model use significantly fewer FLOPs than the unfactorised Model 1, as the attention is computed separately over spatial- and temporal-dimensions. Model 4 adds no additional parameters to the unfactorised Model 1, and uses the least compute. The temporal transformer encoder in Model 2 operates on only nttokens, which is why there is a barely a change in compute and runtime over the average pooling baseline, even though it improves the accuracy substantially (3% on Kinetics and 4.9% on Epic Kitchens). Finally, Model 3 requires more compute and parameters than the other factorised models, as its additional self-attention

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

31

the other factorised models, as its additional self-attention block means that it performs another query-, key-, valueand output-projection in each transformer layer [68]. Model regularisation Pure-transformer architectures such as ViT [18] are known to require large training datasets, and we observed overfitting on smaller datasets like Epic Kitchens and SSv2, even when using an ImageNet pretrained model. In order to effectively train our modelsTable 5: The effect of spatial resolution on the performance of ViViT-L/16x2 and spatio-temporal attention on Kinetics 400. Crop size 224 288 320 Accuracy 80.3 80.7 81.0 GFLOPs 1446 2919 3992 Runtime 58.9 147.6 238.8 on such datasets, we employed several regularisation strategies that we ablate using our “Factorised encoder” model in Tab. 4. We note that these regularisers were originally proposed for training CNNs, and that [64] have recently explored them for training ViT for image classification. Each row of Tab. 4 includes all the methods from the rows above it, and we observe progressive improvements

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

32

explored them for training ViT for image classification. Each row of Tab. 4 includes all the methods from the rows above it, and we observe progressive improvements from adding each regulariser. Overall, we obtain a substantial overall improvement of 5.3% on Epic Kitchens. We also achieve a similar improvement of 5% on SSv2 by using all the regularisation in Tab. 4. Note that the Kineticspretrained models that we initialise from are from Tab. 2, and that all Epic Kitchens models in Tab. 2 were trained with all the regularisers in Tab. 4. For larger datasets like Kinetics and Moments in Time, we do not use these additional regularisers (we use only the first row of Tab. 4), as we obtain state-of-the-art results without them. The appendix contains hyperparameter values and additional details for all regularisers. Varying the backbone Figure 7 compares the ViViTB and ViViT-L backbones for the unfactorised spatiotemporal model. We observe consistent improvements in accuracy as the backbone capacity increases. As expected, the compute also grows as a function of the backbone size. Varying the number of tokens We first analyse the performance as a function of the number of tokens along the

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

33

the compute also grows as a function of the backbone size. Varying the number of tokens We first analyse the performance as a function of the number of tokens along the temporal dimension in Fig. 8. We observe that using smaller input tubelet sizes (and therefore more tokens) leads to consistent accuracy improvements across all of our model architectures. At the same time, computation in terms of FLOPs increases accordingly, and the unfactorised model (Model 1) is impacted the most. We then vary the number of tokens fed into the model by increasing the spatial crop-size from the default of 224 to 320 in Tab. 5. As expected, there is a consistent increase in both accuracy and computation. We note that when comparing to prior work we consistently obtain state-of-the-art results (Sec. 4.3) using a spatial resolution of 224, but we also highlight that further improvements can be obtained at higher spatial resolutions. Varying the number of input frames In our experiments so far, we have kept the number of input frames fixed at 32. We now increase the number of frames input to the model, thereby increasing the number of tokens proportionally.

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

34

We now increase the number of frames input to the model, thereby increasing the number of tokens proportionally. Table 6: Comparisons to state-of-the-art across multiple datasets. For “views”, xydenotes xtemporal crops and yspatial crops. We report the TFLOPs to process all spatio-temporal views. “FE” denotes our Factorised Encoder model. (a) Kinetics 400 Method Top 1 Top 5 Views TFLOPs blVNet [19] 73.5 91.2 – – STM [33] 73.7 91.6 – – TEA [42] 76.1 92.5 103 2.10 TSM-ResNeXt-101 [43] 76.3 – – – I3D NL [75] 77.7 93.3 103 10.77 CorrNet-101 [70] 79.2 – 103 6.72 ip-CSN-152 [66] 79.2 93.8 103 3.27 LGD-3D R101 [51] 79.4 94.4 – – SlowFast R101-NL [21] 79.8 93.9 103 7.02

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

35

LGD-3D R101 [51] 79.4 94.4 – – SlowFast R101-NL [21] 79.8 93.9 103 7.02 X3D-XXL [20] 80.4 94.6 103 5.82 TimeSformer-L [4] 80.7 94.7 13 7.14 ViViT-L/16x2 FE 80.6 92.7 11 3.98 ViViT-L/16x2 FE 81.7 93.8 13 11.94 Methods with large-scale pretraining ip-CSN-152 [66] (IG [44]) 82.5 95.3 103 3.27 ViViT-L/16x2 FE (JFT) 83.5 94.3 13 11.94 ViViT-H/14x2 (JFT) 84.9 95.8 43 47.77(b) Kinetics 600 Method Top 1 Top 5 AttentionNAS [76] 79.8 94.4 LGD-3D R101 [51] 81.5 95.6 SlowFast R101-NL [21] 81.8 95.1 X3D-XL [20] 81.9 95.5

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

36

SlowFast R101-NL [21] 81.8 95.1 X3D-XL [20] 81.9 95.5 TimeSformer-L [4] 82.2 95.6 ViViT-L/16x2 FE 82.9 94.6 ViViT-L/16x2 FE (JFT) 84.3 94.9 ViViT-H/14x2 (JFT) 85.8 96.5 (c) Moments in Time Top 1 Top 5 TSN [72] 25.3 50.1 TRN [86] 28.3 53.4 I3D [8] 29.5 56.1 blVNet [19] 31.4 59.3 AssembleNet-101 [54] 34.3 62.7 ViViT-L/16x2 FE 38.5 64.1(d) Epic Kitchens 100 Top 1 accuracy Method Action Verb Noun TSN [72] 33.2 60.2 46.0 TRN [86] 35.3 65.9 45.4 TBN [36] 36.7 66.0 47.2 TSM [43] 38.3 67.9 49.0 SlowFast [21] 38.5 65.6 50.0

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

37

TSM [43] 38.3 67.9 49.0 SlowFast [21] 38.5 65.6 50.0 ViViT-L/16x2 FE 44.0 66.4 56.8 (e) Something-Something v2 Method Top 1 Top 5 TRN [86] 48.8 77.6 SlowFast [20, 80] 61.7 – TimeSformer-HR [4] 62.5 – TSM [43] 63.4 88.5 STM [33] 64.2 89.8 TEA [42] 65.1 – blVNet [19] 65.2 90.3 ViVIT-L/16x2 FE 65.9 89.9 1 2 3 4 5 6 7 Number of views767880Top-1 Accuracy 32 stride 2 64 stride 2 128 stride 2 Figure 9: The effect of varying the number of frames input to the network and increasing the number of tokens proportionally. We use ViViT-L/16x2 Factorised Encoder on Kinetics 400. A Kinetics video contains 250 frames (10 seconds sampled at 25 fps) and the accuracy for each model saturates once the number of equidistant

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

38

video contains 250 frames (10 seconds sampled at 25 fps) and the accuracy for each model saturates once the number of equidistant temporal views is sufficient to “see” the whole video clip. Observe how models processing more frames (and thus more tokens) achieve higher single- and multi-view accuracy. Figure 9 shows that as we increase the number of frames input to the network, the accuracy from processing a single view increases, since the network incorporates longer temporal context. However, common practice on datasets such as Kinetics [21, 75, 42] is to average results over multiple, shorter “views” of the same video clip. Figure 9 also shows that the accuracy saturates once the number of views is sufficient to cover the whole video. As a Kinetics video consists of 250 frames, and we sample frames with a stride of 2, our model which processes 128 frames requires just a single view to “see” the whole video and achieve its maximum accuarcy. Note that we used ViViT-L/16x2 Factorised Encoder (Model 2) here. As this model is more efficient it can process more tokens, compared to the unfactorised Model 1 which runs out of memory after 48 frames using tubelet

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

39

which runs out of memory after 48 frames using tubelet lengtht= 2 and a “Large” backbone. Models processingmore frames (and thus more tokens) consistently achieve higher single- and multi-view accuracy, in line with our observations in previous experiments (Tab. 5, Fig. 8). Moroever, observe that by processing more frames (and thus more tokens) with Model 2, we are able to achieve higher accuracy than Model 1 (with fewer total FLOPs as well). Finally, we observed that for Model 2, the number of FLOPs effectively increases linearly with the number of input frames as the overall computation is dominated by the initial Spatial Transformer. As a result, the total number of FLOPs for the number of temporal views required to achieve maximum accuracy is constant across the models. In other words, ViViT-L/16x2 FE with 32 frames requires 995.3 GFLOPs per view, and 4 views to saturate multi-view accuracy. The 128-frame model requires 3980.4 GFLOPs but only a single view. As shown by Fig. 9, the latter model achieves the highest accuracy. 4.3. Comparison to state-of-the-art Based on our ablation studies in the previous section,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

40

achieves the highest accuracy. 4.3. Comparison to state-of-the-art Based on our ablation studies in the previous section, we compare to the current state-of-the-art using two of our model variants. We primarily use our Factorised Encoder model (Model 2), as it can process more tokens than Model 1 to achieve higher accuracy. Kinetics Tables 6a and 6b show that our spatio-temporal attention models outperform the state-of-the-art on Kinetics 400 and 600 respectively. Following standard practice, we take 3 spatial crops (left, centre and right) [21, 20, 66, 75] for each temporal view, and notably, we require significantly fewer views than previous CNN-based methods. We surpass the previous CNN-based state-of-the-art using ViViT-L/16x2 Factorised Encoder (FE) pretrained on ImageNet, and also outperform [4] who concurrently proposed a pure-transformer architecture. Moreover, by initialising our backbones from models pretrained on the larger JFT dataset [58], we obtain further improvements. Although these models are not directly comparable to previous work, we do also outperform [66] who pretrained on the large-scale, Instagram dataset [44]. Our best model uses

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

41

the large-scale, Instagram dataset [44]. Our best model uses a ViViT-H backbone pretrained on JFT and significantly advances the best reported results on Kinetics 400 and 600 to 84.9% and 85.8%, respectively. Moments in Time We surpass the state-of-the-art by a significant margin as shown in Tab. 6c. We note that the videos in this dataset are diverse and contain significant label noise, making this task challenging and leading to lower accuracies than on other datasets. Epic Kitchens 100 Table 6d shows that our Factorised Encoder model outperforms previous methods by a significant margin. In addition, our model obtains substantial improvements for Top-1 accuracy of “noun” classes, and the only method which achieves higher “verb” accuracy used optical flow as an additional input modality [43, 50]. Furthermore, all variants of our model presented in Tab. 2 outperformed the existing state-of-the-art on action accuracy. We note that we use the same model to predict verbs and nouns using two separate “heads”, and for simplicity, we do not use separate loss weights for each head.

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

42

nouns using two separate “heads”, and for simplicity, we do not use separate loss weights for each head. Something-Something v2 (SSv2) Finally, Tab. 6e shows that we achieve state-of-the-art Top-1 accuracy with our Factorised encoder model (Model 2), albeit with a smaller margin compared to previous methods. Notably, our Factorised encoder model significantly outperforms the concurrent TimeSformer [4] method by 2.9%, which also proposes a pure-transformer model, but does not consider our Factorised encoder variant or our additional regularisation. SSv2 differs from other datasets in that the backgrounds and objects are quite similar across different classes, meaning that recognising fine-grained motion patterns is necessary to distinguish classes from each other. Our results suggest that capturing these fine-grained motions is an area of improvement and future work for our model. We also note an inverse correlation between the relative performance of previous methods on SSv2 (Tab. 6e) and Kinetics (Tab. 6a) suggesting that these two datasets evaluate complementary characteristics of a model. 5. Conclusion and Future Work We have presented four pure-transformer models for

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

43

suggesting that these two datasets evaluate complementary characteristics of a model. 5. Conclusion and Future Work We have presented four pure-transformer models for video classification, with different accuracy and efficiency profiles, achieving state-of-the-art results across five popular datasets. Furthermore, we have shown how to effectively regularise such high-capacity models for training on smaller datasets and thoroughly ablated our main design choices. Future work is to remove our dependence onimage-pretrained models. Finally, going beyond video classification towards more complex tasks is a clear next step. References [1] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In ICCV , 2021. 1 [2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. In arXiv preprint arXiv:1607.06450 , 2016. 3 [3] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In ICCV , 2019. 1 [4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

44

[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? InarXiv preprint arXiv:2102.05095 , 2021. 2, 3, 4, 8, 9 [5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, et al. Language models are few-shot learners. In NeurIPS , 2020. 2 [6] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In CVPR Workshops , 2019. 2 [7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-toend object detection with transformers. In ECCV , 2020. 1, 2 [8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR ,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

45

2 [8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR , 2017. 1, 2, 5, 6, 8 [9] Yunpeng Chen, Yannis Kalantidis, Jianshu Li, Shuicheng Yan, and Jiashi Feng. A2-nets: Double attention networks. InNeurIPS , 2018. 2 [10] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. In arXiv preprint arXiv:1904.10509 , 2019. 2 [11] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. In ICLR , 2021. 2 [12] Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V . Le. Randaugment: Practical automated data augmentation with a reduced search space. In NeurIPS , 2020. 7, 13, 14

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

46

Le. Randaugment: Practical automated data augmentation with a reduced search space. In NeurIPS , 2020. 7, 13, 14 [13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision. In arXiv preprint arXiv:2006.13256 , 2020. 1, 6 [14] Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Łukasz Kaiser. Universal transformers. In ICLR , 2019. 2 [15] Mostafa Dehghani, Alexey Gritsenko, Anurag Arnab, Matthias Minderer, and Yi Tay. Scenic: A JAX library for computer vision research and beyond. arXiv preprint arXiv:2110.11403 , 2021. 6 [16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR , 2009. 2, 5, 6

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

47

and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR , 2009. 2, 5, 6 [17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL , 2019. 2, 3, 5 [18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR , 2021. 1, 2, 3, 5, 7 [19] Quanfu Fan, Chun-Fu Chen, Hilde Kuehne, Marco Pistoia, and David Cox. More is less: Learning efficient video representations by big-little network and depthwise temporal aggregation. In NeurIPS , 2019. 8 [20] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In CVPR , 2020. 1, 2, 6, 8

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

48

efficient video recognition. In CVPR , 2020. 1, 2, 6, 8 [21] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV , 2019. 1, 6, 8 [22] Christoph Feichtenhofer, Axel Pinz, and Richard Wildes. Spatiotemporal residual networks for video action recognition. In NeurIPS , 2016. 2, 5, 6 [23] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR , 2019. 1 [24] Rohit Girdhar and Deva Ramanan. Attentional pooling for action recognition. In NeurIPS , 2017. 4 [25] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In AISTATS , 2010. 6 [26] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

49

Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The” something something” video database for learning and evaluating visual common sense. InICCV , 2017. 1, 6 [27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR , 2016. 1, 2, 5 [28] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). In arXiv preprint arXiv:1606.08415 , 2016. 3 [29] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. Axial attention in multidimensional transformers. InarXiv preprint arXiv:1912.12180 , 2019. 4 [30] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR , 2018. 2 [31] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. In ECCV , 2016. 7, 13, 14

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

50

Weinberger. Deep networks with stochastic depth. In ECCV , 2016. 7, 13, 14 [32] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In ICCV , 2019. 2 [33] Boyuan Jiang, Mengmeng Wang, Weihao Gan, Wei Wu, and Junjie Yan. Stm: Spatiotemporal and motion encoding for action recognition. In ICCV , 2019. 8[34] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR , 2014. 2, 4 [35] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. In arXiv preprint arXiv:1705.06950 , 2017. 1, 2, 5, 6

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

51

arXiv:1705.06950 , 2017. 1, 2, 5, 6 [36] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In ICCV , 2019. 8 [37] Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In ICLR , 2020. 2 [38] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS , volume 25, 2012. 1, 2 [39] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Tom Duerig, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV , 2020. 5 [40] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

52

5 [40] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. InICLR , 2020. 2 [41] Ivan Laptev. On space-time interest points. IJCV , 64(2-3), 2005. 2 [42] Yan Li, Bin Ji, Xintian Shi, Jianguo Zhang, Bin Kang, and Limin Wang. Tea: Temporal excitation and aggregation for action recognition. In CVPR , 2020. 8 [43] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV , 2019. 8, 9 [44] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV , 2018. 8, 9 [45] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

53

[45] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl V ondrick, et al. Moments in time dataset: one million videos for event understanding. PAMI , 42(2):502–508, 2019. 1, 6 [46] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In arXiv preprint arXiv:2102.00719 , 2021. 2, 4 [47] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR , 2015. 2 [48] Zizheng Pan, Bohan Zhuang, Jing Liu, Haoyu He, and Jianfei Cai. Scalable visual transformers with hierarchical pooling. In arXiv preprint arXiv:2103.10619 , 2021. 2 [49] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

54

[49] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In ICML , 2018. 1, 2 [50] Will Price and Dima Damen. An evaluation of action recognition models on epic-kitchens. In arXiv preprint arXiv:1908.00867 , 2019. 9 [51] Zhaofan Qiu, Ting Yao, Chong-Wah Ngo, Xinmei Tian, and Tao Mei. Learning spatio-temporal representation with local and global diffusion. In CVPR , 2019. 8 [52] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR , 2020. 2 [53] Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya, and Jonathon Shlens. Stand-alone self-attention in vision models. In NeurIPS , 2019. 1, 2

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

55

Bello, Anselm Levskaya, and Jonathon Shlens. Stand-alone self-attention in vision models. In NeurIPS , 2019. 1, 2 [54] Michael S Ryoo, AJ Piergiovanni, Mingxing Tan, and Anelia Angelova. Assemblenet: Searching for multi-stream neural connectivity in video architectures. In ICLR , 2020. 8 [55] Zhuoran Shen, Irwan Bello, Raviteja Vemulapalli, Xuhui Jia, and Ching-Hui Chen. Global self-attention networks for image recognition. In arXiv preprint arXiv:2010.03019 , 2021. 2 [56] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS , 2014. 2, 4 [57] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In CVPR , 2021. 2 [58] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

56

[58] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In ICCV , 2017. 5, 6, 9 [59] Lin Sun, Kui Jia, Dit-Yan Yeung, and Bertram E Shi. Human action recognition using factorized spatio-temporal convolutional networks. In ICCV , 2015. 2 [60] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR , 2015. 1 [61] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR , 2016. 7, 13, 14 [62] Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. Long range arena: A benchmark for efficient transformers. In arXiv preprint

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

57

benchmark for efficient transformers. In arXiv preprint arXiv:2011.04006 , 2020. 2 [63] Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. Efficient transformers: A survey. In arXiv preprint arXiv:2009.06732 , 2020. 2 [64] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through attention. In arXiv preprint arXiv:2012.12877 , 2020. 1, 2, 7 [65] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV , 2015. 2 [66] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In ICCV , 2019. 2, 8, 9[67] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

58

LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR , 2018. 2 [68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS , 2017. 1, 2, 3, 4, 5, 7 [69] Heng Wang, Alexander Kl ¨aser, Cordelia Schmid, and Cheng-Lin Liu. Dense trajectories and motion boundary descriptors for action recognition. IJCV , 103(1), 2013. 2 [70] Heng Wang, Du Tran, Lorenzo Torresani, and Matt Feiszli. Video modeling with correlation networks. In CVPR , 2020. 8 [71] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In arXiv preprint arXiv:2012.00759 , 2020. 2 [72] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

59

arXiv:2012.00759 , 2020. 2 [72] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV , 2016. 4, 8 [73] Sinong Wang, Belinda Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. InarXiv preprint arXiv:2006.04768 , 2020. 2 [74] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In arXiv preprint arXiv:2102.12122 , 2021. 2 [75] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR , 2018. 1, 2, 8 [76] Xiaofang Wang, Xuehan Xiong, Maxim Neumann, AJ Piergiovanni, Michael S Ryoo, Anelia Angelova, Kris M Kitani,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

60

[76] Xiaofang Wang, Xuehan Xiong, Maxim Neumann, AJ Piergiovanni, Michael S Ryoo, Anelia Angelova, Kris M Kitani, and Wei Hua. Attentionnas: Spatiotemporal attention cell search for video classification. In ECCV , 2020. 8 [77] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In arXiv preprint arXiv:2011.14503 , 2020. 2 [78] Dirk Weissenborn, Oscar T ¨ackstr ¨om, and Jakob Uszkoreit. Scaling autoregressive video models. In ICLR , 2020. 4 [79] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR , 2019. 1 [80] Chao-Yuan Wu, Ross Girshick, Kaiming He, Christoph Feichtenhofer, and Philipp Krahenbuhl. A multigrid method

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

61

[80] Chao-Yuan Wu, Ross Girshick, Kaiming He, Christoph Feichtenhofer, and Philipp Krahenbuhl. A multigrid method for efficiently training video models. In CVPR , 2020. 8 [81] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In ECCV , 2018. 2 [82] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. In ICLR , 2018. 7, 13, 14 [83] Li Zhang, Dan Xu, Anurag Arnab, and Philip HS Torr. Dynamic graph message passing networks. In CVPR , 2020. 2 [84] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In arXiv preprint arXiv:2012.09164 , 2020. 2 [85] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

62

arXiv:2012.09164 , 2020. 2 [85] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In arXiv preprint arXiv:2012.15840 , 2020. 2 [86] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos. In ECCV , 2018. 4, 8 Appendix A. Additional experimental details In this appendix, we provide additional experimental details. Section A.1 provides additional details about the regularisers we used and Sec. A.2 details the training hyperparamters used for our experiments. A.1. Further details about regularisers In this section, we provide additional details and list the hyperparameters of the additional regularisers that we employed in Tab. 4. Hyperparameter values for all our experiments are listed in Tab. 7. Stochastic depth Stochastic depth regularisation was originally proposed for training very deep residual networks [31]. Intuitively, the outputs of a layer, `, are

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

63

Stochastic depth Stochastic depth regularisation was originally proposed for training very deep residual networks [31]. Intuitively, the outputs of a layer, `, are “dropped out” with probability, pdrop(`)during training, by setting the output of the layer to be equal to its input. Following [31], we linearly increase the probability of dropping a layer according to its depth within the network, pdrop(`) =` Lpdrop; (10) where`is the index of the layer in the network, and Lis the total number of layers. Random augment Random augment [12] randomly applies data augmentation transformations sequentially to an input example. We follow the public implementation1, but modify the data augmentation operations to be temporally consistent throughout the video (in other words, the same transformation is applied on each frame of the video). The authors define two hyperparameters for Random augment, “number of layers” , the number of augmentation transformations to apply sequentially to a video and “magnitude”, the strength of the transformation that is shared across all augmentation operations. Our values for these parameters are shown in Tab. 7. Label smoothing Label smoothing was proposed by [61]

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

64

across all augmentation operations. Our values for these parameters are shown in Tab. 7. Label smoothing Label smoothing was proposed by [61] originally to regularise training Inception-v3. Concretely, the label distribution used during training, ~y, is a mixture of the one-hot ground-truth label, y, and a uniform distribution,u, to encourage the network to produce less confident predictions during training: ~y= (1)y+u: (11) There is therefore one scalar hyperparamter, 2[0;1]. 1https://github.com/tensorflow/models/blob/ master/official/vision/beta/ops/augment.pyMixup Mixup [82] constructs virtual training examples which are a convex combination of pairs of training examples and their labels. Concretely, given (xi;yi)and(xj;yj) wherexidenotes an input vector and yia one-hot input label, mixup constructs the virtual training example, ~x=xi+ (1)xj ~y=yi+ (1)yj: (12) 2[0;1], and is sampled from a Beta distribution,

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

65

~y=yi+ (1)yj: (12) 2[0;1], and is sampled from a Beta distribution, Beta(;). Our choice of the hyperparameter is detailed in Tab. 7. A.2. Training hyperparameters Table 7 details the hyperparamters for all of our experiments. We use synchronous SGD with momentum, a cosine learning rate schedule with linear warmup, and a batch size of 64 for all experiments. As aforementioned, we only employed additional regularisation when training on the smaller Epic Kitchens and Something-Something v2 datasets. Table 7: Training hyperparamters for experiments in the main paper. “–” indicates that the regularisation method was not used at all. Values which are constant across all columns are listed once. Datasets are denoted as follows: K400: Kinetics 400. K600: Kinetics 600. MiT: Moments in Time. EK: Epic Kitchens. SSv2: Something-Something v2. K400 K600 MiT EK SSv2 Optimisation Optimiser Synchronous SGD Momentum 0.9 Batch size 64 Learning rate schedule cosine with linear warmup Linear warmup epochs 2.5

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

66

Optimiser Synchronous SGD Momentum 0.9 Batch size 64 Learning rate schedule cosine with linear warmup Linear warmup epochs 2.5 Base learning rate 0.1 0.1 0.25 0.5 0.5 Epochs 30 30 10 50 35 Data augmentation Random crop probability 1.0 Random flip probability 0.5 Scale jitter probability 1.0 Maximum scale 1.33 Minimum scale 0.9 Colour jitter probability 0.8 0.8 0.8 – – Rand augment number of layers [12] – – – 2 2 Rand augment magnitude [12] – – – 15 20 Other regularisation Stochastic droplayer rate, pdrop[31] – – – 0.2 0.3 Label smoothing [61] – – – 0.2 0.3 Mixup[82] – – – 0.1 0.3

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2302.01318

0

2023-2-3 Accelerating Large Language Model Decoding with Speculative Sampling Charlie Chen1, Sebastian Borgeaud1, Geoffrey Irving1, Jean-Baptiste Lespiau1, Laurent Sifre1and John Jumper1 1All authors from DeepMind We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model withinhardwarenumerics. WebenchmarkspeculativesamplingwithChinchilla,a70billionparameter language model, achieving a 2–2“5decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself. Introduction Scaling transformer models to 500B+ parameters has led to large performance improvements on many natural language, computer vision and reinforcement learning tasks (Arnab et al., 2021; Brown

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

1

Introduction Scaling transformer models to 500B+ parameters has led to large performance improvements on many natural language, computer vision and reinforcement learning tasks (Arnab et al., 2021; Brown etal.,2020;Chowdheryetal.,2022;Dosovitskiyetal.,2020;Hoffmannetal.,2022;Raeetal.,2021). However, transformer decoding remains a highly costly and inefficient process in this regime. Transformer sampling is typically memory bandwidth bound (Shazeer, 2019), so for a given set of hardware, the time to generate a single token in transformer models is proportional to a first order approximationtothesizeofparametersandthesizeofthetransformermemory. Thesizeoflanguage models also necessitates serving with model parallelism – adding communication overheads (Pope et al., 2022) and multiplying resource requirements. Since each new token depends on the past, many such transformer calls are required to sample a new sequence. We present an algorithm to accelerate transformer sampling for latency critical applications, which we call speculative sampling (SpS). This is achieved by: 1.Generating a short draft of length 𝐾. This can be attained with either a parallel model (Stern

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

2

we call speculative sampling (SpS). This is achieved by: 1.Generating a short draft of length 𝐾. This can be attained with either a parallel model (Stern et al., 2018) or by calling a faster, auto-regressive model 𝐾times. We shall refer to this model as thedraft model , and focus on the case where it is auto-regressive. 2.Scoring the draft using the larger, more powerful model from we wish to sample from. We shall refer to this model as the target model . 3.Using a modified rejection sampling scheme, accept a subset of the 𝐾draft tokens from left to right, recovering the distribution of the target model in the process. Intuitively, there are often sequences where the next token might be “obvious”. Therefore, if there is strongagreementbetweenthedraftandtargetmodel’sdistributionsonagiventokenorsub-sequence of tokens, this setup permits the generation of multiple tokens each time the target model is called . We show that the expected acceptance rate of draft tokens is sufficient to offset the overhead of the Corresponding author(s): ccharlie@deepmind.com

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

3

Corresponding author(s): ccharlie@deepmind.com ©2023 DeepMind. All rights reservedarXiv:2302.01318v1 [cs.CL] 2 Feb 2023 Accelerating Large Language Model Decoding with Speculative Sampling draftingprocessforlargelanguagemodels,resultinginaneffectiveandpracticalmethodforreducing sampling latency without the need for modifying the target model or biasing the sample distribution. Depending on the evaluation domain, SpS leads to a 2–2“5speedup when sampling from Chinchilla (Hoffmann et al., 2022). Notably, the mean tokens per second with SpS often exceeds the idealised ceiling on auto-regressive sampling speed imposed by the memory bandwidth. Related Work There has been a substantial body of work focused on improving sampling latency of large transformers and other auto-regressive models. Since sampling performance is heavily coupled with the model size in memory, quantisation to int8or even int4(Dettmers et al., 2022; Yao et al., 2022) and distillation (Jiao et al., 2020; Sanh et al., 2019) of transformers are effective techniques for reducing sampling latency with little to no

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

4

et al., 2019) of transformers are effective techniques for reducing sampling latency with little to no performance penalty. The observation that model size contributes less to the final performance than expected (Hoffmann et al., 2022) has also encouraged smaller language models in general. During sampling, a cache of the keys and values is maintained for every attention layer, and could become a memory bandwidth bottleneck as the batch size increases. Methods such as multi-query attention (Shazeer, 2019) aims to improve sampling performance by shrinking this cache. However thesetechniquesaremosteffectiveatmaximisingthroughout(atlargerbatchsizes)insteadoflatency, especially for larger models where the majority of the memory bandwidth budget is consumed by the parameters. Using a combination of the above techniques, in addition to a number of low-level optimisations to TPUs, Pope et al. (2022) have greatly improved the serving latency and efficiency of PaLM 540B. There is an existing body of similar work exploiting the efficiency of transformers and sequence models operating in parallel. This includes block parallel sampling (Stern et al., 2018), aggressive decoding (Ge et al., 2022), in addition to some work in parallelizing autoregressive models in the

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

5

decoding (Ge et al., 2022), in addition to some work in parallelizing autoregressive models in the image domain (Song et al., 2021; Wiggers and Hoogeboom, 2020). These methods have yet to be adapted to typical language model use-cases since they either only work with greedy sampling, bias the results or are focused on other modalities. Further, to our knowledge none of these techniques have been scaled to distributed setups, which is necessary for the most expensive decoders with the tens or hundreds of billions of parameters. Coincidentally, the work in this manuscript was undertaken concurrently and independently of the work on speculative decoding from Leviathan et al. (2022). We focus more heavily the distributed serving setting for large models and offer some incremental optimisations, but otherwise the core underlying idea is the same. Auto-regressive Sampling Whilst transformers can be trained efficiently and in parallel on TPUs and GPUs, samples are typically drawn auto-regressively (See algorithm 1). For most applications, auto-regressive sampling (ArS) is highlymemorybandwidthboundandthuscannotmakeeffectiveuseofmodernacceleratorhardware (Shazeer, 2019). A memory bound model call only generates a single token for every sequence in the

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

6

(Shazeer, 2019). A memory bound model call only generates a single token for every sequence in the batch, hence generating multiple tokens introduces a large amount of latency in any system which makes use of it. 2 Accelerating Large Language Model Decoding with Speculative Sampling Thisisespeciallyproblematicasthenumberofparameters inthemodelincreases. Sinceallthemodel parameters need to pass through at least one accelerator chip, the model size divided by the total memory bandwidth across all chips gives us a hard ceiling on the maximum auto-regressive sampling speed. Larger models also require serving on multiple accelerators, introducing a further source of latency due to inter-device communication overheads. Algorithm 1 Auto-regressive (ArS) with Auto-Regressive Models Givenauto-regressivetargetmodel 𝑞¹“j“ºandinitialpromptsequence 𝑥1”“““”𝑥𝑡andtargetsequence length𝑇. Initialise𝑛 𝑡. while𝑛 𝑇do Sample𝑥𝑛¸1𝑞¹𝑥j𝑥1”“““”𝑥𝑛º

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

7

𝑛 𝑛¸1 end while Algorithm 2 Speculative Sampling (SpS) with Auto-Regressive Target and Draft Models Given lookahead 𝐾and minimum target sequence length 𝑇. Given auto-regressive target model 𝑞¹“j“º, and auto-regressive draft model 𝑝¹“j“º, initial prompt sequence𝑥0”“““”𝑥𝑡. Initialise𝑛 𝑡. while𝑛 𝑇do for𝑡=1 :𝐾do Sample draft auto-regressively ˜𝑥𝑡𝑝¹𝑥j”𝑥1”“““”𝑥𝑛”˜𝑥1”“““” ˜𝑥𝑡1º end for In parallel, compute 𝐾¸1sets of logits from drafts ˜𝑥1”“““” ˜𝑥𝐾:

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

8

𝑞¹𝑥j”𝑥1”“““”𝑥𝑛º” 𝑞¹𝑥j”𝑥1”“““”𝑥𝑛”˜𝑥1º” “““” 𝑞¹𝑥j”𝑥1”“““”𝑥𝑛”˜𝑥1”“““” ˜𝑥𝐾º for𝑡=1 :𝐾do Sample𝑟𝑈»0”1¼from a uniform distribution. if𝑟 min 1”𝑞¹𝑥j𝑥1”“““”𝑥𝑛¸𝑡1º 𝑝¹𝑥j𝑥1”“““”𝑥𝑛¸𝑡1º ,then Set𝑥𝑛¸𝑡 ˜𝑥𝑡and𝑛 𝑛¸1.

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

9

,then Set𝑥𝑛¸𝑡 ˜𝑥𝑡and𝑛 𝑛¸1. else sample𝑥𝑛¸𝑡¹𝑞¹𝑥j𝑥1”“““”𝑥𝑛¸𝑡1º𝑝¹𝑥j𝑥1”“““”𝑥𝑛¸𝑡1ºº¸and exit for loop. end if end for If all tokens 𝑥𝑛¸1”“““”𝑥𝑛¸𝐾are accepted, sample extra token 𝑥𝑛¸𝐾¸1𝑞¹𝑥j”𝑥1”“““”𝑥𝑛”𝑥𝑛¸𝐾ºand set𝑛 𝑛¸1. end while Speculative Sampling Conditional Scoring For speculative sampling (See algorithm 2), we first make the observation that computing the logits

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

10

end while Speculative Sampling Conditional Scoring For speculative sampling (See algorithm 2), we first make the observation that computing the logits of a short continuation of 𝐾tokens in parallel has a very similar latency to that of sampling a single 3 Accelerating Large Language Model Decoding with Speculative Sampling token. We focus our attention on large transformers, sharded in the Megatron style (Shoeybi et al., 2019). For these models the majority of sampling time can be attributed to three components: 1.Linear Layers: For small batch sizes, each linear layer only processes a small number of embeddings. This causes the dense matrix multiplies in the feed-forward layers, queries, keys, values computations and the final attention projection to become memory bound. For small 𝐾, this will continue to be memory bound and therefore take a similar amount of time. 2.The Attention Mechanism: Theattentionmechanismisalsomemorybound. Duringsampling, we maintain a cache of all the keys and values of the previous tokens in the sequence to avoid re-computation. These KV-caches are large, and accounts for the majority of the memory

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

11

re-computation. These KV-caches are large, and accounts for the majority of the memory bandwidth utilisation for the attention mechanism. However, since the KV-cache size does not change as we increase 𝐾, there is little to no delta in this component. 3.All-reduces: Asmodelsgrowinsize, itsparametersneedtobedividedacrossmultipleaccelerators, leading to communication overheads. With Megatron, this manifests itself as an all-reduce after every feed-forward and attention layer. Since only the activations for a small number of tokens are transmitted, this operation is typically latency bound instead of throughput bound for both sampling and scoring (for small 𝐾). Again, this results in a similar amount of time spent in the two cases. Other sources of overhead may exist, depending on the exact transformer implementation. Therefore it is still possible that the choice of positioning encoding, decoding method (e.g. a sort might be required for nucleus sampling), hardware limitations etc. can introduce some deltas between scoring and sampling. However, if the conditions are met such that the above components dominate then scoring should not be significantly slower for small 𝐾. Modified Rejection Sampling

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

12

scoring should not be significantly slower for small 𝐾. Modified Rejection Sampling We require a method to recover the distribution of the target model from samples from the draft model, and logits of said tokens from both models. To achieve this, we introduce the following rejection sampling scheme of the drafted tokens. Given a sequence of tokens 𝑥1”“““”𝑥𝑛, and𝐾draft tokens ˜𝑥𝑛¸1”“““” ˜𝑥𝑛¸𝐾generated from 𝑝¹“j“º, we accept ˜𝑥𝑛¸1 with probability: min 1”𝑞¹˜𝑥𝑛¸1j𝑥1”“““”𝑥𝑛º 𝑝¹˜𝑥𝑛¸1j𝑥1”“““”𝑥𝑛º

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

13

𝑝¹˜𝑥𝑛¸1j𝑥1”“““”𝑥𝑛º Where𝑞¹˜𝑥𝑛¸1j𝑥1”“““”𝑥𝑛ºand𝑝¹˜𝑥𝑛¸1j𝑥1”“““”𝑥𝑛ºare the probability of ˜𝑥𝑛¸1according to the target and draft models respectively, conditioned on the context so far. If the token is accepted, we set 𝑥𝑛¸1 ˜𝑥𝑛¸1and repeat the process for ˜𝑥𝑛¸2until either a token is rejected or all tokens have been accepted. If˜𝑥𝑛¸1is rejected, we resample 𝑥𝑛¸1from the following distribution:

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

14

If˜𝑥𝑛¸1is rejected, we resample 𝑥𝑛¸1from the following distribution: 𝑥𝑛¸1¹𝑞¹𝑥j𝑥1”“““”𝑥𝑛º𝑝¹𝑥j𝑥1”“““”𝑥𝑛ºº¸ Where¹“º¸denotes: ¹𝑓¹𝑥ºº¸=max¹0” 𝑓¹𝑥ººÍ 𝑥max¹0” 𝑓¹𝑥ºº By applying this sequentially, we recover the distribution of the target model for the accepted tokens (see proof in Theorem 1) within hardware numerics. Note that: 4 Accelerating Large Language Model Decoding with Speculative Sampling •At least one token will always be generated from a draft-accept loop – if the first token is rejected, a valid token is resampled. •Since the final token of the draft gives us the logits for the next token, if every drafted token is

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

15

rejected, a valid token is resampled. •Since the final token of the draft gives us the logits for the next token, if every drafted token is accepted, we can sample from it normally. This gives us a maximum of 𝐾¸1tokens per loop, over the naive implementation which would only return 𝐾tokens. With standard sampling methods such as nucleus, top-k sampling and adjusting temperature, we can modify the probabilities accordingly before applying this rejection sampling scheme. We have observed that the overall acceptance rate is robust to the exact parameters used. Because we do not interact with the body of the transformer itself, this method can be used in conjunction many other techniques for accelerating or optimising the memory use of sampling, such as quantisation and multi-query attention. Choice of Draft Models Since the acceptance criterion guarantees the distribution of the target model in our samples, we are free to choose the method for drafting a continuation as long as it exposes logits, and there is a high enough acceptance rate and/or low enough latency to break-even. There exist several approaches here: •Incorporating draft generation into the target model, and train the model from the start. This is the strategy used by Stern et al. (2018), which adds multiple heads into the transformer to

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

16

is the strategy used by Stern et al. (2018), which adds multiple heads into the transformer to generate multiple tokens. •Using sequence level distillation (Kim and Rush, 2016) to generate a second model which predicts𝐾tokens in parallel. This strategy was employed by Ge et al. (2022). •Set a portion of the activations of the target model as an input to the draft model, and train the draft model with this input. Although these methods will likely yield powerful drafts, they require a large number of data generated from the target model or changes to the target model. Sequence level distillation in particular would require a large compute budget. This makes them less practical for large scale applications. Whilst large language models produce better samples, intuitively there are "easier" tokens to predict for which smaller models may be sufficient. Therefore we may simply use a smaller version of the target language model as the draft and obtain high acceptance rates. This would also be convenient from an engineering and workflow perspective, since robust tooling for such models should already exist to train the target model in the first place. Results We train a 4 billion parameter draft model optimised for sampling latency on 16 TPU v4s – the same hardware that is typically used to serve Chinchilla for research purposes. This model was trained

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

17

hardware that is typically used to serve Chinchilla for research purposes. This model was trained with the same tokeniser and dataset as Chinchilla, with a slightly smaller width and with only 8 layers. The relatively few number of layers allows it to achieve a sampling speed of 1.8ms/token comparedto 14.1ms/token forChinchilla. Fordetails,pleaserefertothehyperparametersinTable2. For distributed setups it is insufficient to naively choose a small model as the draft, since different models have different optimal inference setups. For example, it is typical to serve Chinchilla 70B 5 Accelerating Large Language Model Decoding with Speculative Sampling Table1jChinchillaperformanceandspeedonXSumandHumanEvalwithnaiveandspeculative sampling at batch size 1 and 𝐾=4. XSum was executed with nucleus parameter 𝑝=0“8, and HumanEval with 𝑝=0“95and temperature 0“8. Sampling Method Benchmark Result Mean Token Time Speed Up ArS (Nucleus)XSum (ROUGE-2)0.112 14.1ms/Token 1

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

18

Sampling Method Benchmark Result Mean Token Time Speed Up ArS (Nucleus)XSum (ROUGE-2)0.112 14.1ms/Token 1 SpS (Nucleus) 0.114 7.52ms/Token 1“92 ArS (Greedy)XSum (ROUGE-2)0.157 14.1ms/Token 1 SpS (Greedy) 0.156 7.00ms/Token 2“01 ArS (Nucleus)HumanEval (100 Shot)45.1% 14.1ms/Token 1 SpS (Nucleus) 47.0% 5.73ms/Token 2“46 on 16 TPU v4s (where it achieves the aforementioned 14.1ms/token ), whereas a chinchilla-optimal 7B achieves its lowest sampling latency on 4 TPU v4s (where it achieves 5ms/token ). For smaller models, the additional memory bandwidth and flops are insufficient to offset the additional communication overhead between more devices – serving a 7B on 16 TPUs actually increases the latency. This means the 7B would provide only a modest speedup if used as a draft with its optimal topology, and

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

19

means the 7B would provide only a modest speedup if used as a draft with its optimal topology, and we will not make full utilisation of the hardware during drafting. We can sidestep this issue by training a wider model with a relatively few number of layers in order to minimise communication overhead. It has been observed that the performance of language models is relatively robust to changes in model aspect ratio (Levine et al., 2020), so this allows us to serve a powerful draft model which can be sampled rapidly on the same hardware as the target model. Evaluation on XSum and HumanEval We evaluate speculative sampling with Chinchilla on two tasks and summarize the results in Table 1: •The XSum (Narayan et al., 2018) benchmark. This is a natural language summarisation task using a 1-shot prompt where we sample a total of 11,305 sequences with a maximum sequence length 128. •The 100-shot HumanEval task (Chen et al., 2021). This is a code generation task involves the generation of 16,400 samples with a maximum sequence length of 512. Even with greedy sampling, a single token deviating due to numerics could result in two sequences diverging wildly. Since pseudo-random seeds are processed differently between ArS and SpS, and

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

20

diverging wildly. Since pseudo-random seeds are processed differently between ArS and SpS, and because the different computation graphs lead to different numerics, we cannot not expect identical outputs. However, we expect the samples to come from the same distribution within numerics and we empirically verify this by evaluating these benchmarks. We run the tasks at batch size 1 with SpS and ArS. The time taken per SpS/ArS loop has low variance, and we can measure it directly from TPU profiles. To obtain the average speedup, standard deviations and other metrics, we log the amount of tokens generated for each speculative loop. In Table 1 we show the performance on the XSum and HumanEval benchmarks for naive and speculative sampling with Chinchilla. 6 Accelerating Large Language Model Decoding with Speculative Sampling We obtain a substantial speedup in both tasks, with HumanEval reaching speedups of almost 2“5. Yet, we have parity in the benchmark metrics – the underlying samples distribution is provably the sameuptonumerics, andthisverifiesthatthedraftmodelisnotbiasingtheresultsempirically. Inthe case of HumanEval and greedy XSum, this speedup exceeded the theoretical memory bandwidth limit

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

21

case of HumanEval and greedy XSum, this speedup exceeded the theoretical memory bandwidth limit of the hardware for autoregressive sampling (model size divided by the total memory bandwidth). Acceptance rate changes per domain It is apparent that the acceptance rate is dependent on the application and the decoding method. HumanEvalachievesasignificantlylargerspeedup—Wehypothesizethatthisisduetoacombination of code containing a lot of common sub-sequences (e.g. for i in range(len(arr)): would be relatively easy for a draft model to guess), is often decomposed into a smaller set of shorter tokens and the temperature value sharpening both the draft and target logits. 01234567 Number of draft tokens (K)40060080010001200140016001800ms Mean Sampling Time (128 tokens) Human Eval XSum 01234567 Number of draft tokens (K)0.50.60.70.80.91.0 Acceptance rate Human Eval XSum 01234567 Number of draft tokens (K)1416182022242628ms Total loop time Figure 1jLeft:The average time to generate 128 tokens, with standard deviation. Note that as 𝐾

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

22

Total loop time Figure 1jLeft:The average time to generate 128 tokens, with standard deviation. Note that as 𝐾 increases, the overall speedup plateaus or even regresses, with XSum being optimal at 𝐾=3. The variance consistently increases with 𝐾.Middle:The average number of tokens accepted divided by 𝐾¸1– this serves as a measure of the overall efficiency of the modified rejection scheme, which decreases with the lookahead. Right:Average time per loop increases approximately linearly with 𝐾 duetotheincreasednumberofmodelcalls. Notethatthegradientisslightlyhigherthanthesampling speed of the draft model, due to additional overheads in nucleus decoding. Trade off between longer drafts and more frequent scoring Wevisualisethetrade-offofincreasing 𝐾,thenumberoftokenssampledbythedraftmodelinFigure1. As𝐾increases, we need fewer scoring calls from the large models to generate the same sequence length, potentially giving us a larger speedup. However, the total loop time increases approximately linearly with the larger number of draft model calls and small increases in the scoring time. The

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

23

length, potentially giving us a larger speedup. However, the total loop time increases approximately linearly with the larger number of draft model calls and small increases in the scoring time. The overall efficiency of the proportion of accepted tokens decreases as 𝐾increases, since later tokens depend on the acceptance of previous tokens. This results in the average speedup plateauing or even degrading with a larger 𝐾(for example, XSum with nucleus’s latency is minimised at 𝐾=3), depending on the domain. 7 Accelerating Large Language Model Decoding with Speculative Sampling Further, even though larger values of 𝐾may yield marginally greater mean speedups in certain circumstances, it also increases variance of the time to generate a full sequence. This could be problematic for settings where the P90, P99 latencies of concern. Conclusion Inthiswork,wedemonstrateanewalgorithmandworkflowforacceleratingthedecodingoflanguage models. Speculative sampling does not require making any modifications to the target language model’s parameters or architecture, is provably lossless within numerics, scales well with the appropriate draft model and complements many existing techniques for reducing latency in the small batch size setting.

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

24

size setting. We optimise and scale the technique to Chinchilla 70B using a draft model which was easy to trainwithexistinginfrastructure, demonstratingthatityieldsalargespeedupacrossbenchmarktasks and common decoding methods in the process. We verify that it is indeed lossless empirically in its downstream tasks. References A.Arnab,M.Dehghani,G.Heigold,C.Sun,M.Lucic,andC.Schmid. Vivit: Avideovisiontransformer. In2021 IEEE/CVF International Conference on Computer Vision (ICCV) , pages 6816–6826. IEEE Computer Society, 2021. T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners. Advances in neural information processing systems , 33:1877–1901, 2020. M.Chen,J.Tworek,H.Jun,Q.Yuan,H.P.deOliveiraPinto,J.Kaplan,H.Edwards,Y.Burda,N.Joseph,

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

25

G.Brockman,A.Ray,R.Puri,G.Krueger,M.Petrov,H.Khlaaf,G.Sastry,P.Mishkin,B.Chan,S.Gray, N. Ryder, M. Pavlov, A. Power, L. Kaiser, M. Bavarian, C. Winter, P. Tillet, F. P. Such, D. Cummings, M.Plappert,F.Chantzis,E.Barnes,A.Herbert-Voss,W.H.Guss,A.Nichol,A.Paino,N.Tezak,J.Tang, I. Babuschkin, S. Balaji, S. Jain, W. Saunders, C. Hesse, A. N. Carr, J. Leike, J. Achiam, V. Misra, E. Morikawa, A. Radford, M. Knight, M. Brundage, M. Murati, K. Mayer, P. Welinder, B. McGrew, D. Amodei, S. McCandlish, I. Sutskever, and W. Zaremba. Evaluating large language models trained

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

26

D. Amodei, S. McCandlish, I. Sutskever, and W. Zaremba. Evaluating large language models trained on code. CoRR, abs/2107.03374, 2021. URL https://arxiv.org/abs/2107.03374 . A. Chowdhery, S. Narang, J. Devlin, M. Bosma, G. Mishra, A. Roberts, P. Barham, H. W. Chung, C. Sutton, S. Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 , 2022. T. Dettmers, M. Lewis, Y. Belkada, and L. Zettlemoyer. Llm. int8 (): 8-bit matrix multiplication for transformers at scale. arXiv preprint arXiv:2208.07339 , 2022. A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer,G.Heigold,S.Gelly,etal. Animageisworth16x16words: Transformersforimagerecognition

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

27

at scale.arXiv preprint arXiv:2010.11929 , 2020. T.Ge,H.Xia,X.Sun,S.Chen,andF.Wei. Losslessaccelerationforseq2seqgenerationwithaggressive decoding. ArXiv, abs/2205.10350, 2022. 8 Accelerating Large Language Model Decoding with Speculative Sampling J. Hoffmann, S. Borgeaud, A. Mensch, E. Buchatskaya, T. Cai, E. Rutherford, D. d. L. Casas, L. A. Hendricks,J.Welbl,A.Clark,etal. Trainingcompute-optimallargelanguagemodels. arXivpreprint arXiv:2203.15556 , 2022. X. Jiao, Y. Yin, L. Shang, X. Jiang, X. Chen, L. Li, F. Wang, and Q. Liu. TinyBERT: Distilling BERT

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

28

for natural language understanding. In Findings of the Association for Computational Linguistics: EMNLP 2020 , pages 4163–4174, Online, Nov. 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.findings-emnlp.372. URL https://aclanthology.org/2020. findings-emnlp.372 . Y. Kim and A. M. Rush. Sequence-level knowledge distillation. CoRR, abs/1606.07947, 2016. URL http://arxiv.org/abs/1606.07947 . Y. Leviathan, M. Kalman, and Y. Matias. Fast inference from transformers via speculative decoding. ArXiv, abs/2211.17192, 2022. Y. Levine, N. Wies, O. Sharir, H. Bata, and A. Shashua. The depth-to-width interplay in self-attention. arXiv preprint arXiv:2006.12467 , 2020. S. Narayan, S. B. Cohen, and M. Lapata. Don’t give me the details, just the summary! topicaware convolutional neural networks for extreme summarization. In Proceedings of the 2018

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

29

Conference on Empirical Methods in Natural Language Processing , pages 1797–1807, Brussels, Belgium, Oct.-Nov. 2018. Association for Computational Linguistics. doi: 10.18653/v1/D18-1206. URLhttps://aclanthology.org/D18-1206 . R. Pope, S. Douglas, A. Chowdhery, J. Devlin, J. Bradbury, A. Levskaya, J. Heek, K. Xiao, S. Agrawal, and J. Dean. Efficiently scaling transformer inference. arXiv preprint arXiv:2211.05102 , 2022. J. W. Rae, S. Borgeaud, T. Cai, K. Millican, J. Hoffmann, F. Song, J. Aslanides, S. Henderson, R. Ring, S. Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446 , 2021. V. Sanh, L. Debut, J. Chaumond, and T. Wolf. Distilbert, a distilled version of bert: smaller, faster,

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

30

V. Sanh, L. Debut, J. Chaumond, and T. Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108 , 2019. N. Shazeer. Fast transformer decoding: One write-head is all you need. CoRR, abs/1911.02150, 2019. URLhttp://arxiv.org/abs/1911.02150 . M. Shoeybi, M. Patwary, R. Puri, P. LeGresley, J. Casper, and B. Catanzaro. Megatron-lm: Training multi-billionparameterlanguagemodelsusingmodelparallelism. arXivpreprintarXiv:1909.08053 , 2019. Y.Song, C.Meng, R.Liao, andS.Ermon. Acceleratingfeedforwardcomputationviaparallelnonlinear equationsolving. InM.MeilaandT.Zhang,editors, Proceedingsofthe38thInternationalConference on Machine Learning , volume 139 of Proceedings of Machine Learning Research , pages 9791–9800.

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

31

on Machine Learning , volume 139 of Proceedings of Machine Learning Research , pages 9791–9800. PMLR, 18–24 Jul 2021. URL https://proceedings.mlr.press/v139/song21a.html . M. Stern, N. Shazeer, and J. Uszkoreit. Blockwise parallel decoding for deep autoregressive models. CoRR, abs/1811.03115, 2018. URL http://arxiv.org/abs/1811.03115 . A. Wiggers and E. Hoogeboom. Predictive sampling with forecasting autoregressive models. In H. D. III and A. Singh, editors, Proceedings of the 37th International Conference on Machine Learning , volume 119 of Proceedings of Machine Learning Research , pages 10260–10269. PMLR, 13–18 Jul 2020. URL https://proceedings.mlr.press/v119/wiggers20a.html . 9 Accelerating Large Language Model Decoding with Speculative Sampling Z. Yao, R. Y. Aminabadi, M. Zhang, X. Wu, C. Li, and Y. He. Zeroquant: Efficient and affordable

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

32

post-training quantization for large-scale transformers. arXiv preprint arXiv:2206.01861 , 2022. Supplementary Materials Author Contributions •Initial proposal: Charlie Chen, John Jumper and Geoffrey Irving •Initial Implementation, Optimisation and Scaling: Charlie Chen •Modified Rejection Sampling Scheme: John Jumper •Engineering Improvements: Jean-Baptiste Lespiau and Charlie Chen •Experiments: Charlie Chen, Sebastian Borgeaud and Laurent Sifre •Draft of Manuscript: Charlie Chen and Sebastian Borgeaud •Manuscript Feedback: Laurent Sifre, Geoffrey Irving and John Jumper Acknowledgements We’d like to thank Oriol Vinyals and Koray Kavukcuoglu for your kind advice and leadership. We’d also like to thank Evan Senter for your additional feedback on the manuscript and Amelia Glaese for your support in navigating the publishing process. Finally, we’d like to thank Blake Hechtman, Berkin Ilbeyi for your valuable advice on XLA and Nikolai Grigoriev for our discussions on the various tricks that can be applied to the transformer architecture. Hyperparams Table 2jHyperparameters for the draft model Model 𝑑modelHeads Layers Params

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

33

that can be applied to the transformer architecture. Hyperparams Table 2jHyperparameters for the draft model Model 𝑑modelHeads Layers Params Target (Chinchilla) 8192 64 80 70B Draft 6144 48 8 4B Proofs Theorem1 (ModifiedRejectionSamplingrecoversthetargetdistribution) .Givendiscretedistributions 𝑞,𝑝and a single draft sample ˜𝑥𝑝, let𝑋be the final resulting sample. For 𝑋=𝑥to be true, we must either sample ˜𝑥=𝑥and then accept it, or resample it after ˜𝑥(of any value) is rejected. Hence: ℙ¹𝑋=𝑥º =ℙ¹˜𝑥=𝑥ºℙ¹˜𝑥acceptedj˜𝑥=𝑥º¸ℙ¹˜𝑥rejectedºℙ¹𝑋=𝑥j˜𝑥rejectedº For the first term, we apply the acceptance rule:

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

34

For the first term, we apply the acceptance rule: ℙ¹˜𝑥=𝑥ºℙ¹˜𝑥acceptedj˜𝑥=𝑥º =𝑝¹𝑥ºmin 1”𝑞¹𝑥º 𝑝¹𝑥º 10 Accelerating Large Language Model Decoding with Speculative Sampling =min¹𝑝¹𝑥º”𝑞¹𝑥ºº For the second conditional term, we apply the resampling rule: ℙ¹𝑋=𝑥j˜𝑥rejectedº=¹𝑞¹𝑥º𝑝¹𝑥ºº¸ Where¹“º¸denotes: ¹𝑓¹𝑥ºº¸=max¹0” 𝑓¹𝑥ººÍ 𝑥max¹0” 𝑓¹𝑥ºº Finally, we calculate the probability of rejection: ℙ¹˜𝑥rejectedº=1ℙ¹˜𝑥acceptedº

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

35

Finally, we calculate the probability of rejection: ℙ¹˜𝑥rejectedº=1ℙ¹˜𝑥acceptedº =1∑︁ 𝑥0ℙ¹𝑋=𝑥0”˜𝑥acceptedº =1∑︁ 𝑥0min¹𝑝¹𝑥0º”𝑞¹𝑥0ºº =∑︁ 𝑥0max¹0”𝑞¹𝑥0º𝑝¹𝑥0ºº =∑︁ 𝑥0𝑞¹𝑥0ºmin¹𝑝¹𝑥0º”𝑞¹𝑥0ºº =∑︁ 𝑥0max¹0”𝑞¹𝑥0º𝑝¹𝑥0ºº This is equal to the denominator of ¹𝑞¹𝑥º𝑝¹𝑥ºº¸, so:

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2302.01318

36

This is equal to the denominator of ¹𝑞¹𝑥º𝑝¹𝑥ºº¸, so: ℙ¹˜𝑥rejectedºℙ¹𝑋=𝑥j˜𝑥rejectedº=max¹0”𝑞¹𝑥º𝑝¹𝑥ºº Hence: ℙ¹𝑋=𝑥º =min¹𝑝¹𝑥º”𝑞¹𝑥ºº¸max¹0”𝑞¹𝑥º𝑝¹𝑥ºº =𝑞¹𝑥º and we have recovered the desired target. 11

2302.01318

Accelerating Large Language Model Decoding with Speculative Sampling

We present speculative sampling, an algorithm for accelerating transformer decoding by enabling the generation of multiple tokens from each transformer call. Our algorithm relies on the observation that the latency of parallel scoring of short continuations, generated by a faster but less powerful draft model, is comparable to that of sampling a single token from the larger target model. This is combined with a novel modified rejection sampling scheme which preserves the distribution of the target model within hardware numerics. We benchmark speculative sampling with Chinchilla, a 70 billion parameter language model, achieving a 2-2.5x decoding speedup in a distributed setup, without compromising the sample quality or making modifications to the model itself.

http://arxiv.org/pdf/2302.01318

cs.CL

20230202

20230202

2305.13245

0

GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints Joshua Ainslie, James Lee-Thorp, Michiel de Jongyy Yury Zemlyanskiy ,Federico Lebrón ,Sumit Sanghai Google Research Abstract Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA. 1 Introduction Autoregressive decoder inference is a severe bottleneck for Transformer models due to the memory bandwidth overhead from loading decoder weights and all attention keys and values at every decoding step (Shazeer, 2019; Pope et al., 2022; de Jong et al., 2022). The memory bandwidth from loading

2305.13245

GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints

Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.

http://arxiv.org/pdf/2305.13245

cs.CL

20230522

20230522

2305.13245

1

and all attention keys and values at every decoding step (Shazeer, 2019; Pope et al., 2022; de Jong et al., 2022). The memory bandwidth from loading keys and values can be sharply reduced through multi-query attention (Shazeer, 2019), which uses multiple query heads but single key and value heads. However, multi-query attention ( MQA ) can lead to quality degradation and training instability, and it may not be feasible to train separate models optimized for quality and inference. Moreover, while some language models already use multiquery attention, such as PaLM (Chowdhery et al., 2022), many do not, including publicly available language models such as T5 (Raffel et al., 2020) and LLaMA (Touvron et al., 2023). This work contains two contributions for faster inference with large language models. First, we Equal contribution. yUniversity of Southern California. Work done at Google Research.show that language model checkpoints with multihead attention ( MHA ) can be uptrained (Komatsuzaki et al., 2022) to use MQA with a small fraction of original training compute. This presents a cost-effective method to obtain fast multi-query as well as high-quality MHA checkpoints. Second, we propose grouped-query attention

2305.13245

GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints

Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.

http://arxiv.org/pdf/2305.13245

cs.CL

20230522

20230522

2305.13245

2

cost-effective method to obtain fast multi-query as well as high-quality MHA checkpoints. Second, we propose grouped-query attention (GQA ), an interpolation between multi-head and multi-query attention with single key and value heads per subgroup of query heads . We show that uptrained GQA achieves quality close to multihead attention while being almost as fast as multiquery attention. 2 Method 2.1 Uptraining Generating a multi-query model from a multi-head model takes place in two steps: first, converting the checkpoint, and second, additional pre-training to allow the model to adapt to its new structure. Figure 1 shows the process for converting a multi-head checkpoint into a multi-query checkpoint. The projection matrices for key and value heads are mean pooled into single projection matrices, which we find works better than selecting a single key and value head or randomly initializing new key and value heads from scratch. Figure 1: Overview of conversion from multi-head to multi-query attention. Key and value projection matrices from all heads are mean pooled into a single head. The converted checkpoint is then pre-trained forarXiv:2305.13245v1 [cs.CL] 22 May 2023

2305.13245

GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints

Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.

http://arxiv.org/pdf/2305.13245

cs.CL

20230522

20230522

2305.13245

3

The converted checkpoint is then pre-trained forarXiv:2305.13245v1 [cs.CL] 22 May 2023 Figure 2: Overview of grouped-query method. Multi-head attention has H query, key, and value heads. Multi-query attention shares single key and value heads across all query heads. Grouped-query attention instead shares single key and value heads for each group of query heads, interpolating between multi-head and multi-query attention. a small proportion of its original training steps on the same pre-training recipe. 2.2 Grouped-query attention Grouped-query attention divides query heads into Ggroups , each of which shares a single key head and value head. GQA -Grefers to grouped-query withGgroups. GQA -1, with a single group and therefore single key and value head, is equivalent to MQA , while GQA -H, with groups equal to number of heads, is equivalent to MHA . Figure 2 shows a comparison of grouped-query attention and multihead/multi-query attention. When converting a multi-head checkpoint to a GQA checkpoint, we construct each group key and value head by meanpooling all the original heads within that group. An intermediate number of groups leads to an

2305.13245

GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints

Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.

http://arxiv.org/pdf/2305.13245

cs.CL

20230522

20230522

2305.13245

4

construct each group key and value head by meanpooling all the original heads within that group. An intermediate number of groups leads to an interpolated model that is higher quality than MQA but faster than MHA , and, as we will show, represents a favorable trade-off. Going from MHA toMQA reducesHkey and value heads to a single key and value head, reducing the size of the key-value cache and therefore amount of data that needs to be loaded by a factor of H. However, larger models generally scale the number of heads, such that multi-query attention represents a more aggressive cut in both memory bandwidth and capacity. GQA lets us keep the same proportional decrease in bandwidth and capacity as model size increases. Moreover, larger models suffer relatively less from memory bandwidth overhead from attention, as the KV-cache scales with model dimension while model FLOPs and parameters scale with the square of model dimension. Finally, standard sharding for large models replicates the single key and value head by the number of model partitions (Popeet al., 2022); GQA removes the waste from such partitioning. Therefore, we expect GQA to present a particularly good trade-off for larger models. 3 Experiments 3.1 Experimental setup Configurations All models are based on the

2305.13245

GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints

Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.

http://arxiv.org/pdf/2305.13245

cs.CL

20230522

20230522