Model Details
This is an unofficial implementation of ALIGN trained on COYO-700M. The official ALIGN is trained on its dataset of 1.8B samples. That dataset is not released to the public. Instead, we trained our implementation of ALIGN model on COYO-700M.
It's developed by Kakao Brain to validate the performance of COYO-700M dataset on a large-scale model.
The training took about 8 days on TPU V3-512.
Model Date
April 2022
Model Type
This is dual encoder model where
- image encoder is using EfficientNet-B7 architecture
- text encoder is using BERT-base architecture
Training data
This model is trained on COYO-700M dataset.
Evaluation results
| Dataset | ImageNet | Flickr30k | MsCOCO | |||
|---|---|---|---|---|---|---|
| KNN | I2T R@1 | T2I R@1 | I2T R@1 | T2I R@1 | ||
| ALIGN-L2-Large(Google) | ALIGN 1.8B | 76.4 | 88.6 | 75.7 | 58.6 | 45.6 |
| ALIGN-B7-Base(Google) | ALIGN 1.8B | 69.3 | - | - | 55.4 | 41.7 |
| COYO-ALIGN-B7-Base(Kakao Brain) | COYO-700M | 68.6 | 88.1 | 73.2 | 61.2 | 43.1 |