File size: 8,774 Bytes
72fd365
 
 
 
 
89943cf
72fd365
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f819bef
72fd365
 
 
 
f819bef
72fd365
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e6d384c
 
 
72fd365
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f819bef
72fd365
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f819bef
72fd365
 
f819bef
72fd365
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
import torch
import torch.nn as nn
from torch.utils.checkpoint import checkpoint
import math

from .diffusion import create_diffusion


class DiffLoss(nn.Module):
    """Diffusion Loss"""
    def __init__(self, target_channels, z_channels, depth, width, num_sampling_steps, grad_checkpointing=False):
        super(DiffLoss, self).__init__()
        self.in_channels = target_channels
        self.net = SimpleMLPAdaLN(
            in_channels=target_channels,
            model_channels=width,
            out_channels=target_channels * 2,  # for vlb loss
            z_channels=z_channels,
            num_res_blocks=depth,
            grad_checkpointing=grad_checkpointing
        )

        self.train_diffusion = create_diffusion(timestep_respacing="", noise_schedule="cosine")
        self.gen_diffusion = create_diffusion(timestep_respacing=num_sampling_steps, noise_schedule="cosine")

    def forward(self, target, z, mask=None):
        t = torch.randint(0, self.train_diffusion.num_timesteps, (target.shape[0],), device=target.device)
        model_kwargs = dict(c=z)
        loss_dict = self.train_diffusion.training_losses(self.net, target, t, model_kwargs)
        loss = loss_dict["loss"]
        if mask is not None:
            loss = (loss * mask).sum() / mask.sum()
        return loss.mean()

    def sample(self, z, temperature=1.0, cfg=1.0):
        # diffusion loss sampling
        if not cfg == 1.0:
            noise = torch.randn(z.shape[0] // 2, self.in_channels).cuda()
            noise = torch.cat([noise, noise], dim=0)
            model_kwargs = dict(c=z, cfg_scale=cfg)
            sample_fn = self.net.forward_with_cfg
        else:
            noise = torch.randn(z.shape[0], self.in_channels).cuda()
            model_kwargs = dict(c=z)
            sample_fn = self.net.forward

        sampled_token_latent = self.gen_diffusion.p_sample_loop(
            sample_fn, noise.shape, noise, clip_denoised=False, model_kwargs=model_kwargs, progress=False,
            temperature=temperature
        )

        return sampled_token_latent


def modulate(x, shift, scale):
    return x * (1 + scale) + shift


class TimestepEmbedder(nn.Module):
    """
    Embeds scalar timesteps into vector representations.
    """
    def __init__(self, hidden_size, frequency_embedding_size=256):
        super().__init__()
        self.mlp = nn.Sequential(
            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size, bias=True),
        )
        self.frequency_embedding_size = frequency_embedding_size

    @staticmethod
    def timestep_embedding(t, dim, max_period=10000):
        """
        Create sinusoidal timestep embeddings.
        :param t: a 1-D Tensor of N indices, one per batch element.
                          These may be fractional.
        :param dim: the dimension of the output.
        :param max_period: controls the minimum frequency of the embeddings.
        :return: an (N, D) Tensor of positional embeddings.
        """
        # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
        half = dim // 2
        freqs = torch.exp(
            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
        ).to(device=t.device)
        args = t[:, None].float() * freqs[None]
        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
        if dim % 2:
            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
        return embedding

    def forward(self, t):
        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
        t_emb = self.mlp(t_freq)
        return t_emb


class ResBlock(nn.Module):
    """
    A residual block that can optionally change the number of channels.
    :param channels: the number of input channels.
    """

    def __init__(
        self,
        channels
    ):
        super().__init__()
        self.channels = channels

        self.in_ln = nn.LayerNorm(channels, eps=1e-6)
        self.mlp = nn.Sequential(
            nn.Linear(channels, channels, bias=True),
            nn.SiLU(),
            nn.Linear(channels, channels, bias=True),
        )

        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(channels, 3 * channels, bias=True)
        )

    def forward(self, x, y):
        shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(y).chunk(3, dim=-1)
        h = modulate(self.in_ln(x), shift_mlp, scale_mlp)
        h = self.mlp(h)
        return x + gate_mlp * h


class FinalLayer(nn.Module):
    """
    The final layer adopted from DiT.
    """
    def __init__(self, model_channels, out_channels):
        super().__init__()
        self.norm_final = nn.LayerNorm(model_channels, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(model_channels, out_channels, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(model_channels, 2 * model_channels, bias=True)
        )

    def forward(self, x, c):
        shift, scale = self.adaLN_modulation(c).chunk(2, dim=-1)
        x = modulate(self.norm_final(x), shift, scale)
        x = self.linear(x)
        return x


class SimpleMLPAdaLN(nn.Module):
    """
    The MLP for Diffusion Loss.
    :param in_channels: channels in the input Tensor.
    :param model_channels: base channel count for the model.
    :param out_channels: channels in the output Tensor.
    :param z_channels: channels in the condition.
    :param num_res_blocks: number of residual blocks per downsample.
    """

    def __init__(
        self,
        in_channels,
        model_channels,
        out_channels,
        z_channels,
        num_res_blocks,
        grad_checkpointing=False
    ):
        super().__init__()

        self.in_channels = in_channels
        self.model_channels = model_channels
        self.out_channels = out_channels
        self.num_res_blocks = num_res_blocks
        self.grad_checkpointing = grad_checkpointing

        self.time_embed = TimestepEmbedder(model_channels)
        self.cond_embed = nn.Linear(z_channels, model_channels)

        self.input_proj = nn.Linear(in_channels, model_channels)

        res_blocks = []
        for i in range(num_res_blocks):
            res_blocks.append(ResBlock(
                model_channels,
            ))

        self.res_blocks = nn.ModuleList(res_blocks)
        self.final_layer = FinalLayer(model_channels, out_channels)

        self.initialize_weights()

    def initialize_weights(self):
        def _basic_init(module):
            if isinstance(module, nn.Linear):
                torch.nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)
        self.apply(_basic_init)

        # Initialize timestep embedding MLP
        nn.init.normal_(self.time_embed.mlp[0].weight, std=0.02)
        nn.init.normal_(self.time_embed.mlp[2].weight, std=0.02)

        # Zero-out adaLN modulation layers
        for block in self.res_blocks:
            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)

        # Zero-out output layers
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
        nn.init.constant_(self.final_layer.linear.weight, 0)
        nn.init.constant_(self.final_layer.linear.bias, 0)

    def forward(self, x, t, c):
        """
        Apply the model to an input batch.
        :param x: an [N x C] Tensor of inputs.
        :param t: a 1-D batch of timesteps.
        :param c: conditioning from AR transformer.
        :return: an [N x C] Tensor of outputs.
        """
        x = self.input_proj(x)
        t = self.time_embed(t)
        c = self.cond_embed(c)

        y = t + c

        if self.grad_checkpointing and not torch.jit.is_scripting():
            for block in self.res_blocks:
                x = checkpoint(block, x, y)
        else:
            for block in self.res_blocks:
                x = block(x, y)

        return self.final_layer(x, y)

    def forward_with_cfg(self, x, t, c, cfg_scale):
        half = x[: len(x) // 2]
        combined = torch.cat([half, half], dim=0)
        model_out = self.forward(combined, t, c)
        eps, rest = model_out[:, :self.in_channels], model_out[:, self.in_channels:]
        cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
        half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
        eps = torch.cat([half_eps, half_eps], dim=0)
        return torch.cat([eps, rest], dim=1)