wpeebles/DiT

1# Copyright (c) Meta Platforms, Inc. and affiliates.2# All rights reserved.3 4# This source code is licensed under the license found in the5# LICENSE file in the root directory of this source tree.6 7import torch8import torch.nn as nn9import numpy as np10import math11from timm.models.vision_transformer import PatchEmbed, Attention, Mlp12 13 14def modulate(x, shift, scale):15    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)16 17 18#################################################################################19#               Embedding Layers for Timesteps and Class Labels                 #20#################################################################################21 22class TimestepEmbedder(nn.Module):23    """24    Embeds scalar timesteps into vector representations.25    """26    def __init__(self, hidden_size, frequency_embedding_size=256):27        super().__init__()28        self.mlp = nn.Sequential(29            nn.Linear(frequency_embedding_size, hidden_size, bias=True),30            nn.SiLU(),31            nn.Linear(hidden_size, hidden_size, bias=True),32        )33        self.frequency_embedding_size = frequency_embedding_size34 35    @staticmethod36    def timestep_embedding(t, dim, max_period=10000):37        """38        Create sinusoidal timestep embeddings.39        :param t: a 1-D Tensor of N indices, one per batch element.40                          These may be fractional.41        :param dim: the dimension of the output.42        :param max_period: controls the minimum frequency of the embeddings.43        :return: an (N, D) Tensor of positional embeddings.44        """45        half = dim // 246        freqs = torch.exp(47            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half48        ).to(device=t.device)49        args = t[:, None].float() * freqs[None]50        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)51        if dim % 2:52            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)53        return embedding54 55    def forward(self, t):56        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)57        t_emb = self.mlp(t_freq)58        return t_emb59 60 61class LabelEmbedder(nn.Module):62    """63    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.64    """65    def __init__(self, num_classes, hidden_size, dropout_prob):66        super().__init__()67        use_cfg_embedding = dropout_prob > 068        self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)69        self.num_classes = num_classes70        self.dropout_prob = dropout_prob71 72    def token_drop(self, labels, force_drop_ids=None):73        """74        Drops labels to enable classifier-free guidance.75        """76        if force_drop_ids is None:77            drop_ids = torch.rand(labels.shape[0]) < self.dropout_prob78        else:79            drop_ids = force_drop_ids == 180        labels = torch.where(drop_ids, self.num_classes, labels)81        return labels82 83    def forward(self, labels, train, force_drop_ids=None):84        use_dropout = self.dropout_prob > 085        if (train and use_dropout) or (force_drop_ids is not None):86            labels = self.token_drop(labels, force_drop_ids)87        embeddings = self.embedding_table(labels)88        return embeddings89 90 91#################################################################################92#                                 Core DiT Model                                #93#################################################################################94 95class DiTBlock(nn.Module):96    """97    A DiT block with gated adaptive layer norm (adaLN) conditioning.98    """99    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, **block_kwargs):100        super().__init__()101        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)102        self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, **block_kwargs)103        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)104        mlp_hidden_dim = int(hidden_size * mlp_ratio)105        approx_gelu = lambda: nn.GELU(approximate="tanh")106        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)107        self.adaLN_modulation = nn.Sequential(108            nn.SiLU(),109            nn.Linear(hidden_size, 6 * hidden_size, bias=True)110        )111 112    def forward(self, x, c):113        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)114        x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))115        x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))116        return x117 118 119class FinalLayer(nn.Module):120    """121    The final layer of DiT.122    """123    def __init__(self, hidden_size, patch_size, out_channels):124        super().__init__()125        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)126        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)127        self.adaLN_modulation = nn.Sequential(128            nn.SiLU(),129            nn.Linear(hidden_size, 2 * hidden_size, bias=True)130        )131 132    def forward(self, x, c):133        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)134        x = modulate(self.norm_final(x), shift, scale)135        x = self.linear(x)136        return x137 138 139class DiT(nn.Module):140    """141    Diffusion model with a Transformer backbone.142    """143    def __init__(144        self,145        input_size=32,146        patch_size=2,147        in_channels=4,148        hidden_size=1152,149        depth=28,150        num_heads=16,151        mlp_ratio=4.0,152        class_dropout_prob=0.1,153        num_classes=1000,154        learn_sigma=True,155    ):156        super().__init__()157        self.learn_sigma = learn_sigma158        self.in_channels = in_channels159        self.out_channels = in_channels * 2 if learn_sigma else in_channels160        self.patch_size = patch_size161        self.num_heads = num_heads162 163        self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)164        self.t_embedder = TimestepEmbedder(hidden_size)165        self.y_embedder = LabelEmbedder(num_classes, hidden_size, class_dropout_prob)166        num_patches = self.x_embedder.num_patches167        # Will use fixed sin-cos embedding:168        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, hidden_size), requires_grad=False)169 170        self.blocks = nn.ModuleList([171            DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(depth)172        ])173        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)174        self.initialize_weights()175 176    def initialize_weights(self):177        # Initialize transformer layers:178        def _basic_init(module):179            if isinstance(module, nn.Linear):180                torch.nn.init.xavier_uniform_(module.weight)181                if module.bias is not None:182                    nn.init.constant_(module.bias, 0)183        self.apply(_basic_init)184 185        # Initialize (and freeze) pos_embed by sin-cos embedding186        pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches ** 0.5))187        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))188 189        # Initialize patch_embed like nn.Linear (instead of nn.Conv2d)190        w = self.x_embedder.proj.weight.data191        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))192 193        # Initialize label embedding table:194        nn.init.normal_(self.y_embedder.embedding_table.weight, std=0.02)195 196        # Initialize timestep embedding MLP:197        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)198        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)199 200        # Zero-out adaLN modulation layers in DiT blocks:201        for block in self.blocks:202            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)203            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)204 205        # Zero-out output layers:206        nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)207        nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)208        nn.init.constant_(self.final_layer.linear.weight, 0)209        nn.init.constant_(self.final_layer.linear.bias, 0)210 211    def unpatchify(self, x):212        """213        x: (N, T, patch_size**2 * C)214        imgs: (N, H, W, C)215        """216        c = self.out_channels217        p = self.x_embedder.patch_size[0]218        h = w = int(x.shape[1] ** 0.5)219        assert h * w == x.shape[1]220 221        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))222        x = torch.einsum('nhwpqc->nchpwq', x)223        imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p))224        return imgs225 226    def forward(self, x, t, y):227        """228        Forward pass of DiT.229        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)230        t: (N,) tensor of diffusion timesteps231        y: (N,) tensor of class labels232        """233        x = self.x_embedder(x) + self.pos_embed    # (N, T, D), where T = H * W / patch_size ** 2234        t = self.t_embedder(t)                     # (N, D)235        y = self.y_embedder(y, self.training)      # (N, D)236        c = t + y                                  # (N, D)237        for block in self.blocks:238            x = block(x, c)                        # (N, T, D)239        x = self.final_layer(x, c)                  # (N, T, patch_size ** 2 * out_channels)240        x = self.unpatchify(x)                     # (N, out_channels, H, W)241        return x242 243    def forward_with_cfg(self, x, t, y, cfg_scale):244        """245        Forward pass of DiT, but also batches the unconditional forward pass for classifier-free guidance.246        """247        # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb248        half = x[: len(x) // 2]249        combined = torch.cat([half, half], dim=0)250        model_out = self.forward(combined, t, y)251        eps, rest = model_out[:, :3], model_out[:, 3:]252        cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)253        half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)254        eps = torch.cat([half_eps, half_eps], dim=0)255        return torch.cat([eps, rest], dim=1)256 257 258#################################################################################259#                   Sine/Cosine Positional Embedding Functions                  #260#################################################################################261 262def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):263    """264    grid_size: int of the grid height and width265    return:266    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)267    """268    grid_h = np.arange(grid_size, dtype=np.float32)269    grid_w = np.arange(grid_size, dtype=np.float32)270    grid = np.meshgrid(grid_w, grid_h)  # here w goes first271    grid = np.stack(grid, axis=0)272 273    grid = grid.reshape([2, 1, grid_size, grid_size])274    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)275    if cls_token and extra_tokens > 0:276        pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)277    return pos_embed278 279 280def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):281    assert embed_dim % 2 == 0282 283    # use half of dimensions to encode grid_h284    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)285    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)286 287    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)288    return emb289 290 291def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):292    """293    embed_dim: output dimension for each position294    pos: a list of positions to be encoded: size (M,)295    out: (M, D)296    """297    assert embed_dim % 2 == 0298    omega = np.arange(embed_dim // 2, dtype=np.float32)299    omega /= embed_dim / 2.300    omega = 1. / 10000**omega  # (D/2,)301 302    pos = pos.reshape(-1)  # (M,)303    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product304 305    emb_sin = np.sin(out) # (M, D/2)306    emb_cos = np.cos(out) # (M, D/2)307 308    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)309    return emb310 311 312#################################################################################313#                                   DiT Configs                                  #314#################################################################################315 316def DiT_XL_2(**kwargs):317    return DiT(depth=28, hidden_size=1152, patch_size=2, num_heads=16, **kwargs)318 319def DiT_XL_4(**kwargs):320    return DiT(depth=28, hidden_size=1152, patch_size=4, num_heads=16, **kwargs)321 322def DiT_XL_8(**kwargs):323    return DiT(depth=28, hidden_size=1152, patch_size=8, num_heads=16, **kwargs)324 325def DiT_L_2(**kwargs):326    return DiT(depth=24, hidden_size=1024, patch_size=2, num_heads=16, **kwargs)327 328def DiT_L_4(**kwargs):329    return DiT(depth=24, hidden_size=1024, patch_size=4, num_heads=16, **kwargs)330 331def DiT_L_8(**kwargs):332    return DiT(depth=24, hidden_size=1024, patch_size=8, num_heads=16, **kwargs)333 334def DiT_B_2(**kwargs):335    return DiT(depth=12, hidden_size=768, patch_size=2, num_heads=12, **kwargs)336 337def DiT_B_4(**kwargs):338    return DiT(depth=12, hidden_size=768, patch_size=4, num_heads=12, **kwargs)339 340def DiT_B_8(**kwargs):341    return DiT(depth=12, hidden_size=768, patch_size=8, num_heads=12, **kwargs)342 343def DiT_S_2(**kwargs):344    return DiT(depth=12, hidden_size=384, patch_size=2, num_heads=6, **kwargs)345 346def DiT_S_4(**kwargs):347    return DiT(depth=12, hidden_size=384, patch_size=4, num_heads=6, **kwargs)348 349def DiT_S_8(**kwargs):350    return DiT(depth=12, hidden_size=384, patch_size=8, num_heads=6, **kwargs)351