models.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# References:
# GLIDE: https://github.com/openai/glide-text2im
# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py
# --------------------------------------------------------

import torch
import torch.nn as nn
import numpy as np
import math
from timm.models.vision_transformer import PatchEmbed, Attention, Mlp


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


#################################################################################
#               Embedding Layers for Timesteps and Class Labels                 #
#################################################################################

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 LabelEmbedder(nn.Module):
    """
    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
    """
    def __init__(self, num_classes, hidden_size, dropout_prob):
        super().__init__()
        use_cfg_embedding = dropout_prob > 0
        self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)
        self.num_classes = num_classes
        self.dropout_prob = dropout_prob

    def token_drop(self, labels, force_drop_ids=None):
        """
        Drops labels to enable classifier-free guidance.
        """
        if force_drop_ids is None:
            drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
        else:
            drop_ids = force_drop_ids == 1
        labels = torch.where(drop_ids, self.num_classes, labels)
        return labels

    def forward(self, labels, train, force_drop_ids=None):
        use_dropout = self.dropout_prob > 0
        if (train and use_dropout) or (force_drop_ids is not None):
            labels = self.token_drop(labels, force_drop_ids)
        embeddings = self.embedding_table(labels)
        return embeddings


#################################################################################
#                                 Core DiT Model                                #
#################################################################################

class CrossAttention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=True):
        super().__init__()
        assert dim % num_heads == 0
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.scale = self.head_dim ** -0.5

        self.q = nn.Linear(dim, dim, bias=qkv_bias)
        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
        self.proj = nn.Linear(dim, dim)

    def forward(self, x, context):
        b, n, c = x.shape
        _, m, _ = context.shape

        q = self.q(x).reshape(b, n, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        kv = self.kv(context).reshape(b, m, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
        k, v = kv[0], kv[1]

        attn = (q * self.scale) @ k.transpose(-2, -1)
        attn = attn.softmax(dim=-1)
        out = attn @ v
        out = out.transpose(1, 2).reshape(b, n, c)
        out = self.proj(out)
        return out


class DiTBlock(nn.Module):
    """
    A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning.
    """
    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, has_cross_attn=False, **block_kwargs):
        super().__init__()
        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, **block_kwargs)

        self.has_cross_attn = bool(has_cross_attn)
        if self.has_cross_attn:
            self.norm_cross = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
            self.cross_attn = CrossAttention(hidden_size, num_heads=num_heads, qkv_bias=True)
        else:
            self.norm_cross = None
            self.cross_attn = None

        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        approx_gelu = lambda: nn.GELU(approximate="tanh")
        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)

        if self.has_cross_attn:
            mod_out = 9 * hidden_size
        else:
            mod_out = 6 * hidden_size

        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, mod_out, bias=True)
        )

    def forward(self, x, c, cond_tokens=None):
        if self.has_cross_attn:
            shift_msa, scale_msa, gate_msa, shift_ca, scale_ca, gate_ca, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(9, dim=1)
            x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
            if cond_tokens is not None:
                x = x + gate_ca.unsqueeze(1) * self.cross_attn(modulate(self.norm_cross(x), shift_ca, scale_ca), cond_tokens)
            x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
            return x

        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)
        x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
        x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
        return x


class FinalLayer(nn.Module):
    """
    The final layer of DiT.
    """
    def __init__(self, hidden_size, patch_size, out_channels):
        super().__init__()
        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 2 * hidden_size, 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 DiT(nn.Module):
    """
    Diffusion model with a Transformer backbone.
    """
    def __init__(
        self,
        input_size=32,
        patch_size=2,
        in_channels=4,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4.0,
        class_dropout_prob=0.1,
        num_classes=1000,
        learn_sigma=True,
        conditioning_mode="none",
        x_channels=None,
        cond_channels=None,
        cond_dropout_prob=0.0,
        num_modalities=None,
        modality_dropout_prob=0.0,
        use_modality_embedding=False,
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.patch_size = patch_size
        self.num_heads = num_heads

        self.conditioning_mode = conditioning_mode
        self.x_channels = in_channels if x_channels is None else int(x_channels)
        self.cond_channels = self.x_channels if cond_channels is None else int(cond_channels)
        self.cond_dropout_prob = float(cond_dropout_prob)

        if conditioning_mode in ["concat", "hybrid"]:
            model_in_channels = self.x_channels + self.cond_channels
        else:
            model_in_channels = self.x_channels

        self.in_channels = self.x_channels
        self.out_channels = self.x_channels * 2 if learn_sigma else self.x_channels

        self.x_embedder = PatchEmbed(input_size, patch_size, model_in_channels, hidden_size, bias=True)
        self.t_embedder = TimestepEmbedder(hidden_size)
        self.y_embedder = None if (num_classes is None or int(num_classes) <= 0) else LabelEmbedder(int(num_classes), hidden_size, class_dropout_prob)

        self.use_modality_embedding = bool(use_modality_embedding)
        if self.use_modality_embedding:
            if num_modalities is None:
                raise ValueError("num_modalities must be provided when use_modality_embedding=True")
            self.src_mod_embedder = LabelEmbedder(int(num_modalities), hidden_size, modality_dropout_prob)
            self.tgt_mod_embedder = LabelEmbedder(int(num_modalities), hidden_size, modality_dropout_prob)
        else:
            self.src_mod_embedder = None
            self.tgt_mod_embedder = None

        if conditioning_mode in ["crossattn", "hybrid"]:
            self.cond_x_embedder = PatchEmbed(input_size, patch_size, self.cond_channels, hidden_size, bias=True)
        else:
            self.cond_x_embedder = None

        if conditioning_mode == "global":
            self.cond_embedder = nn.Sequential(
                nn.Linear(self.cond_channels, hidden_size, bias=True),
                nn.SiLU(),
                nn.Linear(hidden_size, hidden_size, bias=True),
            )
        else:
            self.cond_embedder = None

        num_patches = self.x_embedder.num_patches
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, hidden_size), requires_grad=False)

        has_cross_attn = conditioning_mode in ["crossattn", "hybrid"]
        self.blocks = nn.ModuleList([
            DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio, has_cross_attn=has_cross_attn) for _ in range(depth)
        ])
        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)
        self.initialize_weights()

    def initialize_weights(self):
        # Initialize transformer layers:
        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 (and freeze) pos_embed by sin-cos embedding:
        pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches ** 0.5))
        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))

        # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
        w = self.x_embedder.proj.weight.data
        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
        nn.init.constant_(self.x_embedder.proj.bias, 0)
        if self.cond_x_embedder is not None:
            w = self.cond_x_embedder.proj.weight.data
            nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
            nn.init.constant_(self.cond_x_embedder.proj.bias, 0)
        if self.cond_x_embedder is not None:
            w = self.cond_x_embedder.proj.weight.data
            nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
            nn.init.constant_(self.cond_x_embedder.proj.bias, 0)

        if self.y_embedder is not None:
            nn.init.normal_(self.y_embedder.embedding_table.weight, std=0.02)
        if self.src_mod_embedder is not None:
            nn.init.normal_(self.src_mod_embedder.embedding_table.weight, std=0.02)
        if self.tgt_mod_embedder is not None:
            nn.init.normal_(self.tgt_mod_embedder.embedding_table.weight, std=0.02)

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

        # Zero-out adaLN modulation layers in DiT blocks:
        for block in self.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 unpatchify(self, x):
        """
        x: (N, T, patch_size**2 * C)
        imgs: (N, H, W, C)
        """
        c = self.out_channels
        p = self.x_embedder.patch_size[0]
        h = w = int(x.shape[1] ** 0.5)
        assert h * w == x.shape[1]

        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
        x = torch.einsum('nhwpqc->nchpwq', x)
        imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p))
        return imgs

    def forward(
        self,
        x,
        t,
        y=None,
        cond=None,
        src_mod=None,
        tgt_mod=None,
        force_drop_ids=None,
        force_drop_cond=None,
        force_drop_src_mod=None,
        force_drop_tgt_mod=None,
    ):
        if force_drop_cond is None and self.training and self.cond_dropout_prob > 0:
            force_drop_cond = (torch.rand(x.shape[0], device=x.device) < self.cond_dropout_prob).to(torch.int64)

        if cond is None:
            cond_x = torch.zeros(x.shape[0], self.cond_channels, x.shape[2], x.shape[3], device=x.device, dtype=x.dtype)
        else:
            cond_x = cond

        if force_drop_cond is not None:
            drop_mask = (force_drop_cond.to(x.device).view(-1, 1, 1, 1) == 1)
            cond_x = torch.where(drop_mask, torch.zeros_like(cond_x), cond_x)

        x_in = x
        if self.conditioning_mode in ["concat", "hybrid"]:
            x_in = torch.cat([x_in, cond_x], dim=1)

        x_tok = self.x_embedder(x_in) + self.pos_embed
        t_emb = self.t_embedder(t)
        c = t_emb

        if self.y_embedder is not None:
            if y is None:
                raise ValueError("y must be provided when num_classes > 0")
            c = c + self.y_embedder(y, self.training, force_drop_ids=force_drop_ids)

        if self.src_mod_embedder is not None and src_mod is not None:
            c = c + self.src_mod_embedder(src_mod, self.training, force_drop_ids=force_drop_src_mod)
        if self.tgt_mod_embedder is not None and tgt_mod is not None:
            c = c + self.tgt_mod_embedder(tgt_mod, self.training, force_drop_ids=force_drop_tgt_mod)

        cond_tokens = None
        if self.cond_x_embedder is not None and cond is not None:
            cond_tokens = self.cond_x_embedder(cond) + self.pos_embed

        if self.cond_embedder is not None:
            if cond is None:
                cond_vec_in = torch.zeros(x.shape[0], self.cond_channels, device=x.device, dtype=x.dtype)
            else:
                cond_vec_in = cond.mean(dim=(2, 3))
            cond_vec = self.cond_embedder(cond_vec_in)
            if force_drop_cond is None and self.training and self.cond_dropout_prob > 0:
                force_drop_cond = (torch.rand(x.shape[0], device=x.device) < self.cond_dropout_prob).to(torch.int64)
            if force_drop_cond is not None:
                drop_mask = (force_drop_cond.to(cond_vec.device).view(-1, 1) == 1)
                cond_vec = torch.where(drop_mask, torch.zeros_like(cond_vec), cond_vec)
            c = c + cond_vec

        for block in self.blocks:
            x_tok = block(x_tok, c, cond_tokens=cond_tokens)
        x_tok = self.final_layer(x_tok, c)
        x_out = self.unpatchify(x_tok)
        return x_out

    def forward_with_cfg(self, x, t, cfg_scale, **model_kwargs):
        half = x[: len(x) // 2]
        combined = torch.cat([half, half], dim=0)

        model_kwargs = dict(model_kwargs)
        cond = model_kwargs.get("cond", None)
        if isinstance(cond, torch.Tensor):
            cond_half = cond[: len(cond) // 2]
            cond_null = torch.zeros_like(cond_half)
            model_kwargs["cond"] = torch.cat([cond_half, cond_null], dim=0)

        for k in ["y", "src_mod", "tgt_mod"]:
            v = model_kwargs.get(k, None)
            if isinstance(v, torch.Tensor):
                v_half = v[: len(v) // 2]
                model_kwargs[k] = torch.cat([v_half, v_half], dim=0)

        if self.conditioning_mode == "global":
            force_drop_cond = torch.zeros(combined.shape[0], device=combined.device, dtype=torch.int64)
            force_drop_cond[len(force_drop_cond) // 2 :] = 1
            model_kwargs["force_drop_cond"] = force_drop_cond

        model_out = self.forward(combined, t, **model_kwargs)
        eps, rest = model_out[:, : self.x_channels], model_out[:, self.x_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)


#################################################################################
#                   Sine/Cosine Positional Embedding Functions                  #
#################################################################################
# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py

def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
    """
    grid_size: int of the grid height and width
    return:
    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    grid_h = np.arange(grid_size, dtype=np.float32)
    grid_w = np.arange(grid_size, dtype=np.float32)
    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    grid = np.stack(grid, axis=0)

    grid = grid.reshape([2, 1, grid_size, grid_size])
    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
    if cls_token and extra_tokens > 0:
        pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
    return pos_embed


def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
    assert embed_dim % 2 == 0

    # use half of dimensions to encode grid_h
    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)

    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
    return emb


def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    embed_dim: output dimension for each position
    pos: a list of positions to be encoded: size (M,)
    out: (M, D)
    """
    assert embed_dim % 2 == 0
    omega = np.arange(embed_dim // 2, dtype=np.float64)
    omega /= embed_dim / 2.
    omega = 1. / 10000**omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product

    emb_sin = np.sin(out) # (M, D/2)
    emb_cos = np.cos(out) # (M, D/2)

    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
    return emb


#################################################################################
#                                   DiT Configs                                  #
#################################################################################

def DiT_XL_2(**kwargs):
    return DiT(depth=28, hidden_size=1152, patch_size=2, num_heads=16, **kwargs)

def DiT_XL_4(**kwargs):
    return DiT(depth=28, hidden_size=1152, patch_size=4, num_heads=16, **kwargs)

def DiT_XL_8(**kwargs):
    return DiT(depth=28, hidden_size=1152, patch_size=8, num_heads=16, **kwargs)

def DiT_L_2(**kwargs):
    return DiT(depth=24, hidden_size=1024, patch_size=2, num_heads=16, **kwargs)

def DiT_L_4(**kwargs):
    return DiT(depth=24, hidden_size=1024, patch_size=4, num_heads=16, **kwargs)

def DiT_L_8(**kwargs):
    return DiT(depth=24, hidden_size=1024, patch_size=8, num_heads=16, **kwargs)

def DiT_B_2(**kwargs):
    return DiT(depth=12, hidden_size=768, patch_size=2, num_heads=12, **kwargs)

def DiT_B_4(**kwargs):
    return DiT(depth=12, hidden_size=768, patch_size=4, num_heads=12, **kwargs)

def DiT_B_8(**kwargs):
    return DiT(depth=12, hidden_size=768, patch_size=8, num_heads=12, **kwargs)

def DiT_S_2(**kwargs):
    return DiT(depth=12, hidden_size=384, patch_size=2, num_heads=6, **kwargs)

def DiT_S_4(**kwargs):
    return DiT(depth=12, hidden_size=384, patch_size=4, num_heads=6, **kwargs)

def DiT_S_8(**kwargs):
    return DiT(depth=12, hidden_size=384, patch_size=8, num_heads=6, **kwargs)


DiT_models = {
    'DiT-XL/2': DiT_XL_2,  'DiT-XL/4': DiT_XL_4,  'DiT-XL/8': DiT_XL_8,
    'DiT-L/2':  DiT_L_2,   'DiT-L/4':  DiT_L_4,   'DiT-L/8':  DiT_L_8,
    'DiT-B/2':  DiT_B_2,   'DiT-B/4':  DiT_B_4,   'DiT-B/8':  DiT_B_8,
    'DiT-S/2':  DiT_S_2,   'DiT-S/4':  DiT_S_4,   'DiT-S/8':  DiT_S_8,
}