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)
    # print(x.shape, shift.shape, scale.shape)
    # return x * (1 + scale) + shift


#################################################################################
#               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 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, **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.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)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 6 * hidden_size, bias=True)
        )
    
    def forward(self, x, c):
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)

        B, T, _ = x.shape
        causal_mask = torch.tril(torch.ones(T, T, device=x.device, dtype=torch.bool))
        causal_mask = causal_mask.unsqueeze(0).unsqueeze(0)  # (B, T, T)
        #print(causal_mask.shape)  # 应为 [B, T, T]
        x_1=modulate(self.norm1(x), shift_msa, scale_msa)
        x = x + gate_msa.unsqueeze(1) * self.attn(
            modulate(self.norm1(x), shift_msa, scale_msa),
            attn_mask=causal_mask  # 👈 passed to timm.layers.Attention
        )

        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=4,
        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=False,
        max_gen_len=1000  # Maximum length of generated sequence
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.in_channels = in_channels
        self.out_channels = in_channels * 2 if learn_sigma else in_channels
        self.patch_size = patch_size
        self.num_heads = num_heads

        #self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
        #self.t_embedder = TimestepEmbedder(hidden_size)
        self.patch_proj = nn.Linear(in_channels * patch_size * patch_size, hidden_size)
        self.y_embedder = LabelEmbedder(num_classes, hidden_size, class_dropout_prob)
        self.num_patches = (input_size // patch_size) ** 2  # 使用整数除法  # Total number of patches (assumes square input)
        self.max_gen_len = max_gen_len
        self.hidden_size = hidden_size
        # Will use fixed sin-cos embedding:
        self.pos_embed = nn.Parameter(torch.zeros(1, max_gen_len*self.num_patches, hidden_size), requires_grad=False)

        self.blocks = nn.ModuleList([
            DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio) 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=build_2d_temporal_pos_embed(self.pos_embed.shape[-1], self.max_gen_len, self.num_patches)
        #pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.num_patches ** 0.5))
        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float())

        # 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)
        nn.init.xavier_uniform_(self.patch_proj.weight)
        nn.init.constant_(self.patch_proj.bias, 0)
        # Initialize label embedding table:
        nn.init.normal_(self.y_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)
    
    ###!!! todo
    def forward(self, x,y, return_last=False):
        """
        x: (B, LEN, C, P, P)
        y: (B,)
        """
        B, LEN, C, P, P2 = x.shape
        assert P == self.patch_size and P2 == self.patch_size, "Patch size mismatch"

        # Flatten (C,P,P) -> (C*P*P)
        x = x.view(B, LEN, -1)  # (B, LEN, C*P*P)
        x = self.patch_proj(x)  # (B, LEN, D)

        pos = self.pos_embed[:, :LEN, :]
        token_input = x + pos  # (B, LEN, D)
        #t=self.t_embedder(t)  # (B, D)
        y= self.y_embedder(y, self.training)
        #modified by xjw
        # y = y.unsqueeze(1).expand(-1, LEN, -1)  # (B, LEN, D)
        # y = y + pos  # 叠加位置编码
        cond=y
        for block in self.blocks:
            token_input = block(token_input, cond)  # (B, LEN, D)

        out_token = self.final_layer(token_input, cond)  # (B, LEN, C*P*P)
        out_token = out_token.view(B, LEN, self.out_channels, P, P)

        if return_last:
            return out_token[:, -1:]  # (B, 1, C, P, P)
        return out_token  # (B, LEN, C, P, P)
    
    def forward_with_cfg(self, x, y, is_training=True, cfg_scale=1.0):
        """
        x: (B, LEN, C, P, P)    # 你在推理里已经把 z 做了 to_patch_seq -> start_seq
        y: (B,)
        返回: (B, LEN or 1, C, P, P) 与 self.forward 对齐
        """
        B = x.shape[0]
        assert B % 2 == 0, "Batch size must be even for CFG (cond+uncond)"

        # 拆 cond / 构造 uncond
        half_x = x[: B // 2]       # (B/2, LEN, C, P, P)
        half_y = y[: B // 2]       # (B/2,)

        x_comb = torch.cat([half_x, half_x], dim=0)                # (B, LEN, C, P, P)
        # 无条件标签（通常用 num_classes 作为 null id；若你有别的定义，也可在外面传入）
        null_id = getattr(self.y_embedder, "num_classes", None)
        assert null_id is not None, "LabelEmbedder 里要有 num_classes 作为 null class id"
        y_null = torch.full_like(half_y, fill_value=null_id)
        y_comb = torch.cat([half_y, y_null], dim=0)                # (B,)

        # 走一次模型（和普通 forward 一样）
        out = self.forward(x_comb, y_comb, is_training=is_training)  # 训练返回 (B,LEN,C,P,P)；推理返回 (B,1,C,P,P)

        # 按 batch 维拆 cond/uncond
        cond_out, uncond_out = out.chunk(2, dim=0)  # 与 forward 的输出形状一致

        # 对**所有通道**做 CFG（不要再 :3）
        guided = uncond_out + cfg_scale * (cond_out - uncond_out)

        # 为了兼容你后续逻辑，很多人会把 guided 复制一份拼回 B 的 batch，
        # 也可以只返回 guided（B/2, ...）。这里保持 B 不变：
        out = torch.cat([guided, guided], dim=0)
        return out





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

def build_2d_temporal_pos_embed(D, T, patch_num):
    """
    D: embedding dim
    T: max time steps
    patch_num: number of patches per frame (must be square number)
    return: pos_embed of shape (1, T * patch_num, D)
    """
    grid_size = int(patch_num ** 0.5)
    assert grid_size * grid_size == patch_num, "patch_num must be a square number"

    # 1. Spatial (x, y) positional embedding: (patch_num, D)
    spatial = get_2d_sincos_pos_embed(embed_dim=D, grid_size=grid_size)  # (patch_num, D)
    spatial = np.tile(spatial, (T, 1))  # repeat T times → (T * patch_num, D)

    # 2. Temporal (t) positional embedding: (T, D)
    temporal = get_1d_sincos_pos_embed_from_grid(D, np.arange(T)) # (T, D)
    temporal = np.repeat(temporal, patch_num, axis=0)    # repeat each t → (T * patch_num, D)

    # 3. Add them
    pos_embed = spatial + temporal  # (T * patch_num, D)
    return pos_embed[None]  # (1, LEN, D)


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,
}