code/ppd/models/dit_video.py

import math
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from .mlp import Mlp
from timm.models.vision_transformer import Attention, PatchEmbed


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


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 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, qk_norm=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 = 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, pos=None):
        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_Video(nn.Module):
    """
    Cascade diffusion model with a transformer backbone.
    """

    def __init__(
        self,
        input_size=512,
        in_channels=4,
        out_channels=1,
        hidden_size=1024,
        depth=24,
        num_heads=16,
        patch_size=8,
        mlp_ratio=4.0,
    ):
        super().__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.num_heads = num_heads
        self.patch_size = patch_size
        self.depth = depth

        self.x_embedder = PatchEmbed(input_size, patch_size*2, in_channels, hidden_size, bias=True)
        self.t_embedder = TimestepEmbedder(hidden_size)
        num_patches = (512//16) ** 2
        self.pos_embed = nn.Parameter(
            torch.zeros(1, 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.temporal_blocks = nn.ModuleList(
            [DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(3)]
        )

        self.proj_fusion = nn.Sequential(
                nn.Linear(hidden_size+2048, hidden_size*4),
                nn.SiLU(),
                nn.Linear(hidden_size*4, hidden_size*4),
                nn.SiLU(),
                nn.Linear(hidden_size*4, hidden_size*4),
            )

        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], 
            (512//16, 512//16)
        )
        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)

        # 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, height, width):
        """
        x: (N, T, patch_size**2 * C)
        imgs: (N, H, W, C)
        """
        c = self.out_channels
        p = self.patch_size
        h = height // p
        w = width // p
        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, w * p))
        return imgs

    def forward(self, x=None, semantics=None, timestep=None, dropout=0.1):
        """
        Forward pass of SP-DiT.
        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
        t: (N,) tensor of diffusion timesteps
        """

        N, C, H, W = x.shape
        if len(timestep.shape) == 0:
            timestep = timestep[None]

        x = self.x_embedder(x) + self.pos_embed
        N, L, D = x.shape
        t = self.t_embedder(timestep)  # (N, D)

        # for block in self.blocks:
        j = 0
        for i, block in enumerate(self.blocks):
            if (i+1) > 21:
                x = x.permute(1, 0, 2)
                x = self.temporal_blocks[j](x, t)
                x = x.permute(1, 0, 2)
                j += 1

            if (i+1) > 18:
                ref0 = x[0].permute(1,0).reshape(1, D, H//8, W//8)
                ref0 = F.avg_pool2d(ref0, kernel_size=4)
                ref0 = ref0.flatten(2).permute(0, 2, 1)
                ref0 = ref0.repeat(N, 1, 1)

                ref1 = x[1].permute(1,0).reshape(1, D, H//8, W//8)
                ref1 = F.avg_pool2d(ref1, kernel_size=4)
                ref1 = ref1.flatten(2).permute(0, 2, 1)
                ref1 = ref1.repeat(N, 1, 1)

                ref2 = x[2].permute(1,0).reshape(1, D, H//8, W//8)
                ref2 = F.avg_pool2d(ref2, kernel_size=4)
                ref2 = ref2.flatten(2).permute(0, 2, 1)
                ref2 = ref2.repeat(N, 1, 1)

                x = torch.cat([x, ref0, ref1, ref2], dim=1)
            
            x = block(x, t)

            if (i+1) > 18:
                x = x[:, :4*L]
            if i == (self.depth//2)-1:
                semantics = F.normalize(semantics, dim=-1)
                x = self.proj_fusion(torch.cat([x, semantics], dim=-1))
                p = self.patch_size * 2
                x = x.reshape(shape=(N, H//p, W//p, 2, 2, D))
                x = torch.einsum("nhwpqc->nchpwq", x)
                x = x.reshape(shape=(N, D, (H//p)*2, (W//p)*2))
                x = x.flatten(2).transpose(1, 2)

        x = self.final_layer(x, t)
        x = self.unpatchify(x, height=H, width=W)
        return x

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]
    """
    if isinstance(grid_size, int):
        h, w = grid_size, grid_size
    else:
        h, w = grid_size
    grid_h = np.arange(h, dtype=np.float32)
    grid_w = np.arange(w, 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, h, w])
    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 / 2.0
    omega = 1.0 / 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