microsoft--unilm/LatentLM/models/DiT.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 functools
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import math
from timm.models.vision_transformer import PatchEmbed
from .RMSNorm import RMSNorm


def modulate(x, shift, scale):
    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=False),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size, bias=False),
        )
        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(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.to(t.dtype)

    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


class SwiGLU(nn.Module):
    def __init__(
        self,
        embed_dim,
        ffn_dim,
        drop=0.,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.fc1 = nn.Linear(self.embed_dim, ffn_dim, bias=False)
        self.gate = nn.Linear(self.embed_dim, ffn_dim, bias=False)
        self.fc2 = nn.Linear(ffn_dim, self.embed_dim, bias=False)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x_shape = x.shape
        x = x.reshape(-1, x.size(-1))
        x = F.silu(self.fc1(x)) * self.gate(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        output = x.view(x_shape)
        return output
    
#################################################################################
#                                 Core DiT Model                                #
#################################################################################
class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, num_kv_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.dim = dim
        self.head_dim = dim // num_heads
        self.num_heads = num_heads
        self.num_kv_heads = num_kv_heads
        self.n_rep = num_heads // num_kv_heads
        head_dim = dim // num_heads
        self.scale = head_dim ** -0.5

        self.qkv = nn.Linear(dim, dim + 2 * self.num_kv_heads * self.head_dim, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim, bias=False)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, self.num_heads + 2 * self.num_kv_heads, self.head_dim)
        q, k, v = torch.split(qkv, [self.num_heads, self.num_kv_heads, self.num_kv_heads], dim=2)
        q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
        x = F.scaled_dot_product_attention(
            q, k, v,
            dropout_p=self.attn_drop.p if self.training else 0.,
        )

        x = x.transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x
    
class DiTBlock(nn.Module):
    """
    A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning.
    """
    def __init__(self, hidden_size, num_heads, num_kv_heads, mlp_ratio=4.0, proj_drop=0., attn_drop=0., **block_kwargs):
        super().__init__()
        self.norm1 = RMSNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.attn = Attention(hidden_size, num_heads=num_heads, num_kv_heads=num_kv_heads, qkv_bias=False, proj_drop=proj_drop, attn_drop=attn_drop, **block_kwargs)
        self.norm2 = RMSNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        mlp_hidden_dim = int(hidden_size * mlp_ratio * 2 / 3 / 64) * 64
        self.mlp = SwiGLU(hidden_size, mlp_hidden_dim, drop=proj_drop)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 6 * hidden_size, bias=False)
        )

    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)
        x = x + gate_msa * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
        x = x + gate_mlp * 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, output_size):
        super().__init__()
        self.norm_final = RMSNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_size, output_size, bias=False)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 2 * hidden_size, bias=False)
        )

    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=1,
        flatten_input=False,
        in_channels=4,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        num_kv_heads=None,
        mlp_ratio=4.0,
        class_dropout_prob=0.1,
        num_classes=1000,
        drop=0.0,
        norm_layer=None
    ):
        super().__init__()
        self.in_channels = in_channels
        self.out_channels = in_channels
        self.input_size = input_size
        self.patch_size = patch_size if not flatten_input else 1
        self.num_heads = num_heads
        self.num_kv_heads = num_kv_heads if num_kv_heads is not None else num_heads
        self.flatten_input_size = input_size * input_size // self.patch_size // self.patch_size
        self.flatten_input = flatten_input

        self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, strict_img_size=False, norm_layer=norm_layer) if not flatten_input else nn.Linear(in_channels, hidden_size, bias=False)
        self.t_embedder = TimestepEmbedder(hidden_size)
        self.y_embedder = LabelEmbedder(num_classes, hidden_size, class_dropout_prob)
        # Will use fixed sin-cos embedding:
        self.pos_embed = nn.Parameter(torch.zeros(1, self.flatten_input_size, hidden_size), requires_grad=False)

        self.blocks = nn.ModuleList([
            DiTBlock(hidden_size, self.num_heads, self.num_kv_heads, mlp_ratio=mlp_ratio, proj_drop=drop, attn_drop=drop) for _ in range(depth)
        ])
        self.final_layer = FinalLayer(hidden_size, self.patch_size * self.patch_size * self.out_channels)
        self.initialize_weights()
        
    @property
    def device(self):
        return next(self.parameters()).device

    @property
    def dtype(self):
        return next(self.parameters()).dtype

    def initialize_weights(self):
        # Initialize transformer layers:
        def _basic_init(module):
            if isinstance(module, nn.Linear):
                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)) if not self.flatten_input \
            else get_1d_sincos_pos_embed(self.pos_embed.shape[-1], self.flatten_input_size)
        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))

        # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
        if not self.flatten_input:
            nn.init.constant_(self.x_embedder.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)

        # Zero-out output layers:
        nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
        nn.init.constant_(self.final_layer.linear.weight, 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_noise, t, y, **kwargs):
        """
        Forward pass of DiT.
        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
        t: (N,) tensor of diffusion timesteps
        y: (N,) tensor of class labels
        """
        x = self.x_embedder(x_noise) + self.pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
        t = self.t_embedder(t)                   # (N, D)
        y = self.y_embedder(y, self.training)    # (N, D)
        c = (t + y).unsqueeze(1)                 # (N, D)
        for block in self.blocks:
            x = block(x, c)                      # (N, T, D)
        x = self.final_layer(x, c)               # (N, T, patch_size ** 2 * out_channels)
        if not self.flatten_input:
            x = self.unpatchify(x)               # (N, out_channels, H, W)
        return x
    
    def sample_with_cfg(self, y, cfg_scale, sample_func):
        bsz = y.shape[0]
        z = torch.randn(bsz, self.in_channels, self.input_size, self.input_size, device=self.device, dtype=self.dtype) if not self.flatten_input else torch.randn(bsz, self.flatten_input_size, self.in_channels, device=self.device, dtype=self.dtype)
        samples = sample_func(functools.partial(self.forward_with_cfg, y=y, cfg_scale=cfg_scale), z)
        samples, _ = samples.chunk(2, dim=0)
        return samples

    def forward_with_cfg(self, x, t, y, cfg_scale):
        """
        Forward pass of DiT, but also batches the unconditional forward pass for classifier-free guidance.
        """
        # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
        half = x[: len(x) // 2]
        combined = torch.cat([half, half], dim=0)
        eps = self.forward(combined, t, y)
        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 eps


#################################################################################
#                   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_1d_sincos_pos_embed(embed_dim, seq_len, cls_token=False, extra_tokens=0):
    """
    seq_len: int of the sequence length
    return:
    pos_embed: [seq_len, embed_dim] or [1+seq_len, embed_dim] (w/ or w/o cls_token)
    """
    pos = np.arange(seq_len, dtype=np.float32)
    pos_embed = get_1d_sincos_pos_embed_from_grid(embed_dim, pos)
    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_13B(**kwargs):
    return DiT(depth=40, hidden_size=5120, num_heads=40, **kwargs)

def DiT_7B(**kwargs):
    return DiT(depth=32, hidden_size=4096, num_heads=32, **kwargs)

def DiT_3B(**kwargs):
    return DiT(depth=32, hidden_size=2560, num_heads=20, **kwargs)

def DiT_XL(**kwargs):
    return DiT(depth=24, hidden_size=2048, num_heads=16, **kwargs)

def DiT_Large(**kwargs):
    return DiT(depth=24, hidden_size=1536, num_heads=12, **kwargs)

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

def DiT_Base(**kwargs):
    return DiT(depth=12, hidden_size=768, num_heads=12, **kwargs)

DiT_models = {
    'DiT-13B': DiT_13B, 'DiT-7B': DiT_7B, 'DiT-3B': DiT_3B, 'DiT-XL': DiT_XL, 'DiT-Large': DiT_Large,
    'DiT-Medium':  DiT_Medium, 'DiT-Base': DiT_Base
}