osu_diffusion/utils/models.py

from functools import partial

import numpy as np
import torch
import torch.nn as nn

from .positional_embedding import position_sequence_embedding
from .positional_embedding import timestep_embedding


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

    def forward(self, t):
        t_freq = 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, class_size, hidden_size):
        super().__init__()
        self.class_embedding = nn.Sequential(
            nn.Linear(class_size, hidden_size),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size),
        )

    def forward(self, labels):
        embeddings = self.class_embedding(labels)
        return embeddings


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


class Mlp(nn.Module):
    """MLP as used in Vision Transformer, MLP-Mixer and related networks"""

    def __init__(
        self,
        in_features,
        hidden_features=None,
        out_features=None,
        act_layer=nn.GELU,
        norm_layer=None,
        bias=True,
        drop=0.0,
        use_conv=False,
    ):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        bias = (bias, bias)
        drop_probs = (drop, drop)
        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear

        self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
        self.act = act_layer()
        self.drop1 = nn.Dropout(drop_probs[0])
        self.norm = (
            norm_layer(hidden_features) if norm_layer is not None else nn.Identity()
        )
        self.fc2 = linear_layer(hidden_features, out_features, bias=bias[1])
        self.drop2 = nn.Dropout(drop_probs[1])

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop1(x)
        x = self.norm(x)
        x = self.fc2(x)
        x = self.drop2(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, mlp_ratio=4.0, **block_kwargs):
        super().__init__()
        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.attn = nn.MultiheadAttention(
            hidden_size,
            num_heads=num_heads,
            batch_first=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")
        # noinspection PyTypeChecker
        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, attn_mask=None, key_padding_mask=None):
        (
            shift_msa,
            scale_msa,
            gate_msa,
            shift_mlp,
            scale_mlp,
            gate_mlp,
        ) = self.adaLN_modulation(c).chunk(6, dim=1)
        modulated = modulate(self.norm1(x), shift_msa, scale_msa)
        x = (
            x
            + gate_msa.unsqueeze(1)
            * self.attn(
                modulated,
                modulated,
                modulated,
                need_weights=False,
                attn_mask=attn_mask,
            )[0]
        )
        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, out_channels):
        super().__init__()
        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_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 FirstLayer(nn.Module):
    """
    Embeds scalar positions into vector representation and concatenates context.
    """

    def __init__(
        self,
        hidden_size,
        context_size,
        in_channels,
        frequency_embedding_size=128,
    ):
        super().__init__()
        self.mlp = nn.Sequential(
            nn.Linear(
                in_channels * frequency_embedding_size
                + context_size,
                hidden_size,
                bias=True,
            ),
        )
        self.frequency_embedding_size = frequency_embedding_size

    def forward(self, x, c):
        x_freq = position_sequence_embedding(
            x * 512,
            self.frequency_embedding_size,
        )
        xc_emb = torch.concatenate((x_freq, c), -1)
        xc_emb = self.mlp(xc_emb)
        return xc_emb


class DiT(nn.Module):
    """
    Diffusion model with a Transformer backbone.
    """

    def __init__(
        self,
        in_channels=2,
        context_size=142,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4.0,
        class_size=256,
        learn_sigma=True,
    ):
        super().__init__()
        self.learn_sigma = learn_sigma
        self.in_channels = in_channels
        self.context_size = context_size
        self.out_channels = in_channels * 2 if learn_sigma else in_channels
        self.num_heads = num_heads

        self.context_embedder = FirstLayer(hidden_size, context_size, in_channels)
        self.t_embedder = TimestepEmbedder(hidden_size)
        self.y_embedder = LabelEmbedder(class_size, hidden_size)

        self.blocks = nn.ModuleList(
            [
                DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio)
                for _ in range(depth)
            ],
        )
        self.final_layer = FinalLayer(hidden_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 position embedding MLP:
        nn.init.normal_(self.context_embedder.mlp[0].weight, std=0.02)

        # Initialize label embedding table:
        nn.init.normal_(self.y_embedder.class_embedding[0].weight, std=0.02)
        nn.init.normal_(self.y_embedder.class_embedding[2].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 forward(self, x, t, c, y, attn_mask=None, key_padding_mask=None):
        """
        Forward pass of DiT.
        x: (N, C, T) tensor of sequence inputs
        t: (N) tensor of diffusion timesteps
        c: (N, E, T) tensor of sequence context
        y: (N, C) tensor of class labels
        """
        x = torch.swapaxes(x, 1, 2)  # (N, T, C)
        c = torch.swapaxes(c, 1, 2)  # (N, T, E)
        x = self.context_embedder(x, c)  # (N, T, D), where T = seq_len
        t = self.t_embedder(t)  # (N, D)
        y = self.y_embedder(y)  # (N, D)
        b = t + y  # (N, D)
        for block in self.blocks:
            x = block(x, b, attn_mask, key_padding_mask)  # (N, T, D)
        x = self.final_layer(x, b)  # (N, T, out_channels)
        x = torch.swapaxes(x, 1, 2)  # (N, out_channels, T)
        return x

    def forward_with_cfg(self, x, t, c, y, cfg_scale, attn_mask=None, key_padding_mask=None):
        """
        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)
        model_out = self.forward(combined, t, c, y, attn_mask, key_padding_mask)
        # For exact reproducibility reasons, we apply classifier-free guidance on only
        # three channels by default. The standard approach to cfg applies it to all channels.
        # This can be done by uncommenting the following line and commenting-out the line following that.
        eps, rest = model_out[:, : self.in_channels], model_out[:, self.in_channels:]
        # eps, rest = model_out[:, :3], model_out[:, 3:]
        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 / 2.0
    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


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


def DiT_XL(**kwargs: dict) -> DiT:
    return DiT(depth=28, hidden_size=1152, num_heads=16, **kwargs)


def DiT_L(**kwargs: dict) -> DiT:
    return DiT(depth=24, hidden_size=1024, num_heads=16, **kwargs)


def DiT_B(**kwargs: dict) -> DiT:
    return DiT(depth=12, hidden_size=768, num_heads=12, **kwargs)


def DiT_S(**kwargs: dict) -> DiT:
    return DiT(depth=12, hidden_size=384, num_heads=6, **kwargs)


DiT_models = {
    "DiT-XL": DiT_XL,
    "DiT-L": DiT_L,
    "DiT-B": DiT_B,
    "DiT-S": DiT_S,
}