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 math

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
# from torch.nn.attention.flex_attention import BlockMask, flex_attention


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.SiLU,
        norm_layer=None,
        bias=True,
    ):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        linear_layer = nn.Linear

        self.fc1 = linear_layer(in_features, hidden_features, bias=bias)
        self.act = act_layer()
        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)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.norm(x)
        x = self.fc2(x)
        return x


class RMSNorm(torch.nn.Module):
    def __init__(self, dim: int, eps: float = 1e-5):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        output = self._norm(x.float()).type_as(x)
        return output * self.weight


class Attention(nn.Module):
    def __init__(
        self,
        dim: int,
        num_heads: int = 8,
        qkv_bias: bool = False,
    ) -> None:
        super().__init__()
        assert dim % num_heads == 0, "dim should be divisible by num_heads"
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.scale = self.head_dim**-0.5

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.proj = nn.Linear(dim, dim)
        self.q_norm = RMSNorm(self.head_dim, eps=1e-5)
        self.k_norm = RMSNorm(self.head_dim, eps=1e-5)

    def forward(self, x: torch.Tensor, attn_mask=None) -> torch.Tensor:
        B, N, C = x.shape
        qkv = (
            self.qkv(x)
            .reshape(B, N, 3, self.num_heads, self.head_dim)
            .permute(2, 0, 3, 1, 4)
        )
        q, k, v = qkv.unbind(0)

        if isinstance(attn_mask, torch.Tensor) or attn_mask is None:
            q = self.q_norm(q)
            k = self.k_norm(k)
            # v = v
            x = F.scaled_dot_product_attention(
                q,
                k,
                v,
                attn_mask=attn_mask,
            )
        elif isinstance(attn_mask, BlockMask):
            with torch.autocast(enabled=False, device_type="cuda"):
                q = self.q_norm(q).half()
                k = self.k_norm(k).half()
                v = v.half()
                x = flex_attention(q, k, v, block_mask=attn_mask)
        x = x.transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        return x


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=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 last class as unconditional value
        use_cfg_embedding = dropout_prob > 0
        if use_cfg_embedding:
            self.unconditional_value = num_classes - 1
        self.speaker_id_table = nn.Embedding(num_classes, hidden_size)
        self.phone_table = nn.Embedding(num_classes, hidden_size)
        self.phone_kind_table = nn.Embedding(num_classes, hidden_size)
        self.num_classes = num_classes
        self.dropout_prob = dropout_prob

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

    def forward(self, speaker_id, phone, phone_kind, train, force_drop_ids=None):
        use_dropout = self.dropout_prob > 0
        if (train and use_dropout) or (force_drop_ids is not None):
            speaker_id, phone, phone_kind = self.token_drop(
                speaker_id, phone, phone_kind, force_drop_ids
            )
        speaker_id_embeddings = self.speaker_id_table(speaker_id)
        phone_embeddings = self.phone_table(phone)
        phone_kind_embeddings = self.phone_kind_table(phone_kind)
        return speaker_id_embeddings, phone_embeddings, phone_kind_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)
        self.mlp = Mlp(
            in_features=hidden_size,
            hidden_features=mlp_hidden_dim,
            act_layer=nn.SiLU,
        )
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 6 * hidden_size, bias=True),
        )

    def forward(self, x, c, attn_mask=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 * self.attn(
            modulate(self.norm1(x), shift_msa, scale_msa), attn_mask=attn_mask
        )
        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, 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=256,
        in_channels=1024,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4.0,
        class_dropout_prob=0.1,
        learn_sigma=True,
        embedding_vocab_size=1024,
    ):
        super().__init__()
        self.input_size = input_size
        self.learn_sigma = learn_sigma
        self.in_channels = in_channels
        self.hidden_size = hidden_size
        self.out_channels = in_channels * 2 if learn_sigma else in_channels
        self.num_heads = num_heads

        self.x_embedder = nn.Linear(in_channels, hidden_size, bias=True)
        self.t_embedder = TimestepEmbedder(hidden_size)
        self.y_embedder = LabelEmbedder(
            embedding_vocab_size, hidden_size, class_dropout_prob
        )
        # Will use fixed sin-cos embedding:
        self.register_buffer("pos_embed", torch.zeros(1, self.input_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, 1, 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_1d_sincos_pos_embed(self.pos_embed.shape[-1], self.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):
        w = self.x_embedder.weight.data
        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
        nn.init.constant_(self.x_embedder.bias, 0)

        # Initialize label embedding table:
        scale = 1.0 / math.sqrt(self.hidden_size)
        nn.init.trunc_normal_(self.y_embedder.speaker_id_table.weight, std=scale)
        nn.init.trunc_normal_(self.y_embedder.phone_table.weight, std=scale)

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

        # 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, speaker_id, phone, phone_kind, attn_mask=None):
        """
        Forward pass of DiT.
        x: (N, C, L) tensor of spatial inputs
        t: (N,) tensor of diffusion timesteps
        speaker_id: (N,) tensor of speaker IDs
        phone: (N, L) tensor of phone labels
        phone_kind: (N, L) tensor of phone kinds
        """
        # (N, D), (N, L, D)
        speaker_id_embedding, phone_embedding, phone_kind_embedding = self.y_embedder(
            speaker_id, phone, phone_kind, self.training
        )
        t = self.t_embedder(t)  # (N, D)
        c = t  # (N, D)
        c = (
            c[:, None, :]
            + speaker_id_embedding
            + phone_embedding
            + phone_kind_embedding
        )  # (N, L, D)

        x = x.transpose(-1, -2)  # Swap last two dimensions
        x = self.x_embedder(x) + self.pos_embed[:, : x.shape[1], :]  # (N, L, D)
        for block in self.blocks:
            x = block(x, c, attn_mask=attn_mask)  # (N, L, D)
        x = self.final_layer(x, c)  # (N, L, 2 * out_channels)
        x = x.transpose(-1, -2)  # Swap last two dimensions
        return x

    def forward_with_cfg(
        self, x, t, speaker_id, phone, phone_kind, cfg_scale, attn_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, speaker_id, phone, phone_kind, attn_mask=attn_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_1d_sincos_pos_embed(embed_dim, length, cls_token=False, extra_tokens=0):
    """
    length: int of the length
    return:
    pos_embed: [length, embed_dim] or [1+length, embed_dim] (w/ or w/o cls_token)
    """
    grid = np.arange(length, dtype=np.float32)
    pos_embed = get_1d_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_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):
    return DiT(depth=28, hidden_size=1152, num_heads=16, **kwargs)


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


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


def DiT_S(**kwargs):
    return DiT(depth=6, hidden_size=256, num_heads=4, **kwargs)


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