# 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 import xformers.ops 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_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 MultiHeadCrossAttention(nn.Module): def __init__(self, d_model, num_heads, attn_drop=0., proj_drop=0., **block_kwargs): super(MultiHeadCrossAttention, self).__init__() assert d_model % num_heads == 0, "d_model must be divisible by num_heads" self.d_model = d_model self.num_heads = num_heads self.head_dim = d_model // num_heads self.q_linear = nn.Linear(d_model, d_model) self.kv_linear = nn.Linear(d_model, d_model*2) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(d_model, d_model) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x, cond, mask=None): # query: img tokens; key/value: condition; mask: if padding tokens B, N, C = x.shape q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim) kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim) k, v = kv.unbind(2) attn_bias = None if mask is not None: attn_bias = xformers.ops.fmha.BlockDiagonalMask.from_seqlens([N] * B, mask) x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias) x = x.view(B, -1, C) x = self.proj(x) x = self.proj_drop(x) return x ################################################################################# # Core DiT Model # ################################################################################# class DiTBlock(nn.Module): """ A DiT block with cross attention for conditioning. Adapted from PixArt implementation. """ 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.cross_attn = MultiHeadCrossAttention(hidden_size, num_heads, **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) #) self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5) def forward(self, x, y, t, mask=None): B, N, C = x.shape shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None] + t.reshape(B, 6, -1)).chunk(6, dim=1) x = x + gate_msa * self.attn(modulate(self.norm1(x), shift_msa, scale_msa)).reshape(B, N, C) x = x + self.cross_attn(x, y, 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, 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.scale_shift_table = nn.Parameter(torch.randn(2, hidden_size) / hidden_size ** 0.5) self.out_channels = out_channels def forward(self, x, t): shift, scale = (self.scale_shift_table[None] + t[:, None]).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, in_channels=1, hidden_size=128, depth=12, num_heads=6, mlp_ratio=4.0, condition_channels=768, learn_sigma=True, ): super().__init__() self.learn_sigma = learn_sigma self.input_size = input_size self.in_channels = in_channels 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) approx_gelu = lambda: nn.GELU(approximate="tanh") self.t_block = nn.Sequential( nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True) ) self.y_embedder = Mlp(in_features=condition_channels, hidden_features=hidden_size, out_features=hidden_size, act_layer=approx_gelu, drop=0) # Will use fixed sin-cos embedding: self.pos_embed = nn.Parameter(torch.zeros(1, input_size, 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, 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: grid_1d = np.arange(self.input_size, dtype=np.float32) pos_embed = get_1d_sincos_pos_embed_from_grid(self.pos_embed.shape[-1], grid_1d) self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0)) # Initialize patch_embed like nn.Linear (instead of nn.Conv2d): nn.init.xavier_uniform_(self.x_embedder.weight) nn.init.constant_(self.x_embedder.bias, 0) # Initialize label embedding table: nn.init.normal_(self.y_embedder.fc1.weight, std=0.02) nn.init.normal_(self.y_embedder.fc2.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.cross_attn.proj.weight, 0) nn.init.constant_(block.cross_attn.proj.bias, 0) # Zero-out output layers: nn.init.constant_(self.final_layer.linear.weight, 0) nn.init.constant_(self.final_layer.linear.bias, 0) def ckpt_wrapper(self, module): def ckpt_forward(*inputs): outputs = module(*inputs) return outputs return ckpt_forward def forward(self, x, t, y): """ Forward pass of DiT. x: (N, 1, T) tensor of PCG params t: (N,) tensor of diffusion timesteps y: (N, 1, C) or (N, M, C) tensor of condition image features """ x = x.permute(0, 2, 1) x = self.x_embedder(x) + self.pos_embed # (N, T, D), where T is the input token number (params number) t = self.t_embedder(t) # (N, D) t0 = self.t_block(t) y = self.y_embedder(y) # (N, M, D) # mask for batch cross-attention y_lens = [y.shape[1]] * y.shape[0] y = y.view(1, -1, x.shape[-1]) for block in self.blocks: x = torch.utils.checkpoint.checkpoint(self.ckpt_wrapper(block), x, y, t0, y_lens) # (N, T, D) x = self.final_layer(x, t) # (N, T, out_channels) return x.permute(0, 2, 1) ################################################################################# # Sine/Cosine Positional Embedding Functions # ################################################################################# # https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py 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_S(**kwargs): # 39M return DiT(depth=16, hidden_size=384, num_heads=6, **kwargs) def DiT_mini(**kwargs): # 7.6M return DiT(depth=12, hidden_size=192, num_heads=6, **kwargs) def DiT_tiny(**kwargs): # 1.3M return DiT(depth=8, hidden_size=96, num_heads=6, **kwargs) DiT_models = { 'DiT_S': DiT_S, 'DiT_mini': DiT_mini, 'DiT_tiny': DiT_tiny }