472 lines
19 KiB
Python
472 lines
19 KiB
Python
# 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
|
|
# --------------------------------------------------------
|
|
from typing import Final, Optional
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import numpy as np
|
|
import math
|
|
import torch.utils.checkpoint as cp
|
|
from einops import rearrange
|
|
from timm.layers import use_fused_attn, to_2tuple
|
|
# from timm.models.vision_transformer import PatchEmbed, Attention, Mlp
|
|
from timm.models.vision_transformer import PatchEmbed, Mlp
|
|
from torch.nn import functional as F
|
|
|
|
|
|
class Attention(nn.Module):
|
|
fused_attn: Final[bool]
|
|
|
|
def __init__(
|
|
self,
|
|
dim: int,
|
|
num_heads: int = 8,
|
|
qkv_bias: bool = False,
|
|
qk_norm: bool = False,
|
|
attn_drop: float = 0.,
|
|
proj_drop: float = 0.,
|
|
norm_layer: nn.Module = nn.LayerNorm,
|
|
) -> 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.fused_attn = use_fused_attn()
|
|
|
|
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
|
self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
|
|
self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
|
|
self.attn_drop = nn.Dropout(attn_drop)
|
|
self.proj = nn.Linear(dim, dim)
|
|
self.proj_drop = nn.Dropout(proj_drop)
|
|
|
|
def forward(self, x: torch.Tensor, attention_mask) -> 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)
|
|
q, k = self.q_norm(q), self.k_norm(k)
|
|
|
|
# if self.fused_attn:
|
|
# x = F.scaled_dot_product_attention(
|
|
# q, k, v,
|
|
# attn_mask=attention_mask,
|
|
# dropout_p=self.attn_drop.p if self.training else 0.,
|
|
# )
|
|
# else:
|
|
q = q * self.scale
|
|
attn = q @ k.transpose(-2, -1)
|
|
if attention_mask is not None:
|
|
attention_mask = attention_mask.flatten(1).unsqueeze(-1) # bs t h w -> bs thw 1
|
|
attention_mask = attention_mask @ attention_mask.transpose(1, 2) # bs thw 1 @ bs 1 thw = bs thw thw
|
|
attention_mask = attention_mask.unsqueeze(1)
|
|
assert attn.shape[0] == attention_mask.shape[0]
|
|
attention_mask = attention_mask.masked_fill(attention_mask == 0, torch.finfo(attn.dtype).min)
|
|
attn = attn + attention_mask
|
|
attn = attn.softmax(dim=-1)
|
|
if torch.any(torch.isnan(attn)):
|
|
print('torch.any(torch.isnan(attn))')
|
|
attn = attn.masked_fill(torch.isnan(attn), float(0.))
|
|
attn = self.attn_drop(attn)
|
|
x = attn @ v
|
|
|
|
x = x.transpose(1, 2).reshape(B, N, C)
|
|
x = self.proj(x)
|
|
x = self.proj_drop(x)
|
|
return x
|
|
|
|
|
|
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
|
|
|
|
@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
|
|
|
|
|
|
#################################################################################
|
|
# 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)
|
|
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)
|
|
)
|
|
|
|
def forward(self, x, c, attention_mask):
|
|
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), attention_mask)
|
|
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_t, 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_t * 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=32,
|
|
patch_size=2,
|
|
patch_size_t=1,
|
|
in_channels=256,
|
|
hidden_size=1152,
|
|
depth=28,
|
|
num_heads=16,
|
|
mlp_ratio=4.0,
|
|
class_dropout_prob=0.1,
|
|
num_classes=1000,
|
|
learn_sigma=True,
|
|
):
|
|
super().__init__()
|
|
self.gradient_checkpointing = False
|
|
|
|
self.learn_sigma = learn_sigma
|
|
self.in_channels = in_channels
|
|
self.out_channels = in_channels * 2 if learn_sigma else in_channels
|
|
self.patch_size = patch_size
|
|
self.patch_size_t = patch_size_t
|
|
self.num_heads = num_heads
|
|
self.hidden_size = hidden_size
|
|
# import ipdb;ipdb.set_trace()
|
|
self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
|
|
self.t_embedder = TimestepEmbedder(hidden_size)
|
|
self.y_embedder = LabelEmbedder(num_classes, hidden_size, class_dropout_prob)
|
|
num_patches = self.x_embedder.num_patches
|
|
# Will use fixed sin-cos embedding:
|
|
# 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.final_layer = FinalLayer(hidden_size, patch_size_t, 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], int(self.x_embedder.num_patches ** 0.5))
|
|
# 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 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)
|
|
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):
|
|
"""
|
|
x: (B, N, patch_size_t*patch_size**2 * C)
|
|
imgs: (B, C, T, H, W)
|
|
"""
|
|
c = self.out_channels
|
|
p = self.x_embedder.patch_size[0]
|
|
o = self.patch_size_t
|
|
assert self.t * self.h * self.w == x.shape[1]
|
|
|
|
x = x.reshape(shape=(x.shape[0], self.t, self.h, self.w, o, p, p, c))
|
|
x = torch.einsum('nthwopqc->nctohpwq', x)
|
|
imgs = x.reshape(shape=(x.shape[0], c, self.t * o, self.h * p, self.w * p))
|
|
return imgs
|
|
|
|
def ckpt_wrapper(self, module):
|
|
def ckpt_forward(*inputs):
|
|
outputs = module(*inputs)
|
|
return outputs
|
|
return ckpt_forward
|
|
|
|
def forward(self, x, t, y, attention_mask):
|
|
"""
|
|
Forward pass of DiT.
|
|
x: (B, D, T, H, W) tensor of spatial inputs (images or latent representations of images)
|
|
t: (B,) tensor of diffusion timesteps
|
|
y: (B,) tensor of class labels
|
|
"""
|
|
if x.ndim == 4:
|
|
raise NotImplementedError
|
|
B, D, T, H, W = x.shape
|
|
self.h = num_patches_height = H // self.patch_size
|
|
self.w = num_patches_width = W // self.patch_size
|
|
self.t = num_tubes_length = T // self.patch_size_t # 4 // 1
|
|
|
|
pos_embed = get_2d_sincos_pos_embed(self.hidden_size, [num_patches_height, num_patches_width])
|
|
pos_embed = nn.Parameter(torch.from_numpy(pos_embed).float().unsqueeze(0), requires_grad=False).to(x.device)
|
|
|
|
pos_embed_1d = get_2d_sincos_pos_embed(self.hidden_size, [num_tubes_length, 1])
|
|
pos_embed_1d = nn.Parameter(torch.from_numpy(pos_embed_1d).float().unsqueeze(0), requires_grad=False).to(x.device)
|
|
|
|
# print(num_patches_height, num_patches_width, x.shape, pos_embed.shape)
|
|
self.x_embedder.img_size = [H, W]
|
|
x = rearrange(x, 'b d t h w -> (b t) d h w')
|
|
x = self.x_embedder(x) + pos_embed # (BT, N, D), where N = H * W / patch_size ** 2
|
|
x = rearrange(x, '(b t) n d -> (b n) t d', t=self.t)
|
|
x = x + pos_embed_1d
|
|
x = rearrange(x, '(b n) t d -> b (t n) d', b=B)
|
|
|
|
t = self.t_embedder(t) # (B, D)
|
|
y = self.y_embedder(y, self.training) # (B, D)
|
|
c = t + y # (B, D)
|
|
for block in self.blocks: # (B, N, D)
|
|
if self.gradient_checkpointing and self.training:
|
|
x = torch.utils.checkpoint.checkpoint(self.ckpt_wrapper(block), x, c, attention_mask) # (B, N, D)
|
|
else:
|
|
x = block(x, c, attention_mask)
|
|
x = self.final_layer(x, c) # (B, N, patch_size_t * patch_size ** 2 * out_channels)
|
|
x = self.unpatchify(x) # (B, out_channels, T, H, W)
|
|
return x
|
|
|
|
def forward_with_cfg(self, x, t, y, cfg_scale, attention_mask):
|
|
"""
|
|
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, y, attention_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)
|
|
"""
|
|
if isinstance(grid_size, int):
|
|
grid_size = to_2tuple(grid_size)
|
|
grid_h = np.arange(grid_size[0], dtype=np.float32)
|
|
grid_w = np.arange(grid_size[1], 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[0], grid_size[1]])
|
|
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.
|
|
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_XL_122(**kwargs):
|
|
return DiT(depth=28, hidden_size=1152, patch_size_t=1, patch_size=2, num_heads=16, **kwargs)
|
|
|
|
def DiT_XL_144(**kwargs):
|
|
return DiT(depth=28, hidden_size=1152, patch_size_t=1, patch_size=4, num_heads=16, **kwargs)
|
|
|
|
def DiT_XL_188(**kwargs):
|
|
return DiT(depth=28, hidden_size=1152, patch_size_t=1, patch_size=8, num_heads=16, **kwargs)
|
|
|
|
def DiT_L_122(**kwargs):
|
|
return DiT(depth=24, hidden_size=1024, patch_size_t=1, patch_size=2, num_heads=16, **kwargs)
|
|
|
|
def DiT_L_144(**kwargs):
|
|
return DiT(depth=24, hidden_size=1024, patch_size_t=1, patch_size=4, num_heads=16, **kwargs)
|
|
|
|
def DiT_L_188(**kwargs):
|
|
return DiT(depth=24, hidden_size=1024, patch_size_t=1, patch_size=8, num_heads=16, **kwargs)
|
|
|
|
def DiT_B_122(**kwargs):
|
|
return DiT(depth=12, hidden_size=768, patch_size_t=1, patch_size=2, num_heads=12, **kwargs)
|
|
|
|
def DiT_B_144(**kwargs):
|
|
return DiT(depth=12, hidden_size=768, patch_size_t=1, patch_size=4, num_heads=12, **kwargs)
|
|
|
|
def DiT_B_188(**kwargs):
|
|
return DiT(depth=12, hidden_size=768, patch_size_t=1, patch_size=8, num_heads=12, **kwargs)
|
|
|
|
def DiT_S_122(**kwargs):
|
|
return DiT(depth=12, hidden_size=384, patch_size_t=1, patch_size=2, num_heads=6, **kwargs)
|
|
|
|
def DiT_S_144(**kwargs):
|
|
return DiT(depth=12, hidden_size=384, patch_size_t=1, patch_size=4, num_heads=6, **kwargs)
|
|
|
|
def DiT_S_188(**kwargs):
|
|
return DiT(depth=12, hidden_size=384, patch_size_t=1, patch_size=8, num_heads=6, **kwargs)
|
|
|
|
|
|
DiT_models = {
|
|
'DiT-XL/122': DiT_XL_122, 'DiT-XL/144': DiT_XL_144, 'DiT-XL/188': DiT_XL_188,
|
|
'DiT-L/122': DiT_L_122, 'DiT-L/144': DiT_L_144, 'DiT-L/188': DiT_L_188,
|
|
'DiT-B/122': DiT_B_122, 'DiT-B/144': DiT_B_144, 'DiT-B/188': DiT_B_188,
|
|
'DiT-S/122': DiT_S_122, 'DiT-S/144': DiT_S_144, 'DiT-S/188': DiT_S_188,
|
|
}
|