ActionDiffusion_WACV2025/model/dit.py

513 lines
20 KiB
Python
Raw Permalink Normal View History

2024-12-02 15:42:58 +01:00
# 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 Attention, Mlp # PatchEmbed
from torch.nn import TransformerEncoder, TransformerEncoderLayer
from .helpers import SinusoidalPosEmb
class PatchEmbed(nn.Module):
""" 2D Image to Patch Embedding
"""
def __init__(
self,
img_size=224,
patch_size=16,
in_chans=3,
embed_dim=768,
norm_layer=None,
flatten=True,
bias=True,
):
super().__init__()
img_size = (img_size, 1)
patch_size = (patch_size, 1)
self.img_size = img_size
self.patch_size = patch_size
self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
self.num_patches = self.grid_size[0] * self.grid_size[1]
self.flatten = flatten
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x):
B, C, H, W = x.shape
#_assert(H == self.img_size[0], f"Input image height ({H}) doesn't match model ({self.img_size[0]}).")
#_assert(W == self.img_size[1], f"Input image width ({W}) doesn't match model ({self.img_size[1]}).")
x = self.proj(x)
if self.flatten:
x = x.flatten(2).transpose(1, 2) # BCHW -> BNC
x = self.norm(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):
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))
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, 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 * 1 * 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,
in_channels=4,
hidden_size=384,
depth=12,
num_heads=6,
mlp_ratio=4.0,
class_dropout_prob=0.1,
num_classes=1000,
learn_sigma=False,
):
super().__init__()
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.num_heads = num_heads
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, 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
#print('pos_embed', pos_embed.shape, self.x_embedder.num_patches)
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: (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]'''
h = x.shape[1]
w = 1
#print(x.shape)
x = x.reshape(shape=(x.shape[0], h, w, p, 1, c))
x = torch.einsum('nhwpqc->nchpwq', x)
imgs = x.reshape(shape=(x.shape[0], c, h * p, w * 1))
return imgs
def forward(self, x, t):
"""
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
"""
#print('x', x.shape, 'x_embedder', self.x_embedder(x).shape, 'pos_embed', self.pos_embed.shape)
x = self.x_embedder(x) + 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 # (N, D)
c = t
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)
x = self.unpatchify(x) # (N, out_channels, H, W)
return x
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)
model_out = self.forward(combined, t, y)
# 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(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, 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
def get_emb(sin_inp):
"""
Gets a base embedding for one dimension with sin and cos intertwined
"""
emb = torch.stack((sin_inp.sin(), sin_inp.cos()), dim=-1)
return torch.flatten(emb, -2, -1)
class PositionalEncoding1D(nn.Module):
def __init__(self, channels, dtype_override=None):
"""
:param channels: The last dimension of the tensor you want to apply pos emb to.
:param dtype_override: If set, overrides the dtype of the output embedding.
"""
super(PositionalEncoding1D, self).__init__()
self.org_channels = channels
channels = int(np.ceil(channels / 2) * 2)
inv_freq = 1.0 / (10000 ** (torch.arange(0, channels, 2).float() / channels))
self.register_buffer("inv_freq", inv_freq)
self.register_buffer("cached_penc", None, persistent=False)
self.channels = channels
self.dtype_override = dtype_override
def forward(self, tensor):
"""
:param tensor: A 3d tensor of size (batch_size, ch, x)
:return: Positional Encoding Matrix of size (batch_size, ch, x)
"""
if len(tensor.shape) != 3:
raise RuntimeError("The input tensor has to be 3d!")
if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
return self.cached_penc
self.cached_penc = None
batch_size, orig_ch, x = tensor.shape
pos_x = torch.arange(x, device=tensor.device, dtype=self.inv_freq.dtype)
sin_inp_x = torch.einsum("i,j->ij", pos_x, self.inv_freq)
emb_x = get_emb(sin_inp_x)
#print('emb_x', emb_x.shape)
emb = torch.zeros(
(self.channels, x),
device=tensor.device,
dtype=(
self.dtype_override if self.dtype_override is not None else tensor.dtype
),
)
emb[:self.channels, :] = emb_x.permute(1,0)
self.cached_penc = emb[None, :orig_ch, :].repeat(batch_size, 1, 1)
return self.cached_penc
class TransformerModel(nn.Module):
def __init__(self, ntoken, ninp, nhead, nhid, nlayers=6, dropout=0.0):
super(TransformerModel, self).__init__()
self.model_type = 'Transformer'
self.pos_encoder = PositionalEncoding1D(ninp)
encoder_layers = TransformerEncoderLayer(ninp, nhead, nhid, dropout, batch_first=True)
self.transformer_encoder = TransformerEncoder(encoder_layers, nlayers)
#self.encoder = nn.Embedding(ntoken, ninp)
self.ninp = ninp
#self.decoder = nn.Linear(ninp, ntoken)
self.time_mlp = nn.Sequential( # should be removed for Noise and Deterministic Baselines
SinusoidalPosEmb(ninp),
nn.Linear(ninp-1, ninp * 4),
nn.Mish(),
nn.Linear(ninp * 4, ninp),
)
#self.init_weights()
def generate_square_subsequent_mask(self, sz):
mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))
return mask
'''def init_weights(self):
initrange = 0.1
self.encoder.weight.data.uniform_(-initrange, initrange)
self.decoder.bias.data.zero_()
self.decoder.weight.data.uniform_(-initrange, initrange)'''
def forward(self, src, t):
#print('self.ninp', self.ninp)
#src = self.encoder(src) * math.sqrt(self.ninp)
#print('src', src.shape)
#t = self.time_mlp(t).unsqueeze(1)
emb = self.pos_encoder(src)
#time = torch.cat((t,t,t), dim=1)
#print('time', time.shape)
output = self.transformer_encoder(src+emb)
#print('shape after transformer', output.shape)
#output = self.decoder(output)
return output
#################################################################################
# DiT Configs #
#################################################################################
def DiT_XL_2(**kwargs):
return DiT(depth=28, hidden_size=1152, patch_size=2, num_heads=16, **kwargs)
def DiT_XL_4(**kwargs):
return DiT(depth=28, hidden_size=1152, patch_size=4, num_heads=16, **kwargs)
def DiT_XL_8(**kwargs):
return DiT(depth=28, hidden_size=1152, patch_size=8, num_heads=16, **kwargs)
def DiT_L_2(**kwargs):
return DiT(depth=24, hidden_size=1024, patch_size=2, num_heads=16, **kwargs)
def DiT_L_4(**kwargs):
return DiT(depth=24, hidden_size=1024, patch_size=4, num_heads=16, **kwargs)
def DiT_L_8(**kwargs):
return DiT(depth=24, hidden_size=1024, patch_size=8, num_heads=16, **kwargs)
def DiT_B_2(**kwargs):
return DiT(depth=12, hidden_size=768, patch_size=2, num_heads=12, **kwargs)
def DiT_B_4(**kwargs):
return DiT(depth=12, hidden_size=768, patch_size=4, num_heads=12, **kwargs)
def DiT_B_8(**kwargs):
return DiT(depth=12, hidden_size=768, patch_size=8, num_heads=12, **kwargs)
def DiT_S_2(**kwargs):
return DiT(depth=12, hidden_size=384, patch_size=2, num_heads=6, **kwargs)
def DiT_S_4(**kwargs):
return DiT(depth=12, hidden_size=384, patch_size=4, num_heads=6, **kwargs)
def DiT_S_8(**kwargs):
return DiT(depth=12, hidden_size=384, patch_size=8, num_heads=6, **kwargs)
DiT_models = {
'DiT-XL/2': DiT_XL_2, 'DiT-XL/4': DiT_XL_4, 'DiT-XL/8': DiT_XL_8,
'DiT-L/2': DiT_L_2, 'DiT-L/4': DiT_L_4, 'DiT-L/8': DiT_L_8,
'DiT-B/2': DiT_B_2, 'DiT-B/4': DiT_B_4, 'DiT-B/8': DiT_B_8,
'DiT-S/2': DiT_S_2, 'DiT-S/4': DiT_S_4, 'DiT-S/8': DiT_S_8,
}