PixArt support
This commit is contained in:
City
@@ -0,0 +1,292 @@
|
||||
# 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 torch
|
||||
import torch.nn as nn
|
||||
import os
|
||||
import numpy as np
|
||||
from timm.models.layers import DropPath
|
||||
from timm.models.vision_transformer import PatchEmbed, Mlp
|
||||
|
||||
|
||||
from .utils import auto_grad_checkpoint, to_2tuple
|
||||
from .PixArt_blocks import t2i_modulate, CaptionEmbedder, WindowAttention, MultiHeadCrossAttention, T2IFinalLayer, TimestepEmbedder, LabelEmbedder, FinalLayer
|
||||
|
||||
|
||||
class PixArtBlock(nn.Module):
|
||||
"""
|
||||
A PixArt block with adaptive layer norm (adaLN-single) conditioning.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., window_size=0, input_size=None, use_rel_pos=False, **block_kwargs):
|
||||
super().__init__()
|
||||
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
self.attn = WindowAttention(hidden_size, num_heads=num_heads, qkv_bias=True,
|
||||
input_size=input_size if window_size == 0 else (window_size, window_size),
|
||||
use_rel_pos=use_rel_pos, **block_kwargs)
|
||||
self.cross_attn = MultiHeadCrossAttention(hidden_size, num_heads, **block_kwargs)
|
||||
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
# to be compatible with lower version pytorch
|
||||
approx_gelu = lambda: nn.GELU(approximate="tanh")
|
||||
self.mlp = Mlp(in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu, drop=0)
|
||||
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
self.window_size = window_size
|
||||
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 + self.drop_path(gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)).reshape(B, N, C))
|
||||
x = x + self.cross_attn(x, y, mask)
|
||||
x = x + self.drop_path(gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
|
||||
|
||||
return x
|
||||
|
||||
|
||||
#############################################################################
|
||||
# Core PixArt Model #
|
||||
#################################################################################
|
||||
class PixArt(nn.Module):
|
||||
"""
|
||||
Diffusion model with a Transformer backbone.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
input_size=32,
|
||||
patch_size=2,
|
||||
in_channels=4,
|
||||
hidden_size=1152,
|
||||
depth=28,
|
||||
num_heads=16,
|
||||
mlp_ratio=4.0,
|
||||
class_dropout_prob=0.1,
|
||||
pred_sigma=True,
|
||||
drop_path: float = 0.,
|
||||
window_size=0,
|
||||
window_block_indexes=[],
|
||||
use_rel_pos=False,
|
||||
caption_channels=4096,
|
||||
lewei_scale=1.0,
|
||||
config=None,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__()
|
||||
self.pred_sigma = pred_sigma
|
||||
self.in_channels = in_channels
|
||||
self.out_channels = in_channels * 2 if pred_sigma else in_channels
|
||||
self.patch_size = patch_size
|
||||
self.num_heads = num_heads
|
||||
self.lewei_scale = lewei_scale,
|
||||
self.dtype = torch.get_default_dtype()
|
||||
|
||||
self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
|
||||
self.t_embedder = TimestepEmbedder(hidden_size)
|
||||
num_patches = self.x_embedder.num_patches
|
||||
self.base_size = input_size // self.patch_size
|
||||
# Will use fixed sin-cos embedding:
|
||||
self.register_buffer("pos_embed", torch.zeros(1, num_patches, 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 = CaptionEmbedder(in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob, act_layer=approx_gelu)
|
||||
drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)] # stochastic depth decay rule
|
||||
self.blocks = nn.ModuleList([
|
||||
PixArtBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
|
||||
input_size=(input_size // patch_size, input_size // patch_size),
|
||||
window_size=window_size if i in window_block_indexes else 0,
|
||||
use_rel_pos=use_rel_pos if i in window_block_indexes else False)
|
||||
for i in range(depth)
|
||||
])
|
||||
self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels)
|
||||
|
||||
self.initialize_weights()
|
||||
|
||||
print(f'Warning: lewei scale: {self.lewei_scale}, base size: {self.base_size}')
|
||||
|
||||
def forward_raw(self, x, t, y, mask=None, data_info=None):
|
||||
"""
|
||||
Original forward pass of PixArt.
|
||||
x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
|
||||
t: (N,) tensor of diffusion timesteps
|
||||
y: (N, 1, 120, C) tensor of class labels
|
||||
"""
|
||||
self.h, self.w = x.shape[-2]//self.patch_size, x.shape[-1]//self.patch_size
|
||||
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)
|
||||
t0 = self.t_block(t)
|
||||
y = self.y_embedder(y, self.training) # (N, 1, L, D)
|
||||
if self.training:
|
||||
mask = mask.squeeze(1).squeeze(1)
|
||||
y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
|
||||
y_lens = mask.sum(dim=1).tolist()
|
||||
else:
|
||||
y_lens = [y.shape[2]] * y.shape[0]
|
||||
y = y.squeeze(1).view(1, -1, x.shape[-1])
|
||||
for block in self.blocks:
|
||||
x = auto_grad_checkpoint(block, x, y, t0, y_lens) # (N, T, D) #support grad checkpoint
|
||||
x = self.final_layer(x, t) # (N, T, patch_size ** 2 * out_channels)
|
||||
x = self.unpatchify(x) # (N, out_channels, H, W)
|
||||
return x
|
||||
|
||||
def forward(self, x, timesteps, context, y=None, **kwargs):
|
||||
"""
|
||||
Forward pass that adapts comfy input to original forward function
|
||||
x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
|
||||
timesteps: (N,) tensor of diffusion timesteps
|
||||
context: (N, 1, 120, C) conditioning
|
||||
y: extra conditioning.
|
||||
"""
|
||||
## Still accepts the input w/o that dim but returns garbage
|
||||
if len(context.shape) == 3:
|
||||
context = context.unsqueeze(1)
|
||||
|
||||
## run original forward pass
|
||||
out = self.forward_raw(
|
||||
x = x.to(self.dtype),
|
||||
t = timesteps.to(self.dtype),
|
||||
y = context.to(self.dtype),
|
||||
)
|
||||
|
||||
## only return EPS
|
||||
out = out.to(torch.float)
|
||||
eps, rest = out[:, :self.in_channels], out[:, self.in_channels:]
|
||||
return eps
|
||||
|
||||
def forward_with_dpmsolver(self, x, t, y, mask=None, **kwargs):
|
||||
"""
|
||||
dpm solver donnot need variance prediction
|
||||
"""
|
||||
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
|
||||
model_out = self.forward(x, t, y, mask)
|
||||
return model_out.chunk(2, dim=1)[0]
|
||||
|
||||
def forward_with_cfg(self, x, t, y, cfg_scale, **kwargs):
|
||||
"""
|
||||
Forward pass of PixArt, 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, kwargs)
|
||||
model_out = model_out['x'] if isinstance(model_out, dict) else model_out
|
||||
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)
|
||||
|
||||
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]
|
||||
|
||||
x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
|
||||
x = torch.einsum('nhwpqc->nchpwq', x)
|
||||
imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p))
|
||||
return imgs
|
||||
|
||||
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), lewei_scale=self.lewei_scale, base_size=self.base_size)
|
||||
self.pos_embed.data.copy_(torch.from_numpy(pos_embed).unsqueeze(0).to(self.dtype))
|
||||
|
||||
# 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]))
|
||||
|
||||
# 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)
|
||||
nn.init.normal_(self.t_block[1].weight, std=0.02)
|
||||
|
||||
# Initialize caption embedding MLP:
|
||||
nn.init.normal_(self.y_embedder.y_proj.fc1.weight, std=0.02)
|
||||
nn.init.normal_(self.y_embedder.y_proj.fc2.weight, std=0.02)
|
||||
|
||||
# Zero-out adaLN modulation layers in PixArt 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 get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0, lewei_scale=1.0, base_size=16):
|
||||
"""
|
||||
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_size[0]/base_size) / lewei_scale
|
||||
grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1]/base_size) / lewei_scale
|
||||
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], grid_size[0]])
|
||||
|
||||
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
|
||||
@@ -0,0 +1,295 @@
|
||||
# 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
|
||||
from tqdm import tqdm
|
||||
from timm.models.layers import DropPath
|
||||
from timm.models.vision_transformer import Mlp
|
||||
|
||||
from .utils import auto_grad_checkpoint, to_2tuple
|
||||
from .PixArt_blocks import t2i_modulate, CaptionEmbedder, WindowAttention, MultiHeadCrossAttention, T2IFinalLayer, TimestepEmbedder, SizeEmbedder
|
||||
from .PixArt import PixArt, get_2d_sincos_pos_embed
|
||||
|
||||
|
||||
class PatchEmbed(nn.Module):
|
||||
""" 2D Image to Patch Embedding
|
||||
"""
|
||||
def __init__(
|
||||
self,
|
||||
patch_size=16,
|
||||
in_chans=3,
|
||||
embed_dim=768,
|
||||
norm_layer=None,
|
||||
flatten=True,
|
||||
bias=True,
|
||||
):
|
||||
super().__init__()
|
||||
patch_size = to_2tuple(patch_size)
|
||||
self.patch_size = patch_size
|
||||
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):
|
||||
x = self.proj(x)
|
||||
if self.flatten:
|
||||
x = x.flatten(2).transpose(1, 2) # BCHW -> BNC
|
||||
x = self.norm(x)
|
||||
return x
|
||||
|
||||
|
||||
class PixArtMSBlock(nn.Module):
|
||||
"""
|
||||
A PixArt block with adaptive layer norm zero (adaLN-Zero) conditioning.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., window_size=0, input_size=None, use_rel_pos=False, **block_kwargs):
|
||||
super().__init__()
|
||||
self.hidden_size = hidden_size
|
||||
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
self.attn = WindowAttention(hidden_size, num_heads=num_heads, qkv_bias=True,
|
||||
input_size=input_size if window_size == 0 else (window_size, window_size),
|
||||
use_rel_pos=use_rel_pos, **block_kwargs)
|
||||
self.cross_attn = MultiHeadCrossAttention(hidden_size, num_heads, **block_kwargs)
|
||||
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
# to be compatible with lower version pytorch
|
||||
approx_gelu = lambda: nn.GELU(approximate="tanh")
|
||||
self.mlp = Mlp(in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu, drop=0)
|
||||
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
self.window_size = window_size
|
||||
self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
|
||||
|
||||
def forward(self, x, y, t, mask=None, **kwargs):
|
||||
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 + self.drop_path(gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)))
|
||||
x = x + self.cross_attn(x, y, mask)
|
||||
x = x + self.drop_path(gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
|
||||
|
||||
return x
|
||||
|
||||
|
||||
#############################################################################
|
||||
# Core PixArt Model #
|
||||
#################################################################################
|
||||
class PixArtMS(PixArt):
|
||||
"""
|
||||
Diffusion model with a Transformer backbone.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
input_size=32,
|
||||
patch_size=2,
|
||||
in_channels=4,
|
||||
hidden_size=1152,
|
||||
depth=28,
|
||||
num_heads=16,
|
||||
mlp_ratio=4.0,
|
||||
class_dropout_prob=0.1,
|
||||
learn_sigma=True,
|
||||
pred_sigma=True,
|
||||
drop_path: float = 0.,
|
||||
window_size=0,
|
||||
window_block_indexes=[],
|
||||
use_rel_pos=False,
|
||||
caption_channels=4096,
|
||||
lewei_scale=1.,
|
||||
config=None,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__(
|
||||
input_size=input_size,
|
||||
patch_size=patch_size,
|
||||
in_channels=in_channels,
|
||||
hidden_size=hidden_size,
|
||||
depth=depth,
|
||||
num_heads=num_heads,
|
||||
mlp_ratio=mlp_ratio,
|
||||
class_dropout_prob=class_dropout_prob,
|
||||
learn_sigma=learn_sigma,
|
||||
pred_sigma=pred_sigma,
|
||||
drop_path=drop_path,
|
||||
window_size=window_size,
|
||||
window_block_indexes=window_block_indexes,
|
||||
use_rel_pos=use_rel_pos,
|
||||
lewei_scale=lewei_scale,
|
||||
config=config,
|
||||
**kwargs,
|
||||
)
|
||||
self.dtype = torch.get_default_dtype()
|
||||
self.h = self.w = 0
|
||||
approx_gelu = lambda: nn.GELU(approximate="tanh")
|
||||
self.t_block = nn.Sequential(
|
||||
nn.SiLU(),
|
||||
nn.Linear(hidden_size, 6 * hidden_size, bias=True)
|
||||
)
|
||||
self.x_embedder = PatchEmbed(patch_size, in_channels, hidden_size, bias=True)
|
||||
self.y_embedder = CaptionEmbedder(in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob, act_layer=approx_gelu)
|
||||
self.csize_embedder = SizeEmbedder(hidden_size//3) # c_size embed
|
||||
self.ar_embedder = SizeEmbedder(hidden_size//3) # aspect ratio embed
|
||||
drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)] # stochastic depth decay rule
|
||||
self.blocks = nn.ModuleList([
|
||||
PixArtMSBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
|
||||
input_size=(input_size // patch_size, input_size // patch_size),
|
||||
window_size=window_size if i in window_block_indexes else 0,
|
||||
use_rel_pos=use_rel_pos if i in window_block_indexes else False)
|
||||
for i in range(depth)
|
||||
])
|
||||
self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels)
|
||||
self.training = False
|
||||
self.initialize()
|
||||
|
||||
def forward_raw(self, x, t, y, mask=None, data_info=None, **kwargs):
|
||||
"""
|
||||
Original forward pass of PixArt.
|
||||
x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
|
||||
t: (N,) tensor of diffusion timesteps
|
||||
y: (N, 1, 120, C) tensor of class labels
|
||||
"""
|
||||
bs = x.shape[0]
|
||||
c_size, ar = data_info['img_hw'], data_info['aspect_ratio']
|
||||
self.h, self.w = x.shape[-2]//self.patch_size, x.shape[-1]//self.patch_size
|
||||
pos_embed = torch.from_numpy(get_2d_sincos_pos_embed(self.pos_embed.shape[-1], (self.h, self.w), lewei_scale=self.lewei_scale, base_size=self.base_size)).unsqueeze(0).to(x.device).to(self.dtype)
|
||||
x = self.x_embedder(x) + pos_embed # (N, T, D), where T = H * W / patch_size ** 2
|
||||
t = self.t_embedder(t) # (N, D)
|
||||
csize = self.csize_embedder(c_size, bs) # (N, D)
|
||||
ar = self.ar_embedder(ar, bs) # (N, D)
|
||||
t = t + torch.cat([csize, ar], dim=1)
|
||||
t0 = self.t_block(t)
|
||||
y = self.y_embedder(y, self.training) # (N, D)
|
||||
if self.training:
|
||||
mask = mask.squeeze(1).squeeze(1)
|
||||
y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
|
||||
y_lens = mask.sum(dim=1).tolist()
|
||||
else:
|
||||
y_lens = [y.shape[2]] * y.shape[0]
|
||||
y = y.squeeze(1).view(1, -1, x.shape[-1])
|
||||
for block in self.blocks:
|
||||
x = auto_grad_checkpoint(block, x, y, t0, y_lens, **kwargs) # (N, T, D) #support grad checkpoint
|
||||
x = self.final_layer(x, t) # (N, T, patch_size ** 2 * out_channels)
|
||||
x = self.unpatchify(x) # (N, out_channels, H, W)
|
||||
return x
|
||||
|
||||
def forward(self, x, timesteps, context, y=None, **kwargs):
|
||||
"""
|
||||
Forward pass that adapts comfy input to original forward function
|
||||
x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
|
||||
timesteps: (N,) tensor of diffusion timesteps
|
||||
context: (N, 1, 120, C) conditioning
|
||||
y: extra conditioning.
|
||||
"""
|
||||
## aspect ratio based on the latent image shape.
|
||||
# Ideally, these should only be used as a fallback with the real ones
|
||||
# passed in `y` to allow different values to be used for cont/uncond.
|
||||
bs = x.shape[0]
|
||||
data_info = {
|
||||
"img_hw" : torch.tensor(
|
||||
[[x.shape[2]*8, x.shape[3]*8]],
|
||||
dtype=self.dtype,
|
||||
device=x.device
|
||||
).repeat(bs, 1),
|
||||
"aspect_ratio" : torch.tensor(
|
||||
[[x.shape[2]/x.shape[3]]],
|
||||
dtype=self.dtype,
|
||||
device=x.device
|
||||
).repeat(bs, 1),
|
||||
}
|
||||
|
||||
## Still accepts the input w/o that dim but returns garbage
|
||||
if len(context.shape) == 3:
|
||||
context = context.unsqueeze(1)
|
||||
|
||||
## run original forward pass
|
||||
out = self.forward_raw(
|
||||
x = x.to(self.dtype),
|
||||
t = timesteps.to(self.dtype),
|
||||
y = context.to(self.dtype),
|
||||
data_info=data_info,
|
||||
)
|
||||
|
||||
## only return EPS
|
||||
out = out.to(torch.float)
|
||||
eps, rest = out[:, :self.in_channels], out[:, self.in_channels:]
|
||||
return eps
|
||||
|
||||
def forward_with_dpmsolver(self, x, t, y, data_info, **kwargs):
|
||||
"""
|
||||
dpm solver donnot need variance prediction
|
||||
"""
|
||||
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
|
||||
model_out = self.forward_raw(x, t, y, data_info=data_info, **kwargs)
|
||||
return model_out.chunk(2, dim=1)[0]
|
||||
|
||||
def forward_with_cfg(self, x, t, y, cfg_scale, data_info, **kwargs):
|
||||
"""
|
||||
Forward pass of PixArt, 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_raw(combined, t, y, data_info=data_info)
|
||||
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)
|
||||
|
||||
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]
|
||||
assert self.h * self.w == x.shape[1]
|
||||
|
||||
x = x.reshape(shape=(x.shape[0], self.h, self.w, p, p, c))
|
||||
x = torch.einsum('nhwpqc->nchpwq', x)
|
||||
imgs = x.reshape(shape=(x.shape[0], c, self.h * p, self.w * p))
|
||||
return imgs
|
||||
|
||||
def initialize(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 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]))
|
||||
|
||||
# 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)
|
||||
nn.init.normal_(self.t_block[1].weight, std=0.02)
|
||||
nn.init.normal_(self.csize_embedder.mlp[0].weight, std=0.02)
|
||||
nn.init.normal_(self.csize_embedder.mlp[2].weight, std=0.02)
|
||||
nn.init.normal_(self.ar_embedder.mlp[0].weight, std=0.02)
|
||||
nn.init.normal_(self.ar_embedder.mlp[2].weight, std=0.02)
|
||||
|
||||
# Initialize caption embedding MLP:
|
||||
nn.init.normal_(self.y_embedder.y_proj.fc1.weight, std=0.02)
|
||||
nn.init.normal_(self.y_embedder.y_proj.fc2.weight, std=0.02)
|
||||
|
||||
# Zero-out adaLN modulation layers in PixArt 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)
|
||||
@@ -0,0 +1,435 @@
|
||||
# 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 torch
|
||||
import torch.nn as nn
|
||||
from timm.models.vision_transformer import Mlp, Attention as Attention_
|
||||
from einops import rearrange, repeat
|
||||
import xformers.ops
|
||||
|
||||
from .utils import add_decomposed_rel_pos
|
||||
from comfy import model_management
|
||||
|
||||
if model_management.xformers_enabled():
|
||||
import xformers
|
||||
import xformers.ops
|
||||
|
||||
def modulate(x, shift, scale):
|
||||
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
|
||||
|
||||
|
||||
def t2i_modulate(x, shift, scale):
|
||||
return x * (1 + scale) + shift
|
||||
|
||||
competent_attention_implementation = False
|
||||
|
||||
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/value: img tokens; key: condition; mask: if padding tokens
|
||||
B, N, C = x.shape
|
||||
|
||||
if model_management.xformers_enabled():
|
||||
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
|
||||
else:
|
||||
global competent_attention_implementation
|
||||
if not competent_attention_implementation:
|
||||
print("""\nYou should REALLY consider installing/enabling xformers.\nAlternatively, open up ExtraModels/PixArt/models/PixArt_blocks.py and\n- Fix the attention map on line 77 if you know how to\n- Add scaled_dot_product_attention on line 150\n- Send a PR and remove this message on line 32/66-69\n""")
|
||||
competent_attention_implementation = True
|
||||
|
||||
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)
|
||||
q, k, v = map(lambda t: t.permute(0, 2, 1, 3),(q, k, v),)
|
||||
|
||||
attn_mask = None
|
||||
if mask is not None:
|
||||
# This is probably wrong
|
||||
attn_mask = torch.zeros(
|
||||
[1, q.shape[1], q.shape[2], v.shape[2]],
|
||||
dtype=q.dtype,
|
||||
device=q.device
|
||||
)
|
||||
attn_mask[:, :, (q.shape[2]//2):, mask[0]:] = True
|
||||
attn_mask[:, :, :(q.shape[2]//2), :mask[1]] = True
|
||||
|
||||
x = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, dropout_p=self.attn_drop.p)
|
||||
x = x.permute(0, 2, 1, 3).contiguous()
|
||||
x = x.view(B, -1, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
|
||||
|
||||
class WindowAttention(Attention_):
|
||||
"""Multi-head Attention block with relative position embeddings."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
num_heads=8,
|
||||
qkv_bias=True,
|
||||
use_rel_pos=False,
|
||||
rel_pos_zero_init=True,
|
||||
input_size=None,
|
||||
**block_kwargs,
|
||||
):
|
||||
"""
|
||||
Args:
|
||||
dim (int): Number of input channels.
|
||||
num_heads (int): Number of attention heads.
|
||||
qkv_bias (bool: If True, add a learnable bias to query, key, value.
|
||||
rel_pos (bool): If True, add relative positional embeddings to the attention map.
|
||||
rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
|
||||
input_size (int or None): Input resolution for calculating the relative positional
|
||||
parameter size.
|
||||
"""
|
||||
super().__init__(dim, num_heads=num_heads, qkv_bias=qkv_bias, **block_kwargs)
|
||||
|
||||
self.use_rel_pos = use_rel_pos
|
||||
if self.use_rel_pos:
|
||||
# initialize relative positional embeddings
|
||||
self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, self.head_dim))
|
||||
self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, self.head_dim))
|
||||
|
||||
if not rel_pos_zero_init:
|
||||
nn.init.trunc_normal_(self.rel_pos_h, std=0.02)
|
||||
nn.init.trunc_normal_(self.rel_pos_w, std=0.02)
|
||||
|
||||
def forward(self, x, mask=None):
|
||||
B, N, C = x.shape # 2 4096 1152
|
||||
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
|
||||
|
||||
if model_management.xformers_enabled():
|
||||
q, k, v = qkv.unbind(2)
|
||||
|
||||
if getattr(self, 'fp32_attention', False):
|
||||
q, k, v = q.float(), k.float(), v.float()
|
||||
|
||||
attn_bias = None
|
||||
if mask is not None:
|
||||
attn_bias = torch.zeros([B * self.num_heads, q.shape[1], k.shape[1]], dtype=q.dtype, device=q.device)
|
||||
attn_bias.masked_fill_(mask.squeeze(1).repeat(self.num_heads, 1, 1) == 0, float('-inf'))
|
||||
# Switch between torch / xformers attention
|
||||
x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)
|
||||
x = x.view(B, N, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
else:
|
||||
q, k, v = qkv.permute(2, 0, 3, 1, 4).unbind(0)
|
||||
|
||||
q = q * self.scale
|
||||
attn = q @ k.transpose(-2, -1)
|
||||
attn = attn.softmax(dim=-1)
|
||||
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
|
||||
|
||||
|
||||
#################################################################################
|
||||
# AMP attention with fp32 softmax to fix loss NaN problem during training #
|
||||
#################################################################################
|
||||
class Attention(Attention_):
|
||||
def forward(self, x):
|
||||
B, N, C = x.shape
|
||||
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
||||
q, k, v = qkv.unbind(0) # make torchscript happy (cannot use tensor as tuple)
|
||||
use_fp32_attention = getattr(self, 'fp32_attention', False)
|
||||
if use_fp32_attention:
|
||||
q, k = q.float(), k.float()
|
||||
with torch.cuda.amp.autocast(enabled=not use_fp32_attention):
|
||||
attn = (q @ k.transpose(-2, -1)) * self.scale
|
||||
attn = attn.softmax(dim=-1)
|
||||
|
||||
attn = self.attn_drop(attn)
|
||||
|
||||
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
|
||||
|
||||
class FinalLayer(nn.Module):
|
||||
"""
|
||||
The final layer of PixArt.
|
||||
"""
|
||||
|
||||
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 T2IFinalLayer(nn.Module):
|
||||
"""
|
||||
The final layer of PixArt.
|
||||
"""
|
||||
|
||||
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.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 = t2i_modulate(self.norm_final(x), shift, scale)
|
||||
x = self.linear(x)
|
||||
return x
|
||||
|
||||
|
||||
class MaskFinalLayer(nn.Module):
|
||||
"""
|
||||
The final layer of PixArt.
|
||||
"""
|
||||
|
||||
def __init__(self, final_hidden_size, c_emb_size, patch_size, out_channels):
|
||||
super().__init__()
|
||||
self.norm_final = nn.LayerNorm(final_hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
self.linear = nn.Linear(final_hidden_size, patch_size * patch_size * out_channels, bias=True)
|
||||
self.adaLN_modulation = nn.Sequential(
|
||||
nn.SiLU(),
|
||||
nn.Linear(c_emb_size, 2 * final_hidden_size, bias=True)
|
||||
)
|
||||
def forward(self, x, t):
|
||||
shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
|
||||
x = modulate(self.norm_final(x), shift, scale)
|
||||
x = self.linear(x)
|
||||
return x
|
||||
|
||||
|
||||
class DecoderLayer(nn.Module):
|
||||
"""
|
||||
The final layer of PixArt.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, decoder_hidden_size):
|
||||
super().__init__()
|
||||
self.norm_decoder = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
self.linear = nn.Linear(hidden_size, decoder_hidden_size, bias=True)
|
||||
self.adaLN_modulation = nn.Sequential(
|
||||
nn.SiLU(),
|
||||
nn.Linear(hidden_size, 2 * hidden_size, bias=True)
|
||||
)
|
||||
def forward(self, x, t):
|
||||
shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
|
||||
x = modulate(self.norm_decoder(x), shift, scale)
|
||||
x = self.linear(x)
|
||||
return x
|
||||
|
||||
|
||||
#################################################################################
|
||||
# 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, device=t.device) / half)
|
||||
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.to(t.dtype))
|
||||
return t_emb
|
||||
|
||||
|
||||
class SizeEmbedder(TimestepEmbedder):
|
||||
"""
|
||||
Embeds scalar timesteps into vector representations.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, frequency_embedding_size=256):
|
||||
super().__init__(hidden_size=hidden_size, frequency_embedding_size=frequency_embedding_size)
|
||||
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
|
||||
self.outdim = hidden_size
|
||||
|
||||
def forward(self, s, bs):
|
||||
if s.ndim == 1:
|
||||
s = s[:, None]
|
||||
assert s.ndim == 2
|
||||
if s.shape[0] != bs:
|
||||
s = s.repeat(bs//s.shape[0], 1)
|
||||
assert s.shape[0] == bs
|
||||
b, dims = s.shape[0], s.shape[1]
|
||||
s = rearrange(s, "b d -> (b d)")
|
||||
s_freq = self.timestep_embedding(s, self.frequency_embedding_size)
|
||||
s_emb = self.mlp(s_freq.to(s.dtype))
|
||||
s_emb = rearrange(s_emb, "(b d) d2 -> b (d d2)", b=b, d=dims, d2=self.outdim)
|
||||
return s_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]).cuda() < 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 CaptionEmbedder(nn.Module):
|
||||
"""
|
||||
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
|
||||
"""
|
||||
|
||||
def __init__(self, in_channels, hidden_size, uncond_prob, act_layer=nn.GELU(approximate='tanh'), token_num=120):
|
||||
super().__init__()
|
||||
self.y_proj = Mlp(in_features=in_channels, hidden_features=hidden_size, out_features=hidden_size, act_layer=act_layer, drop=0)
|
||||
self.register_buffer("y_embedding", nn.Parameter(torch.randn(token_num, in_channels) / in_channels ** 0.5))
|
||||
self.uncond_prob = uncond_prob
|
||||
|
||||
def token_drop(self, caption, force_drop_ids=None):
|
||||
"""
|
||||
Drops labels to enable classifier-free guidance.
|
||||
"""
|
||||
if force_drop_ids is None:
|
||||
drop_ids = torch.rand(caption.shape[0]).cuda() < self.uncond_prob
|
||||
else:
|
||||
drop_ids = force_drop_ids == 1
|
||||
caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
|
||||
return caption
|
||||
|
||||
def forward(self, caption, train, force_drop_ids=None):
|
||||
if train:
|
||||
assert caption.shape[2:] == self.y_embedding.shape
|
||||
use_dropout = self.uncond_prob > 0
|
||||
if (train and use_dropout) or (force_drop_ids is not None):
|
||||
caption = self.token_drop(caption, force_drop_ids)
|
||||
caption = self.y_proj(caption)
|
||||
return caption
|
||||
|
||||
|
||||
class CaptionEmbedderDoubleBr(nn.Module):
|
||||
"""
|
||||
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
|
||||
"""
|
||||
|
||||
def __init__(self, in_channels, hidden_size, uncond_prob, act_layer=nn.GELU(approximate='tanh'), token_num=120):
|
||||
super().__init__()
|
||||
self.proj = Mlp(in_features=in_channels, hidden_features=hidden_size, out_features=hidden_size, act_layer=act_layer, drop=0)
|
||||
self.embedding = nn.Parameter(torch.randn(1, in_channels) / 10 ** 0.5)
|
||||
self.y_embedding = nn.Parameter(torch.randn(token_num, in_channels) / 10 ** 0.5)
|
||||
self.uncond_prob = uncond_prob
|
||||
|
||||
def token_drop(self, global_caption, caption, force_drop_ids=None):
|
||||
"""
|
||||
Drops labels to enable classifier-free guidance.
|
||||
"""
|
||||
if force_drop_ids is None:
|
||||
drop_ids = torch.rand(global_caption.shape[0]).cuda() < self.uncond_prob
|
||||
else:
|
||||
drop_ids = force_drop_ids == 1
|
||||
global_caption = torch.where(drop_ids[:, None], self.embedding, global_caption)
|
||||
caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
|
||||
return global_caption, caption
|
||||
|
||||
def forward(self, caption, train, force_drop_ids=None):
|
||||
assert caption.shape[2: ] == self.y_embedding.shape
|
||||
global_caption = caption.mean(dim=2).squeeze()
|
||||
use_dropout = self.uncond_prob > 0
|
||||
if (train and use_dropout) or (force_drop_ids is not None):
|
||||
global_caption, caption = self.token_drop(global_caption, caption, force_drop_ids)
|
||||
y_embed = self.proj(global_caption)
|
||||
return y_embed, caption
|
||||
@@ -0,0 +1,122 @@
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from torch.utils.checkpoint import checkpoint, checkpoint_sequential
|
||||
from collections.abc import Iterable
|
||||
from itertools import repeat
|
||||
|
||||
def _ntuple(n):
|
||||
def parse(x):
|
||||
if isinstance(x, Iterable) and not isinstance(x, str):
|
||||
return x
|
||||
return tuple(repeat(x, n))
|
||||
return parse
|
||||
|
||||
to_1tuple = _ntuple(1)
|
||||
to_2tuple = _ntuple(2)
|
||||
|
||||
def set_grad_checkpoint(model, use_fp32_attention=False, gc_step=1):
|
||||
assert isinstance(model, nn.Module)
|
||||
|
||||
def set_attr(module):
|
||||
module.grad_checkpointing = True
|
||||
module.fp32_attention = use_fp32_attention
|
||||
module.grad_checkpointing_step = gc_step
|
||||
model.apply(set_attr)
|
||||
|
||||
def auto_grad_checkpoint(module, *args, **kwargs):
|
||||
if getattr(module, 'grad_checkpointing', False):
|
||||
if isinstance(module, Iterable):
|
||||
gc_step = module[0].grad_checkpointing_step
|
||||
return checkpoint_sequential(module, gc_step, *args, **kwargs)
|
||||
else:
|
||||
return checkpoint(module, *args, **kwargs)
|
||||
return module(*args, **kwargs)
|
||||
|
||||
def checkpoint_sequential(functions, step, input, *args, **kwargs):
|
||||
|
||||
# Hack for keyword-only parameter in a python 2.7-compliant way
|
||||
preserve = kwargs.pop('preserve_rng_state', True)
|
||||
if kwargs:
|
||||
raise ValueError("Unexpected keyword arguments: " + ",".join(arg for arg in kwargs))
|
||||
|
||||
def run_function(start, end, functions):
|
||||
def forward(input):
|
||||
for j in range(start, end + 1):
|
||||
input = functions[j](input, *args)
|
||||
return input
|
||||
return forward
|
||||
|
||||
if isinstance(functions, torch.nn.Sequential):
|
||||
functions = list(functions.children())
|
||||
|
||||
# the last chunk has to be non-volatile
|
||||
end = -1
|
||||
segment = len(functions) // step
|
||||
for start in range(0, step * (segment - 1), step):
|
||||
end = start + step - 1
|
||||
input = checkpoint(run_function(start, end, functions), input, preserve_rng_state=preserve)
|
||||
return run_function(end + 1, len(functions) - 1, functions)(input)
|
||||
|
||||
def get_rel_pos(q_size, k_size, rel_pos):
|
||||
"""
|
||||
Get relative positional embeddings according to the relative positions of
|
||||
query and key sizes.
|
||||
Args:
|
||||
q_size (int): size of query q.
|
||||
k_size (int): size of key k.
|
||||
rel_pos (Tensor): relative position embeddings (L, C).
|
||||
|
||||
Returns:
|
||||
Extracted positional embeddings according to relative positions.
|
||||
"""
|
||||
max_rel_dist = int(2 * max(q_size, k_size) - 1)
|
||||
# Interpolate rel pos if needed.
|
||||
if rel_pos.shape[0] != max_rel_dist:
|
||||
# Interpolate rel pos.
|
||||
rel_pos_resized = F.interpolate(
|
||||
rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
|
||||
size=max_rel_dist,
|
||||
mode="linear",
|
||||
)
|
||||
rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
|
||||
else:
|
||||
rel_pos_resized = rel_pos
|
||||
|
||||
# Scale the coords with short length if shapes for q and k are different.
|
||||
q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
|
||||
k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
|
||||
relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
|
||||
|
||||
return rel_pos_resized[relative_coords.long()]
|
||||
|
||||
def add_decomposed_rel_pos(attn, q, rel_pos_h, rel_pos_w, q_size, k_size):
|
||||
"""
|
||||
Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
|
||||
https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py # noqa B950
|
||||
Args:
|
||||
attn (Tensor): attention map.
|
||||
q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
|
||||
rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
|
||||
rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
|
||||
q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
|
||||
k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
|
||||
|
||||
Returns:
|
||||
attn (Tensor): attention map with added relative positional embeddings.
|
||||
"""
|
||||
q_h, q_w = q_size
|
||||
k_h, k_w = k_size
|
||||
Rh = get_rel_pos(q_h, k_h, rel_pos_h)
|
||||
Rw = get_rel_pos(q_w, k_w, rel_pos_w)
|
||||
|
||||
B, _, dim = q.shape
|
||||
r_q = q.reshape(B, q_h, q_w, dim)
|
||||
rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
|
||||
rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
|
||||
|
||||
attn = (
|
||||
attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]
|
||||
).view(B, q_h * q_w, k_h * k_w)
|
||||
|
||||
return attn
|
||||
Reference in New Issue
Block a user