434 lines
17 KiB
Python
434 lines
17 KiB
Python
# 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
|
|
|
|
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 and len(mask) > 1:
|
|
# 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 |