Spaces:
Sleeping
Sleeping
File size: 6,546 Bytes
5af269e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 |
import torch
import torch.nn as nn
from collections import OrderedDict
from lvdm.basics import (
zero_module,
conv_nd,
avg_pool_nd
)
from einops import rearrange
from lvdm.modules.attention import register_attn_processor, set_attn_processor, DualCrossAttnProcessor, get_attn_processor
from lvdm.modules.attention import DualCrossAttnProcessorAS
from utils.utils import instantiate_from_config
from lvdm.modules.encoders.arch_transformer import Transformer
class StyleTransformer(nn.Module):
def __init__(self, in_dim=1024, out_dim=1024, num_heads=8, num_tokens=4, n_layers=2):
super().__init__()
scale = in_dim ** -0.5
self.num_tokens = num_tokens
self.style_emb = nn.Parameter(torch.randn(1, num_tokens, in_dim) * scale)
self.transformer_blocks = Transformer(
width=in_dim,
layers=n_layers,
heads=num_heads,
)
self.ln1 = nn.LayerNorm(in_dim)
self.ln2 = nn.LayerNorm(in_dim)
self.proj = nn.Parameter(torch.randn(in_dim, out_dim) * scale)
def forward(self, x):
style_emb = self.style_emb.repeat(x.shape[0], 1, 1)
x = torch.cat([style_emb, x], dim=1)
# x = torch.cat([x, style_emb], dim=1)
x = self.ln1(x)
x = x.permute(1, 0, 2)
x = self.transformer_blocks(x)
x = x.permute(1, 0, 2)
x = self.ln2(x[:, :self.num_tokens, :])
x = x @ self.proj
return x
class ScaleEncoder(nn.Module):
def __init__(self, in_dim=1024, out_dim=1, num_heads=8, num_tokens=16, n_layers=2):
super().__init__()
scale = in_dim ** -0.5
self.num_tokens = num_tokens
self.scale_emb = nn.Parameter(torch.randn(1, num_tokens, in_dim) * scale)
self.transformer_blocks = Transformer(
width=in_dim,
layers=n_layers,
heads=num_heads,
)
self.ln1 = nn.LayerNorm(in_dim)
self.ln2 = nn.LayerNorm(in_dim)
self.out = nn.Sequential(
nn.Linear(in_dim, 32),
nn.GELU(),
nn.Linear(32, out_dim),
nn.Tanh(),
)
def forward(self, x):
scale_emb = self.scale_emb.repeat(x.shape[0], 1, 1)
x = torch.cat([scale_emb, x], dim=1)
x = self.ln1(x)
x = x.permute(1, 0, 2)
x = self.transformer_blocks(x)
x = x.permute(1, 0, 2)
x = self.ln2(x[:, :self.num_tokens, :])
x = self.out(x)
return x
class DropPath(nn.Module):
r"""DropPath but without rescaling and supports optional all-zero and/or all-keep.
"""
def __init__(self, p):
super(DropPath, self).__init__()
self.p = p
def forward(self, *args, zero=None, keep=None):
if not self.training:
return args[0] if len(args) == 1 else args
# params
x = args[0]
b = x.size(0)
n = (torch.rand(b) < self.p).sum()
# non-zero and non-keep mask
mask = x.new_ones(b, dtype=torch.bool)
if keep is not None:
mask[keep] = False
if zero is not None:
mask[zero] = False
# drop-path index
index = torch.where(mask)[0]
index = index[torch.randperm(len(index))[:n]]
if zero is not None:
index = torch.cat([index, torch.where(zero)[0]], dim=0)
# drop-path multiplier
multiplier = x.new_ones(b)
multiplier[index] = 0.0
output = tuple(u * self.broadcast(multiplier, u) for u in args)
return output[0] if len(args) == 1 else output
def broadcast(self, src, dst):
assert src.size(0) == dst.size(0)
shape = (dst.size(0), ) + (1, ) * (dst.ndim - 1)
return src.view(shape)
class ImageContext(nn.Module):
def __init__(self, width=1024, context_dim=768, token_num=1):
super().__init__()
self.width = width
self.token_num = token_num
self.context_dim = context_dim
self.fc = nn.Sequential(
nn.Linear(context_dim, width),
nn.SiLU(),
nn.Linear(width, token_num * context_dim),
)
self.drop_path = DropPath(0.5)
def forward(self, x):
# x shape [B, C]
out = self.drop_path(self.fc(x))
out = rearrange(out, 'b (n c) -> b n c', n=self.token_num)
return out
class StyleAdapterDualAttnAS(nn.Module):
def __init__(self, image_context_config, scale_predictor_config, scale=1.0, use_norm=False, time_embed_dim=1024, mid_dim=32):
super().__init__()
self.image_context_model = instantiate_from_config(image_context_config)
self.scale_predictor = instantiate_from_config(scale_predictor_config)
self.scale = scale
self.use_norm = use_norm
self.time_embed_dim = time_embed_dim
self.mid_dim = mid_dim
def create_cross_attention_adapter(self, unet):
ori_processor = register_attn_processor(unet)
dual_attn_processor = {}
for idx, key in enumerate(ori_processor.keys()):
kv_state_dicts = {
'k': {'weight': unet.state_dict()[key[:-10] + '.to_k.weight']},
'v': {'weight': unet.state_dict()[key[:-10] + '.to_v.weight']},
}
context_dim = kv_state_dicts['k']['weight'].shape[1]
inner_dim = kv_state_dicts['k']['weight'].shape[0]
print(key, context_dim, inner_dim)
dual_attn_processor[key] = DualCrossAttnProcessorAS(
context_dim=context_dim,
inner_dim=inner_dim,
state_dict=kv_state_dicts,
scale=self.scale,
use_norm=self.use_norm,
layer_idx=idx,
)
set_attn_processor(unet, dual_attn_processor)
dual_attn_processor = {key.replace('.', '_'): value for key, value in dual_attn_processor.items()}
self.add_module('kv_attn_layers', nn.ModuleDict(dual_attn_processor))
def set_cross_attention_adapter(self, unet):
dual_attn_processor = get_attn_processor(unet)
for key in dual_attn_processor.keys():
module_key = key.replace('.', '_')
dual_attn_processor[key] = self.kv_attn_layers[module_key]
print('set', key, module_key)
set_attn_processor(unet, dual_attn_processor)
def forward(self, x):
# x shape [B, C]
return self.image_context_model(x)
|