File size: 7,524 Bytes
34097e9 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 |
"""
Copyright (C) 2019 NVIDIA Corporation. All rights reserved.
Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode).
"""
import re
import torch
import torch.nn as nn
from ldm.modules.diffusionmodules.util import normalization, checkpoint
from ldm.modules.diffusionmodules.openaimodel import ResBlock, UNetModel
class SPADE(nn.Module):
def __init__(self, norm_nc, label_nc=256, config_text='spadeinstance3x3'):
super().__init__()
assert config_text.startswith('spade')
parsed = re.search('spade(\D+)(\d)x\d', config_text)
ks = int(parsed.group(2))
self.param_free_norm = normalization(norm_nc)
# The dimension of the intermediate embedding space. Yes, hardcoded.
nhidden = 128
pw = ks // 2
self.mlp_shared = nn.Sequential(
nn.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw),
nn.ReLU()
)
self.mlp_gamma = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw)
self.mlp_beta = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw)
def forward(self, x_dic, segmap_dic):
return checkpoint(
self._forward, (x_dic, segmap_dic), self.parameters(), True
)
def _forward(self, x_dic, segmap_dic):
segmap = segmap_dic[str(x_dic.size(-1))]
x = x_dic
# Part 1. generate parameter-free normalized activations
normalized = self.param_free_norm(x)
# Part 2. produce scaling and bias conditioned on semantic map
# segmap = F.interpolate(segmap, size=x.size()[2:], mode='nearest')
actv = self.mlp_shared(segmap)
repeat_factor = normalized.shape[0]//segmap.shape[0]
if repeat_factor > 1:
out = normalized
out *= (1 + self.mlp_gamma(actv).repeat_interleave(repeat_factor, dim=0))
out += self.mlp_beta(actv).repeat_interleave(repeat_factor, dim=0)
else:
out = normalized
out *= (1 + self.mlp_gamma(actv))
out += self.mlp_beta(actv)
return out
def dual_resblock_forward(self: ResBlock, x, emb, spade: SPADE, get_struct_cond):
if self.updown:
in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
h = in_rest(x)
h = self.h_upd(h)
x = self.x_upd(x)
h = in_conv(h)
else:
h = self.in_layers(x)
emb_out = self.emb_layers(emb).type(h.dtype)
while len(emb_out.shape) < len(h.shape):
emb_out = emb_out[..., None]
if self.use_scale_shift_norm:
out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
scale, shift = torch.chunk(emb_out, 2, dim=1)
h = out_norm(h) * (1 + scale) + shift
h = out_rest(h)
else:
h = h + emb_out
h = self.out_layers(h)
h = spade(h, get_struct_cond())
return self.skip_connection(x) + h
class SPADELayers(nn.Module):
def __init__(self):
'''
A container class for fast SPADE layer loading.
params inferred from the official checkpoint
'''
super().__init__()
self.input_blocks = nn.ModuleList([
nn.Identity(),
SPADE(320),
SPADE(320),
nn.Identity(),
SPADE(640),
SPADE(640),
nn.Identity(),
SPADE(1280),
SPADE(1280),
nn.Identity(),
SPADE(1280),
SPADE(1280),
])
self.middle_block = nn.ModuleList([
SPADE(1280),
nn.Identity(),
SPADE(1280),
])
self.output_blocks = nn.ModuleList([
SPADE(1280),
SPADE(1280),
SPADE(1280),
SPADE(1280),
SPADE(1280),
SPADE(1280),
SPADE(640),
SPADE(640),
SPADE(640),
SPADE(320),
SPADE(320),
SPADE(320),
])
self.input_ids = [1,2,4,5,7,8,10,11]
self.output_ids = list(range(12))
self.mid_ids = [0,2]
self.forward_cache_name = 'org_forward_stablesr'
self.unet = None
def hook(self, unet: UNetModel, get_struct_cond):
# hook all resblocks
self.unet = unet
resblock: ResBlock = None
for i in self.input_ids:
resblock = unet.input_blocks[i][0]
# debug
# assert isinstance(resblock, ResBlock)
if not hasattr(resblock, self.forward_cache_name):
setattr(resblock, self.forward_cache_name, resblock._forward)
resblock._forward = lambda x, timesteps, resblock=resblock, spade=self.input_blocks[i]: dual_resblock_forward(resblock, x, timesteps, spade, get_struct_cond)
for i in self.output_ids:
resblock = unet.output_blocks[i][0]
# debug
# assert isinstance(resblock, ResBlock)
if not hasattr(resblock, self.forward_cache_name):
setattr(resblock, self.forward_cache_name, resblock._forward)
resblock._forward = lambda x, timesteps, resblock=resblock, spade=self.output_blocks[i]: dual_resblock_forward(resblock, x, timesteps, spade, get_struct_cond)
for i in self.mid_ids:
resblock = unet.middle_block[i]
# debug
# assert isinstance(resblock, ResBlock)
if not hasattr(resblock, self.forward_cache_name):
setattr(resblock, self.forward_cache_name, resblock._forward)
resblock._forward = lambda x, timesteps, resblock=resblock, spade=self.middle_block[i]: dual_resblock_forward(resblock, x, timesteps, spade, get_struct_cond)
def unhook(self):
unet = self.unet
if unet is None: return
resblock: ResBlock = None
for i in self.input_ids:
resblock = unet.input_blocks[i][0]
if hasattr(resblock, self.forward_cache_name):
resblock._forward = getattr(resblock, self.forward_cache_name)
delattr(resblock, self.forward_cache_name)
for i in self.output_ids:
resblock = unet.output_blocks[i][0]
if hasattr(resblock, self.forward_cache_name):
resblock._forward = getattr(resblock, self.forward_cache_name)
delattr(resblock, self.forward_cache_name)
for i in self.mid_ids:
resblock = unet.middle_block[i]
if hasattr(resblock, self.forward_cache_name):
resblock._forward = getattr(resblock, self.forward_cache_name)
delattr(resblock, self.forward_cache_name)
self.unet = None
def load_from_dict(self, state_dict):
"""
Load model weights from a dictionary.
:param state_dict: a dict of parameters.
"""
filtered_dict = {}
for k, v in state_dict.items():
if k.startswith("model.diffusion_model."):
key = k[len("model.diffusion_model.") :]
# remove the '.0.spade' within the key
if 'middle_block' not in key:
key = key.replace('.0.spade', '')
else:
key = key.replace('.spade', '')
filtered_dict[key] = v
self.load_state_dict(filtered_dict)
if __name__ == '__main__':
path = '../models/stablesr_sd21.ckpt'
state_dict = torch.load(path)
model = SPADELayers()
model.load_from_dict(state_dict)
print(model) |