Upload EVPRefer
Browse files- config.json +12 -0
- evpconfig.py +10 -0
- model.py +320 -0
- model.safetensors +3 -0
- models.py +349 -0
config.json
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"EVPRefer"
|
4 |
+
],
|
5 |
+
"auto_map": {
|
6 |
+
"AutoConfig": "evpconfig.EVPConfig",
|
7 |
+
"AutoModel": "model.EVPRefer"
|
8 |
+
},
|
9 |
+
"model_type": "EVP",
|
10 |
+
"torch_dtype": "float32",
|
11 |
+
"transformers_version": "4.35.2"
|
12 |
+
}
|
evpconfig.py
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import PretrainedConfig
|
2 |
+
|
3 |
+
class EVPConfig(PretrainedConfig):
|
4 |
+
model_type = "EVP"
|
5 |
+
def __init__(
|
6 |
+
self,
|
7 |
+
**kwargs,
|
8 |
+
):
|
9 |
+
super().__init__(**kwargs)
|
10 |
+
|
model.py
ADDED
@@ -0,0 +1,320 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import os
|
5 |
+
import sys
|
6 |
+
from ldm.util import instantiate_from_config
|
7 |
+
from transformers.models.clip.modeling_clip import CLIPTextModel
|
8 |
+
from omegaconf import OmegaConf
|
9 |
+
from lib.mask_predictor import SimpleDecoding
|
10 |
+
from transformers import PreTrainedModel
|
11 |
+
from .models import UNetWrapper, TextAdapterRefer
|
12 |
+
from evpconfig import EVPConfig
|
13 |
+
from transformers import CLIPTokenizer
|
14 |
+
import torchvision.transforms as transforms
|
15 |
+
|
16 |
+
|
17 |
+
def icnr(x, scale=2, init=nn.init.kaiming_normal_):
|
18 |
+
"""
|
19 |
+
Checkerboard artifact free sub-pixel convolution
|
20 |
+
https://arxiv.org/abs/1707.02937
|
21 |
+
"""
|
22 |
+
ni,nf,h,w = x.shape
|
23 |
+
ni2 = int(ni/(scale**2))
|
24 |
+
k = init(torch.zeros([ni2,nf,h,w])).transpose(0, 1)
|
25 |
+
k = k.contiguous().view(ni2, nf, -1)
|
26 |
+
k = k.repeat(1, 1, scale**2)
|
27 |
+
k = k.contiguous().view([nf,ni,h,w]).transpose(0, 1)
|
28 |
+
x.data.copy_(k)
|
29 |
+
|
30 |
+
|
31 |
+
class PixelShuffle(nn.Module):
|
32 |
+
"""
|
33 |
+
Real-Time Single Image and Video Super-Resolution
|
34 |
+
https://arxiv.org/abs/1609.05158
|
35 |
+
"""
|
36 |
+
def __init__(self, n_channels, scale):
|
37 |
+
super(PixelShuffle, self).__init__()
|
38 |
+
self.conv = nn.Conv2d(n_channels, n_channels*(scale**2), kernel_size=1)
|
39 |
+
icnr(self.conv.weight)
|
40 |
+
self.shuf = nn.PixelShuffle(scale)
|
41 |
+
self.relu = nn.ReLU()
|
42 |
+
|
43 |
+
def forward(self,x):
|
44 |
+
x = self.shuf(self.relu(self.conv(x)))
|
45 |
+
return x
|
46 |
+
|
47 |
+
|
48 |
+
class AttentionModule(nn.Module):
|
49 |
+
def __init__(self, in_channels, out_channels):
|
50 |
+
super(AttentionModule, self).__init__()
|
51 |
+
|
52 |
+
# Convolutional Layers
|
53 |
+
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
|
54 |
+
|
55 |
+
# Group Normalization
|
56 |
+
self.group_norm = nn.GroupNorm(20, out_channels)
|
57 |
+
|
58 |
+
# ReLU Activation
|
59 |
+
self.relu = nn.ReLU()
|
60 |
+
|
61 |
+
# Spatial Attention
|
62 |
+
self.spatial_attention = nn.Sequential(
|
63 |
+
nn.Conv2d(in_channels, 1, kernel_size=1),
|
64 |
+
nn.Sigmoid()
|
65 |
+
)
|
66 |
+
|
67 |
+
def forward(self, x):
|
68 |
+
# Apply spatial attention
|
69 |
+
spatial_attention = self.spatial_attention(x)
|
70 |
+
x = x * spatial_attention
|
71 |
+
|
72 |
+
# Apply convolutional layer
|
73 |
+
x = self.conv1(x)
|
74 |
+
x = self.group_norm(x)
|
75 |
+
x = self.relu(x)
|
76 |
+
|
77 |
+
return x
|
78 |
+
|
79 |
+
|
80 |
+
class AttentionDownsamplingModule(nn.Module):
|
81 |
+
def __init__(self, in_channels, out_channels, scale_factor=2):
|
82 |
+
super(AttentionDownsamplingModule, self).__init__()
|
83 |
+
|
84 |
+
# Spatial Attention
|
85 |
+
self.spatial_attention = nn.Sequential(
|
86 |
+
nn.Conv2d(in_channels, 1, kernel_size=1),
|
87 |
+
nn.Sigmoid()
|
88 |
+
)
|
89 |
+
|
90 |
+
# Channel Attention
|
91 |
+
self.channel_attention = nn.Sequential(
|
92 |
+
nn.AdaptiveAvgPool2d(1),
|
93 |
+
nn.Conv2d(in_channels, in_channels // 8, kernel_size=1),
|
94 |
+
nn.ReLU(inplace=True),
|
95 |
+
nn.Conv2d(in_channels // 8, in_channels, kernel_size=1),
|
96 |
+
nn.Sigmoid()
|
97 |
+
)
|
98 |
+
|
99 |
+
# Convolutional Layers
|
100 |
+
if scale_factor == 2:
|
101 |
+
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
|
102 |
+
elif scale_factor == 4:
|
103 |
+
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=2, padding=1)
|
104 |
+
|
105 |
+
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=2, padding=1)
|
106 |
+
|
107 |
+
# Group Normalization
|
108 |
+
self.group_norm = nn.GroupNorm(20, out_channels)
|
109 |
+
|
110 |
+
# ReLU Activation
|
111 |
+
self.relu = nn.ReLU(inplace=True)
|
112 |
+
|
113 |
+
def forward(self, x):
|
114 |
+
# Apply spatial attention
|
115 |
+
spatial_attention = self.spatial_attention(x)
|
116 |
+
x = x * spatial_attention
|
117 |
+
|
118 |
+
# Apply channel attention
|
119 |
+
channel_attention = self.channel_attention(x)
|
120 |
+
x = x * channel_attention
|
121 |
+
|
122 |
+
# Apply convolutional layers
|
123 |
+
x = self.conv1(x)
|
124 |
+
x = self.group_norm(x)
|
125 |
+
x = self.relu(x)
|
126 |
+
x = self.conv2(x)
|
127 |
+
x = self.group_norm(x)
|
128 |
+
x = self.relu(x)
|
129 |
+
|
130 |
+
return x
|
131 |
+
|
132 |
+
|
133 |
+
class AttentionUpsamplingModule(nn.Module):
|
134 |
+
def __init__(self, in_channels, out_channels):
|
135 |
+
super(AttentionUpsamplingModule, self).__init__()
|
136 |
+
|
137 |
+
# Spatial Attention for outs[2]
|
138 |
+
self.spatial_attention = nn.Sequential(
|
139 |
+
nn.Conv2d(in_channels, 1, kernel_size=1),
|
140 |
+
nn.Sigmoid()
|
141 |
+
)
|
142 |
+
|
143 |
+
# Channel Attention for outs[2]
|
144 |
+
self.channel_attention = nn.Sequential(
|
145 |
+
nn.AdaptiveAvgPool2d(1),
|
146 |
+
nn.Conv2d(in_channels, in_channels // 8, kernel_size=1),
|
147 |
+
nn.ReLU(),
|
148 |
+
nn.Conv2d(in_channels // 8, in_channels, kernel_size=1),
|
149 |
+
nn.Sigmoid()
|
150 |
+
)
|
151 |
+
|
152 |
+
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
|
153 |
+
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
|
154 |
+
|
155 |
+
# Group Normalization
|
156 |
+
self.group_norm = nn.GroupNorm(20, out_channels)
|
157 |
+
|
158 |
+
# ReLU Activation
|
159 |
+
self.relu = nn.ReLU()
|
160 |
+
self.upscale = PixelShuffle(in_channels, 2)
|
161 |
+
|
162 |
+
def forward(self, x):
|
163 |
+
# Apply spatial attention
|
164 |
+
spatial_attention = self.spatial_attention(x)
|
165 |
+
x = x * spatial_attention
|
166 |
+
|
167 |
+
# Apply channel attention
|
168 |
+
channel_attention = self.channel_attention(x)
|
169 |
+
x = x * channel_attention
|
170 |
+
|
171 |
+
# Apply convolutional layers
|
172 |
+
x = self.conv1(x)
|
173 |
+
x = self.group_norm(x)
|
174 |
+
x = self.relu(x)
|
175 |
+
x = self.conv2(x)
|
176 |
+
x = self.group_norm(x)
|
177 |
+
x = self.relu(x)
|
178 |
+
|
179 |
+
# Upsample
|
180 |
+
x = self.upscale(x)
|
181 |
+
|
182 |
+
return x
|
183 |
+
|
184 |
+
|
185 |
+
class ConvLayer(nn.Module):
|
186 |
+
def __init__(self, in_channels, out_channels):
|
187 |
+
super(ConvLayer, self).__init__()
|
188 |
+
|
189 |
+
self.conv1 = nn.Sequential(
|
190 |
+
nn.Conv2d(in_channels, out_channels, 1),
|
191 |
+
nn.GroupNorm(20, out_channels),
|
192 |
+
nn.ReLU(),
|
193 |
+
)
|
194 |
+
|
195 |
+
def forward(self, x):
|
196 |
+
x = self.conv1(x)
|
197 |
+
|
198 |
+
return x
|
199 |
+
|
200 |
+
|
201 |
+
class InverseMultiAttentiveFeatureRefinement(nn.Module):
|
202 |
+
def __init__(self, in_channels_list):
|
203 |
+
super(InverseMultiAttentiveFeatureRefinement, self).__init__()
|
204 |
+
|
205 |
+
self.layer1 = AttentionModule(in_channels_list[0], in_channels_list[0])
|
206 |
+
self.layer2 = AttentionDownsamplingModule(in_channels_list[0], in_channels_list[0]//2, scale_factor = 2)
|
207 |
+
self.layer3 = ConvLayer(in_channels_list[0]//2 + in_channels_list[1], in_channels_list[1])
|
208 |
+
self.layer4 = AttentionDownsamplingModule(in_channels_list[1], in_channels_list[1]//2, scale_factor = 2)
|
209 |
+
self.layer5 = ConvLayer(in_channels_list[1]//2 + in_channels_list[2], in_channels_list[2])
|
210 |
+
self.layer6 = AttentionDownsamplingModule(in_channels_list[2], in_channels_list[2]//2, scale_factor = 2)
|
211 |
+
self.layer7 = ConvLayer(in_channels_list[2]//2 + in_channels_list[3], in_channels_list[3])
|
212 |
+
|
213 |
+
'''
|
214 |
+
self.layer8 = AttentionUpsamplingModule(in_channels_list[3], in_channels_list[3])
|
215 |
+
self.layer9 = ConvLayer(in_channels_list[2] + in_channels_list[3], in_channels_list[2])
|
216 |
+
self.layer10 = AttentionUpsamplingModule(in_channels_list[2], in_channels_list[2])
|
217 |
+
self.layer11 = ConvLayer(in_channels_list[1] + in_channels_list[2], in_channels_list[1])
|
218 |
+
self.layer12 = AttentionUpsamplingModule(in_channels_list[1], in_channels_list[1])
|
219 |
+
self.layer13 = ConvLayer(in_channels_list[0] + in_channels_list[1], in_channels_list[0])
|
220 |
+
'''
|
221 |
+
def forward(self, inputs):
|
222 |
+
x_c4, x_c3, x_c2, x_c1 = inputs
|
223 |
+
x_c4 = self.layer1(x_c4)
|
224 |
+
x_c4_3 = self.layer2(x_c4)
|
225 |
+
x_c3 = torch.cat([x_c4_3, x_c3], dim=1)
|
226 |
+
x_c3 = self.layer3(x_c3)
|
227 |
+
x_c3_2 = self.layer4(x_c3)
|
228 |
+
x_c2 = torch.cat([x_c3_2, x_c2], dim=1)
|
229 |
+
x_c2 = self.layer5(x_c2)
|
230 |
+
x_c2_1 = self.layer6(x_c2)
|
231 |
+
x_c1 = torch.cat([x_c2_1, x_c1], dim=1)
|
232 |
+
x_c1 = self.layer7(x_c1)
|
233 |
+
'''
|
234 |
+
x_c1_2 = self.layer8(x_c1)
|
235 |
+
x_c2 = torch.cat([x_c1_2, x_c2], dim=1)
|
236 |
+
x_c2 = self.layer9(x_c2)
|
237 |
+
x_c2_3 = self.layer10(x_c2)
|
238 |
+
x_c3 = torch.cat([x_c2_3, x_c3], dim=1)
|
239 |
+
x_c3 = self.layer11(x_c3)
|
240 |
+
x_c3_4 = self.layer12(x_c3)
|
241 |
+
x_c4 = torch.cat([x_c3_4, x_c4], dim=1)
|
242 |
+
x_c4 = self.layer13(x_c4)
|
243 |
+
'''
|
244 |
+
return [x_c4, x_c3, x_c2, x_c1]
|
245 |
+
|
246 |
+
|
247 |
+
|
248 |
+
class EVPRefer(PreTrainedModel):
|
249 |
+
"""Encoder Decoder segmentors.
|
250 |
+
|
251 |
+
EncoderDecoder typically consists of backbone, decode_head, auxiliary_head.
|
252 |
+
Note that auxiliary_head is only used for deep supervision during training,
|
253 |
+
which could be dumped during inference.
|
254 |
+
"""
|
255 |
+
config_class = EVPConfig
|
256 |
+
def __init__(self, config,
|
257 |
+
sd_path=None,
|
258 |
+
base_size=512,
|
259 |
+
token_embed_dim=768,
|
260 |
+
neck_dim=[320,680,1320,1280],
|
261 |
+
**args):
|
262 |
+
super().__init__(config)
|
263 |
+
config = OmegaConf.load('./v1-inference.yaml')
|
264 |
+
if os.path.exists(f'{sd_path}'):
|
265 |
+
config.model.params.ckpt_path = f'{sd_path}'
|
266 |
+
else:
|
267 |
+
config.model.params.ckpt_path = None
|
268 |
+
|
269 |
+
sd_model = instantiate_from_config(config.model)
|
270 |
+
self.encoder_vq = sd_model.first_stage_model
|
271 |
+
self.unet = UNetWrapper(sd_model.model, base_size=base_size)
|
272 |
+
del sd_model.cond_stage_model
|
273 |
+
del self.encoder_vq.decoder
|
274 |
+
for param in self.encoder_vq.parameters():
|
275 |
+
param.requires_grad = True
|
276 |
+
|
277 |
+
self.text_adapter = TextAdapterRefer(text_dim=token_embed_dim)
|
278 |
+
|
279 |
+
self.classifier = SimpleDecoding(dims=neck_dim)
|
280 |
+
|
281 |
+
self.gamma = nn.Parameter(torch.ones(token_embed_dim) * 1e-4)
|
282 |
+
self.aggregation = InverseMultiAttentiveFeatureRefinement([320,680,1320,1280])
|
283 |
+
self.clip_model = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14")
|
284 |
+
self.tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
|
285 |
+
|
286 |
+
for param in self.clip_model.parameters():
|
287 |
+
param.requires_grad = True
|
288 |
+
|
289 |
+
|
290 |
+
def forward(self, img, sentences):
|
291 |
+
image_t = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])(img)
|
292 |
+
shape = image_t.shape
|
293 |
+
img = torch.nn.functional.interpolate(image_t, (512,512), mode='bilinear', align_corners=True)
|
294 |
+
|
295 |
+
input_ids = self.tokenizer(text=sentences, truncation=True, max_length=40, return_length=True,
|
296 |
+
return_overflowing_tokens=False, padding="max_length", return_tensors="pt")['input_ids'].to(image_t.device)
|
297 |
+
|
298 |
+
|
299 |
+
input_shape = img.shape[-2:]
|
300 |
+
|
301 |
+
latents = self.encoder_vq.encode(img).mode()
|
302 |
+
latents = latents / 4.7164
|
303 |
+
|
304 |
+
l_feats = self.clip_model(input_ids=input_ids).last_hidden_state
|
305 |
+
c_crossattn = self.text_adapter(latents, l_feats, self.gamma) # NOTE: here the c_crossattn should be expand_dim as latents
|
306 |
+
t = torch.ones((img.shape[0],), device=img.device).long()
|
307 |
+
outs = self.unet(latents, t, c_crossattn=[c_crossattn])
|
308 |
+
|
309 |
+
outs = self.aggregation(outs)
|
310 |
+
|
311 |
+
x_c1, x_c2, x_c3, x_c4 = outs
|
312 |
+
x = self.classifier(x_c4, x_c3, x_c2, x_c1)
|
313 |
+
x = F.interpolate(x, size=input_shape, mode='bilinear', align_corners=True)
|
314 |
+
pred = torch.nn.functional.interpolate(x, shape[2:], mode='bilinear', align_corners=True)
|
315 |
+
output_mask = pred.detach().cpu().argmax(1).data.numpy().squeeze()
|
316 |
+
return output_mask
|
317 |
+
|
318 |
+
|
319 |
+
def get_latent(self, x):
|
320 |
+
return self.encoder_vq.encode(x).mode()
|
model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:07d8a7895e79a4defbf5445cada1a9973bcda51f7c73e4a91c878074a5f758f5
|
3 |
+
size 4317946624
|
models.py
ADDED
@@ -0,0 +1,349 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from omegaconf import OmegaConf
|
2 |
+
|
3 |
+
import torch as th
|
4 |
+
import torch
|
5 |
+
import math
|
6 |
+
import abc
|
7 |
+
|
8 |
+
from torch import nn, einsum
|
9 |
+
|
10 |
+
from einops import rearrange, repeat
|
11 |
+
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
|
12 |
+
from transformers import CLIPTokenizer
|
13 |
+
from transformers.models.clip.modeling_clip import CLIPTextConfig, CLIPTextModel, CLIPTextTransformer#, _expand_mask
|
14 |
+
from inspect import isfunction
|
15 |
+
|
16 |
+
|
17 |
+
def exists(val):
|
18 |
+
return val is not None
|
19 |
+
|
20 |
+
|
21 |
+
def uniq(arr):
|
22 |
+
return{el: True for el in arr}.keys()
|
23 |
+
|
24 |
+
|
25 |
+
def default(val, d):
|
26 |
+
if exists(val):
|
27 |
+
return val
|
28 |
+
return d() if isfunction(d) else d
|
29 |
+
|
30 |
+
|
31 |
+
|
32 |
+
def register_attention_control(model, controller):
|
33 |
+
def ca_forward(self, place_in_unet):
|
34 |
+
def forward(x, context=None, mask=None):
|
35 |
+
h = self.heads
|
36 |
+
|
37 |
+
q = self.to_q(x)
|
38 |
+
is_cross = context is not None
|
39 |
+
context = default(context, x)
|
40 |
+
k = self.to_k(context)
|
41 |
+
v = self.to_v(context)
|
42 |
+
|
43 |
+
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
|
44 |
+
|
45 |
+
sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
|
46 |
+
|
47 |
+
if exists(mask):
|
48 |
+
mask = rearrange(mask, 'b ... -> b (...)')
|
49 |
+
max_neg_value = -torch.finfo(sim.dtype).max
|
50 |
+
mask = repeat(mask, 'b j -> (b h) () j', h=h)
|
51 |
+
sim.masked_fill_(~mask, max_neg_value)
|
52 |
+
|
53 |
+
# attention, what we cannot get enough of
|
54 |
+
attn = sim.softmax(dim=-1)
|
55 |
+
|
56 |
+
attn2 = rearrange(attn, '(b h) k c -> h b k c', h=h).mean(0)
|
57 |
+
controller(attn2, is_cross, place_in_unet)
|
58 |
+
|
59 |
+
out = einsum('b i j, b j d -> b i d', attn, v)
|
60 |
+
out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
|
61 |
+
return self.to_out(out)
|
62 |
+
|
63 |
+
return forward
|
64 |
+
|
65 |
+
class DummyController:
|
66 |
+
def __call__(self, *args):
|
67 |
+
return args[0]
|
68 |
+
|
69 |
+
def __init__(self):
|
70 |
+
self.num_att_layers = 0
|
71 |
+
|
72 |
+
if controller is None:
|
73 |
+
controller = DummyController()
|
74 |
+
|
75 |
+
def register_recr(net_, count, place_in_unet):
|
76 |
+
if net_.__class__.__name__ == 'CrossAttention':
|
77 |
+
net_.forward = ca_forward(net_, place_in_unet)
|
78 |
+
return count + 1
|
79 |
+
elif hasattr(net_, 'children'):
|
80 |
+
for net__ in net_.children():
|
81 |
+
count = register_recr(net__, count, place_in_unet)
|
82 |
+
return count
|
83 |
+
|
84 |
+
cross_att_count = 0
|
85 |
+
sub_nets = model.diffusion_model.named_children()
|
86 |
+
|
87 |
+
for net in sub_nets:
|
88 |
+
if "input_blocks" in net[0]:
|
89 |
+
cross_att_count += register_recr(net[1], 0, "down")
|
90 |
+
elif "output_blocks" in net[0]:
|
91 |
+
cross_att_count += register_recr(net[1], 0, "up")
|
92 |
+
elif "middle_block" in net[0]:
|
93 |
+
cross_att_count += register_recr(net[1], 0, "mid")
|
94 |
+
|
95 |
+
controller.num_att_layers = cross_att_count
|
96 |
+
|
97 |
+
|
98 |
+
class AttentionControl(abc.ABC):
|
99 |
+
|
100 |
+
def step_callback(self, x_t):
|
101 |
+
return x_t
|
102 |
+
|
103 |
+
def between_steps(self):
|
104 |
+
return
|
105 |
+
|
106 |
+
@property
|
107 |
+
def num_uncond_att_layers(self):
|
108 |
+
return 0
|
109 |
+
|
110 |
+
@abc.abstractmethod
|
111 |
+
def forward (self, attn, is_cross: bool, place_in_unet: str):
|
112 |
+
raise NotImplementedError
|
113 |
+
|
114 |
+
def __call__(self, attn, is_cross: bool, place_in_unet: str):
|
115 |
+
attn = self.forward(attn, is_cross, place_in_unet)
|
116 |
+
return attn
|
117 |
+
|
118 |
+
def reset(self):
|
119 |
+
self.cur_step = 0
|
120 |
+
self.cur_att_layer = 0
|
121 |
+
|
122 |
+
def __init__(self):
|
123 |
+
self.cur_step = 0
|
124 |
+
self.num_att_layers = -1
|
125 |
+
self.cur_att_layer = 0
|
126 |
+
|
127 |
+
|
128 |
+
class AttentionStore(AttentionControl):
|
129 |
+
@staticmethod
|
130 |
+
def get_empty_store():
|
131 |
+
return {"down_cross": [], "mid_cross": [], "up_cross": [],
|
132 |
+
"down_self": [], "mid_self": [], "up_self": []}
|
133 |
+
|
134 |
+
def forward(self, attn, is_cross: bool, place_in_unet: str):
|
135 |
+
key = f"{place_in_unet}_{'cross' if is_cross else 'self'}"
|
136 |
+
if attn.shape[1] <= (self.max_size) ** 2: # avoid memory overhead
|
137 |
+
self.step_store[key].append(attn)
|
138 |
+
return attn
|
139 |
+
|
140 |
+
def between_steps(self):
|
141 |
+
if len(self.attention_store) == 0:
|
142 |
+
self.attention_store = self.step_store
|
143 |
+
else:
|
144 |
+
for key in self.attention_store:
|
145 |
+
for i in range(len(self.attention_store[key])):
|
146 |
+
self.attention_store[key][i] += self.step_store[key][i]
|
147 |
+
self.step_store = self.get_empty_store()
|
148 |
+
|
149 |
+
def get_average_attention(self):
|
150 |
+
average_attention = {key: [item for item in self.step_store[key]] for key in self.step_store}
|
151 |
+
return average_attention
|
152 |
+
|
153 |
+
def reset(self):
|
154 |
+
super(AttentionStore, self).reset()
|
155 |
+
self.step_store = self.get_empty_store()
|
156 |
+
self.attention_store = {}
|
157 |
+
|
158 |
+
def __init__(self, base_size=64, max_size=None):
|
159 |
+
super(AttentionStore, self).__init__()
|
160 |
+
self.step_store = self.get_empty_store()
|
161 |
+
self.attention_store = {}
|
162 |
+
self.base_size = base_size
|
163 |
+
if max_size is None:
|
164 |
+
self.max_size = self.base_size // 2
|
165 |
+
else:
|
166 |
+
self.max_size = max_size
|
167 |
+
|
168 |
+
def register_hier_output(model):
|
169 |
+
self = model.diffusion_model
|
170 |
+
from ldm.modules.diffusionmodules.util import checkpoint, timestep_embedding
|
171 |
+
def forward(x, timesteps=None, context=None, y=None,**kwargs):
|
172 |
+
"""
|
173 |
+
Apply the model to an input batch.
|
174 |
+
:param x: an [N x C x ...] Tensor of inputs.
|
175 |
+
:param timesteps: a 1-D batch of timesteps.
|
176 |
+
:param context: conditioning plugged in via crossattn
|
177 |
+
:param y: an [N] Tensor of labels, if class-conditional.
|
178 |
+
:return: an [N x C x ...] Tensor of outputs.
|
179 |
+
"""
|
180 |
+
assert (y is not None) == (
|
181 |
+
self.num_classes is not None
|
182 |
+
), "must specify y if and only if the model is class-conditional"
|
183 |
+
hs = []
|
184 |
+
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
|
185 |
+
emb = self.time_embed(t_emb)
|
186 |
+
|
187 |
+
if self.num_classes is not None:
|
188 |
+
assert y.shape == (x.shape[0],)
|
189 |
+
emb = emb + self.label_emb(y)
|
190 |
+
|
191 |
+
h = x.type(self.dtype)
|
192 |
+
for module in self.input_blocks:
|
193 |
+
# import pdb; pdb.set_trace()
|
194 |
+
if context.shape[1]==2:
|
195 |
+
h = module(h, emb, context[:,0,:].unsqueeze(1))
|
196 |
+
else:
|
197 |
+
h = module(h, emb, context)
|
198 |
+
hs.append(h)
|
199 |
+
if context.shape[1]==2:
|
200 |
+
h = self.middle_block(h, emb, context[:,0,:].unsqueeze(1))
|
201 |
+
else:
|
202 |
+
h = self.middle_block(h, emb, context)
|
203 |
+
out_list = []
|
204 |
+
|
205 |
+
for i_out, module in enumerate(self.output_blocks):
|
206 |
+
h = th.cat([h, hs.pop()], dim=1)
|
207 |
+
if context.shape[1]==2:
|
208 |
+
h = module(h, emb, context[:,1,:].unsqueeze(1))
|
209 |
+
else:
|
210 |
+
h = module(h, emb, context)
|
211 |
+
if i_out in [1, 4, 7]:
|
212 |
+
out_list.append(h)
|
213 |
+
h = h.type(x.dtype)
|
214 |
+
|
215 |
+
out_list.append(h)
|
216 |
+
return out_list
|
217 |
+
|
218 |
+
self.forward = forward
|
219 |
+
|
220 |
+
class UNetWrapper(nn.Module):
|
221 |
+
def __init__(self, unet, use_attn=True, base_size=512, max_attn_size=None, attn_selector='up_cross+down_cross') -> None:
|
222 |
+
super().__init__()
|
223 |
+
self.unet = unet
|
224 |
+
self.attention_store = AttentionStore(base_size=base_size // 8, max_size=max_attn_size)
|
225 |
+
self.size16 = base_size // 32
|
226 |
+
self.size32 = base_size // 16
|
227 |
+
self.size64 = base_size // 8
|
228 |
+
self.use_attn = use_attn
|
229 |
+
if self.use_attn:
|
230 |
+
register_attention_control(unet, self.attention_store)
|
231 |
+
register_hier_output(unet)
|
232 |
+
self.attn_selector = attn_selector.split('+')
|
233 |
+
|
234 |
+
def forward(self, *args, **kwargs):
|
235 |
+
if self.use_attn:
|
236 |
+
self.attention_store.reset()
|
237 |
+
out_list = self.unet(*args, **kwargs)
|
238 |
+
if self.use_attn:
|
239 |
+
avg_attn = self.attention_store.get_average_attention()
|
240 |
+
attn16, attn32, attn64 = self.process_attn(avg_attn)
|
241 |
+
out_list[1] = torch.cat([out_list[1], attn16], dim=1)
|
242 |
+
out_list[2] = torch.cat([out_list[2], attn32], dim=1)
|
243 |
+
if attn64 is not None:
|
244 |
+
out_list[3] = torch.cat([out_list[3], attn64], dim=1)
|
245 |
+
return out_list[::-1]
|
246 |
+
|
247 |
+
def process_attn(self, avg_attn):
|
248 |
+
attns = {self.size16: [], self.size32: [], self.size64: []}
|
249 |
+
for k in self.attn_selector:
|
250 |
+
for up_attn in avg_attn[k]:
|
251 |
+
size = int(math.sqrt(up_attn.shape[1]))
|
252 |
+
attns[size].append(rearrange(up_attn, 'b (h w) c -> b c h w', h=size))
|
253 |
+
attn16 = torch.stack(attns[self.size16]).mean(0)
|
254 |
+
attn32 = torch.stack(attns[self.size32]).mean(0)
|
255 |
+
if len(attns[self.size64]) > 0:
|
256 |
+
attn64 = torch.stack(attns[self.size64]).mean(0)
|
257 |
+
else:
|
258 |
+
attn64 = None
|
259 |
+
return attn16, attn32, attn64
|
260 |
+
|
261 |
+
class TextAdapter(nn.Module):
|
262 |
+
def __init__(self, text_dim=768, hidden_dim=None):
|
263 |
+
super().__init__()
|
264 |
+
if hidden_dim is None:
|
265 |
+
hidden_dim = text_dim
|
266 |
+
self.fc = nn.Sequential(
|
267 |
+
nn.Linear(text_dim, hidden_dim),
|
268 |
+
nn.GELU(),
|
269 |
+
nn.Linear(hidden_dim, text_dim)
|
270 |
+
)
|
271 |
+
|
272 |
+
def forward(self, latents, texts, gamma):
|
273 |
+
n_class, channel = texts.shape
|
274 |
+
bs = latents.shape[0]
|
275 |
+
|
276 |
+
texts_after = self.fc(texts)
|
277 |
+
texts = texts + gamma * texts_after
|
278 |
+
texts = repeat(texts, 'n c -> b n c', b=bs)
|
279 |
+
return texts
|
280 |
+
|
281 |
+
class TextAdapterRefer(nn.Module):
|
282 |
+
def __init__(self, text_dim=768):
|
283 |
+
super().__init__()
|
284 |
+
|
285 |
+
self.fc = nn.Sequential(
|
286 |
+
nn.Linear(text_dim, text_dim),
|
287 |
+
nn.GELU(),
|
288 |
+
nn.Linear(text_dim, text_dim)
|
289 |
+
)
|
290 |
+
|
291 |
+
def forward(self, latents, texts, gamma):
|
292 |
+
texts_after = self.fc(texts)
|
293 |
+
texts = texts + gamma * texts_after
|
294 |
+
return texts
|
295 |
+
|
296 |
+
|
297 |
+
class TextAdapterDepth(nn.Module):
|
298 |
+
def __init__(self, text_dim=768):
|
299 |
+
super().__init__()
|
300 |
+
|
301 |
+
self.fc = nn.Sequential(
|
302 |
+
nn.Linear(text_dim, text_dim),
|
303 |
+
nn.GELU(),
|
304 |
+
nn.Linear(text_dim, text_dim)
|
305 |
+
)
|
306 |
+
|
307 |
+
def forward(self, latents, texts, gamma):
|
308 |
+
# use the gamma to blend
|
309 |
+
n_sen, channel = texts.shape
|
310 |
+
bs = latents.shape[0]
|
311 |
+
|
312 |
+
texts_after = self.fc(texts)
|
313 |
+
texts = texts + gamma * texts_after
|
314 |
+
texts = repeat(texts, 'n c -> n b c', b=1)
|
315 |
+
return texts
|
316 |
+
|
317 |
+
|
318 |
+
class FrozenCLIPEmbedder(nn.Module):
|
319 |
+
"""Uses the CLIP transformer encoder for text (from Hugging Face)"""
|
320 |
+
def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77, pool=True):
|
321 |
+
super().__init__()
|
322 |
+
self.tokenizer = CLIPTokenizer.from_pretrained(version)
|
323 |
+
self.transformer = CLIPTextModel.from_pretrained(version)
|
324 |
+
self.device = device
|
325 |
+
self.max_length = max_length
|
326 |
+
self.freeze()
|
327 |
+
|
328 |
+
self.pool = pool
|
329 |
+
|
330 |
+
def freeze(self):
|
331 |
+
self.transformer = self.transformer.eval()
|
332 |
+
for param in self.parameters():
|
333 |
+
param.requires_grad = False
|
334 |
+
|
335 |
+
def forward(self, text):
|
336 |
+
batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
|
337 |
+
return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
|
338 |
+
tokens = batch_encoding["input_ids"].to(self.device)
|
339 |
+
outputs = self.transformer(input_ids=tokens)
|
340 |
+
|
341 |
+
if self.pool:
|
342 |
+
z = outputs.pooler_output
|
343 |
+
else:
|
344 |
+
z = outputs.last_hidden_state
|
345 |
+
return z
|
346 |
+
|
347 |
+
def encode(self, text):
|
348 |
+
return self(text)
|
349 |
+
|