Spaces:
Runtime error
Runtime error
Mehdi Cherti
commited on
Commit
•
e96a195
1
Parent(s):
2ab447a
update
Browse files- EMA.py +8 -0
- eval_all.sh +12 -2
- run.py +28 -1
- score_sde/models/discriminator.py +2 -2
- score_sde/models/layers.py +1 -0
- score_sde/models/projected_discriminator.py +783 -0
- scripts/init.sh +14 -2
- scripts/run_jurecadc_conda.sh +23 -0
- scripts/run_juwelsbooster_conda.sh +19 -0
- test.py +8 -0
- test_ddgan.py +7 -2
- train_ddgan.py +20 -4
EMA.py
CHANGED
@@ -21,8 +21,16 @@ class EMA(Optimizer):
|
|
21 |
self.optimizer = opt
|
22 |
self.state = opt.state
|
23 |
self.param_groups = opt.param_groups
|
|
|
24 |
|
25 |
def step(self, *args, **kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
retval = self.optimizer.step(*args, **kwargs)
|
27 |
|
28 |
# stop here if we are not applying EMA
|
|
|
21 |
self.optimizer = opt
|
22 |
self.state = opt.state
|
23 |
self.param_groups = opt.param_groups
|
24 |
+
self.defaults = {}
|
25 |
|
26 |
def step(self, *args, **kwargs):
|
27 |
+
# for group in self.optimizer.param_groups:
|
28 |
+
# group.setdefault('amsgrad', False)
|
29 |
+
# group.setdefault('maximize', False)
|
30 |
+
# group.setdefault('foreach', None)
|
31 |
+
# group.setdefault('capturable', False)
|
32 |
+
# group.setdefault('differentiable', False)
|
33 |
+
# group.setdefault('fused', False)
|
34 |
retval = self.optimizer.step(*args, **kwargs)
|
35 |
|
36 |
# stop here if we are not applying EMA
|
eval_all.sh
CHANGED
@@ -1,7 +1,17 @@
|
|
1 |
#!/bin/bash
|
2 |
-
for model in ddgan_sd_v10 ddgan_laion2b_v2 ddgan_ddb_v1 ddgan_ddb_v2 ddgan_ddb_v3;do
|
3 |
-
|
|
|
|
|
|
|
|
|
4 |
bs=32
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
else
|
6 |
bs=64
|
7 |
fi
|
|
|
1 |
#!/bin/bash
|
2 |
+
#for model in ddgan_sd_v10 ddgan_laion2b_v2 ddgan_ddb_v1 ddgan_ddb_v2 ddgan_ddb_v3 ddgan_ddb_v4;do
|
3 |
+
#for model in ddgan_ddb_v2 ddgan_ddb_v3 ddgan_ddb_v4 ddgan_ddb_v5;do
|
4 |
+
#for model in ddgan_ddb_v4 ddgan_ddb_v6 ddgan_ddb_v7 ddgan_laion_aesthetic_v15;do
|
5 |
+
#for model in ddgan_ddb_v6;do
|
6 |
+
for model in ddgan_laion_aesthetic_v15;do
|
7 |
+
if [ "$model" == "ddgan_ddb_v3" ]; then
|
8 |
bs=32
|
9 |
+
elif [ "$model" == "ddgan_laion_aesthetic_v15" ]; then
|
10 |
+
bs=32
|
11 |
+
elif [ "$model" == "ddgan_ddb_v6" ]; then
|
12 |
+
bs=32
|
13 |
+
elif [ "$model" == "ddgan_ddb_v4" ]; then
|
14 |
+
bs=16
|
15 |
else
|
16 |
bs=64
|
17 |
fi
|
run.py
CHANGED
@@ -256,6 +256,28 @@ def ddgan_ddb_v3():
|
|
256 |
cfg['model']['num_timesteps'] = 2
|
257 |
return cfg
|
258 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
259 |
models = [
|
260 |
ddgan_cifar10_cond17, # cifar10, cross attn for discr
|
261 |
ddgan_cifar10_cond18, # cifar10, xl encoder
|
@@ -283,6 +305,7 @@ models = [
|
|
283 |
ddgan_laion_aesthetic_v12,
|
284 |
ddgan_laion_aesthetic_v13,
|
285 |
ddgan_laion_aesthetic_v14,
|
|
|
286 |
ddgan_laion2b_v1,
|
287 |
ddgan_sd_v1,
|
288 |
ddgan_sd_v2,
|
@@ -298,7 +321,11 @@ models = [
|
|
298 |
ddgan_laion2b_v2,
|
299 |
ddgan_ddb_v1,
|
300 |
ddgan_ddb_v2,
|
301 |
-
ddgan_ddb_v3
|
|
|
|
|
|
|
|
|
302 |
]
|
303 |
|
304 |
def get_model(model_name):
|
|
|
256 |
cfg['model']['num_timesteps'] = 2
|
257 |
return cfg
|
258 |
|
259 |
+
def ddgan_ddb_v4():
|
260 |
+
cfg = ddgan_ddb_v1()
|
261 |
+
cfg['model']['num_channels_dae'] = 256
|
262 |
+
cfg['model']['num_timesteps'] = 2
|
263 |
+
return cfg
|
264 |
+
|
265 |
+
def ddgan_ddb_v5():
|
266 |
+
cfg = ddgan_ddb_v2()
|
267 |
+
return cfg
|
268 |
+
|
269 |
+
def ddgan_ddb_v6():
|
270 |
+
cfg = ddgan_ddb_v3()
|
271 |
+
return cfg
|
272 |
+
|
273 |
+
def ddgan_ddb_v7():
|
274 |
+
cfg = ddgan_ddb_v1()
|
275 |
+
return cfg
|
276 |
+
|
277 |
+
def ddgan_laion_aesthetic_v15():
|
278 |
+
cfg = ddgan_ddb_v3()
|
279 |
+
return cfg
|
280 |
+
|
281 |
models = [
|
282 |
ddgan_cifar10_cond17, # cifar10, cross attn for discr
|
283 |
ddgan_cifar10_cond18, # cifar10, xl encoder
|
|
|
305 |
ddgan_laion_aesthetic_v12,
|
306 |
ddgan_laion_aesthetic_v13,
|
307 |
ddgan_laion_aesthetic_v14,
|
308 |
+
ddgan_laion_aesthetic_v15,
|
309 |
ddgan_laion2b_v1,
|
310 |
ddgan_sd_v1,
|
311 |
ddgan_sd_v2,
|
|
|
321 |
ddgan_laion2b_v2,
|
322 |
ddgan_ddb_v1,
|
323 |
ddgan_ddb_v2,
|
324 |
+
ddgan_ddb_v3,
|
325 |
+
ddgan_ddb_v4,
|
326 |
+
ddgan_ddb_v5,
|
327 |
+
ddgan_ddb_v6,
|
328 |
+
ddgan_ddb_v7,
|
329 |
]
|
330 |
|
331 |
def get_model(model_name):
|
score_sde/models/discriminator.py
CHANGED
@@ -181,7 +181,7 @@ class SmallCondAttnDiscriminator(nn.Module):
|
|
181 |
hidden_dim=t_emb_dim,
|
182 |
output_dim=t_emb_dim,
|
183 |
act=act,
|
184 |
-
|
185 |
|
186 |
|
187 |
|
@@ -368,7 +368,7 @@ class CondAttnDiscriminator(nn.Module):
|
|
368 |
hidden_dim=t_emb_dim,
|
369 |
output_dim=t_emb_dim,
|
370 |
act=act,
|
371 |
-
|
372 |
|
373 |
self.start_conv = conv2d(nc,ngf*2,1, padding=0)
|
374 |
self.conv1 = DownConvBlock(ngf*2, ngf*4, t_emb_dim = t_emb_dim, downsample = True, act=act)
|
|
|
181 |
hidden_dim=t_emb_dim,
|
182 |
output_dim=t_emb_dim,
|
183 |
act=act,
|
184 |
+
)
|
185 |
|
186 |
|
187 |
|
|
|
368 |
hidden_dim=t_emb_dim,
|
369 |
output_dim=t_emb_dim,
|
370 |
act=act,
|
371 |
+
)
|
372 |
|
373 |
self.start_conv = conv2d(nc,ngf*2,1, padding=0)
|
374 |
self.conv1 = DownConvBlock(ngf*2, ngf*4, t_emb_dim = t_emb_dim, downsample = True, act=act)
|
score_sde/models/layers.py
CHANGED
@@ -559,6 +559,7 @@ class CondAttnBlock(nn.Module):
|
|
559 |
h = h.permute(0,2,1)
|
560 |
h = h.contiguous()
|
561 |
h_new = self.ca(h, cond, mask=mask)
|
|
|
562 |
h_new = h_new.permute(0,2,1)
|
563 |
h_new = h_new.contiguous()
|
564 |
h_new = h_new.view(B, C, H, W)
|
|
|
559 |
h = h.permute(0,2,1)
|
560 |
h = h.contiguous()
|
561 |
h_new = self.ca(h, cond, mask=mask)
|
562 |
+
# print(h_new.min(), h_new.max())
|
563 |
h_new = h_new.permute(0,2,1)
|
564 |
h_new = h_new.contiguous()
|
565 |
h_new = h_new.view(B, C, H, W)
|
score_sde/models/projected_discriminator.py
ADDED
@@ -0,0 +1,783 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from functools import partial
|
2 |
+
import numpy as np
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
|
7 |
+
|
8 |
+
#from pg_modules.blocks import DownBlock, DownBlockPatch, conv2d
|
9 |
+
import functools
|
10 |
+
import torch
|
11 |
+
import torch.nn as nn
|
12 |
+
import torch.nn.functional as F
|
13 |
+
from torch.nn.utils import spectral_norm
|
14 |
+
from . import layers
|
15 |
+
from .layers import CondAttnBlock
|
16 |
+
from .discriminator import *
|
17 |
+
|
18 |
+
|
19 |
+
def conv2d(*args, **kwargs):
|
20 |
+
return spectral_norm(nn.Conv2d(*args, **kwargs))
|
21 |
+
|
22 |
+
|
23 |
+
def convTranspose2d(*args, **kwargs):
|
24 |
+
return spectral_norm(nn.ConvTranspose2d(*args, **kwargs))
|
25 |
+
|
26 |
+
|
27 |
+
def embedding(*args, **kwargs):
|
28 |
+
return spectral_norm(nn.Embedding(*args, **kwargs))
|
29 |
+
|
30 |
+
|
31 |
+
def linear(*args, **kwargs):
|
32 |
+
return spectral_norm(nn.Linear(*args, **kwargs))
|
33 |
+
|
34 |
+
|
35 |
+
def NormLayer(c, mode='batch'):
|
36 |
+
if mode == 'group':
|
37 |
+
return nn.GroupNorm(c//2, c)
|
38 |
+
elif mode == 'batch':
|
39 |
+
return nn.BatchNorm2d(c)
|
40 |
+
|
41 |
+
|
42 |
+
### Activations
|
43 |
+
|
44 |
+
|
45 |
+
class GLU(nn.Module):
|
46 |
+
def forward(self, x):
|
47 |
+
nc = x.size(1)
|
48 |
+
assert nc % 2 == 0, 'channels dont divide 2!'
|
49 |
+
nc = int(nc/2)
|
50 |
+
return x[:, :nc] * torch.sigmoid(x[:, nc:])
|
51 |
+
|
52 |
+
|
53 |
+
class Swish(nn.Module):
|
54 |
+
def forward(self, feat):
|
55 |
+
return feat * torch.sigmoid(feat)
|
56 |
+
|
57 |
+
|
58 |
+
### Upblocks
|
59 |
+
|
60 |
+
|
61 |
+
class InitLayer(nn.Module):
|
62 |
+
def __init__(self, nz, channel, sz=4):
|
63 |
+
super().__init__()
|
64 |
+
|
65 |
+
self.init = nn.Sequential(
|
66 |
+
convTranspose2d(nz, channel*2, sz, 1, 0, bias=False),
|
67 |
+
NormLayer(channel*2),
|
68 |
+
GLU(),
|
69 |
+
)
|
70 |
+
|
71 |
+
def forward(self, noise):
|
72 |
+
noise = noise.view(noise.shape[0], -1, 1, 1)
|
73 |
+
return self.init(noise)
|
74 |
+
|
75 |
+
|
76 |
+
def UpBlockSmall(in_planes, out_planes):
|
77 |
+
block = nn.Sequential(
|
78 |
+
nn.Upsample(scale_factor=2, mode='nearest'),
|
79 |
+
conv2d(in_planes, out_planes*2, 3, 1, 1, bias=False),
|
80 |
+
NormLayer(out_planes*2), GLU())
|
81 |
+
return block
|
82 |
+
|
83 |
+
|
84 |
+
class UpBlockSmallCond(nn.Module):
|
85 |
+
def __init__(self, in_planes, out_planes, z_dim):
|
86 |
+
super().__init__()
|
87 |
+
self.in_planes = in_planes
|
88 |
+
self.out_planes = out_planes
|
89 |
+
self.up = nn.Upsample(scale_factor=2, mode='nearest')
|
90 |
+
self.conv = conv2d(in_planes, out_planes*2, 3, 1, 1, bias=False)
|
91 |
+
|
92 |
+
which_bn = functools.partial(CCBN, which_linear=linear, input_size=z_dim)
|
93 |
+
self.bn = which_bn(2*out_planes)
|
94 |
+
self.act = GLU()
|
95 |
+
|
96 |
+
def forward(self, x, c):
|
97 |
+
x = self.up(x)
|
98 |
+
x = self.conv(x)
|
99 |
+
x = self.bn(x, c)
|
100 |
+
x = self.act(x)
|
101 |
+
return x
|
102 |
+
|
103 |
+
|
104 |
+
def UpBlockBig(in_planes, out_planes):
|
105 |
+
block = nn.Sequential(
|
106 |
+
nn.Upsample(scale_factor=2, mode='nearest'),
|
107 |
+
conv2d(in_planes, out_planes*2, 3, 1, 1, bias=False),
|
108 |
+
NoiseInjection(),
|
109 |
+
NormLayer(out_planes*2), GLU(),
|
110 |
+
conv2d(out_planes, out_planes*2, 3, 1, 1, bias=False),
|
111 |
+
NoiseInjection(),
|
112 |
+
NormLayer(out_planes*2), GLU()
|
113 |
+
)
|
114 |
+
return block
|
115 |
+
|
116 |
+
|
117 |
+
class UpBlockBigCond(nn.Module):
|
118 |
+
def __init__(self, in_planes, out_planes, z_dim):
|
119 |
+
super().__init__()
|
120 |
+
self.in_planes = in_planes
|
121 |
+
self.out_planes = out_planes
|
122 |
+
self.up = nn.Upsample(scale_factor=2, mode='nearest')
|
123 |
+
self.conv1 = conv2d(in_planes, out_planes*2, 3, 1, 1, bias=False)
|
124 |
+
self.conv2 = conv2d(out_planes, out_planes*2, 3, 1, 1, bias=False)
|
125 |
+
|
126 |
+
which_bn = functools.partial(CCBN, which_linear=linear, input_size=z_dim)
|
127 |
+
self.bn1 = which_bn(2*out_planes)
|
128 |
+
self.bn2 = which_bn(2*out_planes)
|
129 |
+
self.act = GLU()
|
130 |
+
self.noise = NoiseInjection()
|
131 |
+
|
132 |
+
def forward(self, x, c):
|
133 |
+
# block 1
|
134 |
+
x = self.up(x)
|
135 |
+
x = self.conv1(x)
|
136 |
+
x = self.noise(x)
|
137 |
+
x = self.bn1(x, c)
|
138 |
+
x = self.act(x)
|
139 |
+
|
140 |
+
# block 2
|
141 |
+
x = self.conv2(x)
|
142 |
+
x = self.noise(x)
|
143 |
+
x = self.bn2(x, c)
|
144 |
+
x = self.act(x)
|
145 |
+
|
146 |
+
return x
|
147 |
+
|
148 |
+
|
149 |
+
class SEBlock(nn.Module):
|
150 |
+
def __init__(self, ch_in, ch_out):
|
151 |
+
super().__init__()
|
152 |
+
self.main = nn.Sequential(
|
153 |
+
nn.AdaptiveAvgPool2d(4),
|
154 |
+
conv2d(ch_in, ch_out, 4, 1, 0, bias=False),
|
155 |
+
Swish(),
|
156 |
+
conv2d(ch_out, ch_out, 1, 1, 0, bias=False),
|
157 |
+
nn.Sigmoid(),
|
158 |
+
)
|
159 |
+
|
160 |
+
def forward(self, feat_small, feat_big):
|
161 |
+
return feat_big * self.main(feat_small)
|
162 |
+
|
163 |
+
|
164 |
+
### Downblocks
|
165 |
+
|
166 |
+
|
167 |
+
class SeparableConv2d(nn.Module):
|
168 |
+
def __init__(self, in_channels, out_channels, kernel_size, bias=False):
|
169 |
+
super(SeparableConv2d, self).__init__()
|
170 |
+
self.depthwise = conv2d(in_channels, in_channels, kernel_size=kernel_size,
|
171 |
+
groups=in_channels, bias=bias, padding=1)
|
172 |
+
self.pointwise = conv2d(in_channels, out_channels,
|
173 |
+
kernel_size=1, bias=bias)
|
174 |
+
|
175 |
+
def forward(self, x):
|
176 |
+
out = self.depthwise(x)
|
177 |
+
out = self.pointwise(out)
|
178 |
+
return out
|
179 |
+
|
180 |
+
|
181 |
+
class DownBlock(nn.Module):
|
182 |
+
def __init__(self, in_planes, out_planes, separable=False):
|
183 |
+
super().__init__()
|
184 |
+
if not separable:
|
185 |
+
self.main = nn.Sequential(
|
186 |
+
conv2d(in_planes, out_planes, 4, 2, 1),
|
187 |
+
NormLayer(out_planes),
|
188 |
+
nn.LeakyReLU(0.2, inplace=True),
|
189 |
+
)
|
190 |
+
else:
|
191 |
+
self.main = nn.Sequential(
|
192 |
+
SeparableConv2d(in_planes, out_planes, 3),
|
193 |
+
NormLayer(out_planes),
|
194 |
+
nn.LeakyReLU(0.2, inplace=True),
|
195 |
+
nn.AvgPool2d(2, 2),
|
196 |
+
)
|
197 |
+
|
198 |
+
def forward(self, feat):
|
199 |
+
return self.main(feat)
|
200 |
+
|
201 |
+
|
202 |
+
class DownBlockPatch(nn.Module):
|
203 |
+
def __init__(self, in_planes, out_planes, separable=False):
|
204 |
+
super().__init__()
|
205 |
+
self.main = nn.Sequential(
|
206 |
+
DownBlock(in_planes, out_planes, separable),
|
207 |
+
conv2d(out_planes, out_planes, 1, 1, 0, bias=False),
|
208 |
+
NormLayer(out_planes),
|
209 |
+
nn.LeakyReLU(0.2, inplace=True),
|
210 |
+
)
|
211 |
+
|
212 |
+
def forward(self, feat):
|
213 |
+
return self.main(feat)
|
214 |
+
|
215 |
+
|
216 |
+
### CSM
|
217 |
+
|
218 |
+
|
219 |
+
class ResidualConvUnit(nn.Module):
|
220 |
+
def __init__(self, cin, activation, bn):
|
221 |
+
super().__init__()
|
222 |
+
self.conv = nn.Conv2d(cin, cin, kernel_size=3, stride=1, padding=1, bias=True)
|
223 |
+
self.skip_add = nn.quantized.FloatFunctional()
|
224 |
+
|
225 |
+
def forward(self, x):
|
226 |
+
return self.skip_add.add(self.conv(x), x)
|
227 |
+
|
228 |
+
|
229 |
+
class FeatureFusionBlock(nn.Module):
|
230 |
+
def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True, lowest=False):
|
231 |
+
super().__init__()
|
232 |
+
|
233 |
+
self.deconv = deconv
|
234 |
+
self.align_corners = align_corners
|
235 |
+
|
236 |
+
self.expand = expand
|
237 |
+
out_features = features
|
238 |
+
if self.expand==True:
|
239 |
+
out_features = features//2
|
240 |
+
|
241 |
+
self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1)
|
242 |
+
self.skip_add = nn.quantized.FloatFunctional()
|
243 |
+
|
244 |
+
def forward(self, *xs):
|
245 |
+
output = xs[0]
|
246 |
+
|
247 |
+
if len(xs) == 2:
|
248 |
+
output = self.skip_add.add(output, xs[1])
|
249 |
+
|
250 |
+
output = nn.functional.interpolate(
|
251 |
+
output, scale_factor=2, mode="bilinear", align_corners=self.align_corners
|
252 |
+
)
|
253 |
+
|
254 |
+
output = self.out_conv(output)
|
255 |
+
|
256 |
+
return output
|
257 |
+
|
258 |
+
|
259 |
+
### Misc
|
260 |
+
|
261 |
+
|
262 |
+
class NoiseInjection(nn.Module):
|
263 |
+
def __init__(self):
|
264 |
+
super().__init__()
|
265 |
+
self.weight = nn.Parameter(torch.zeros(1), requires_grad=True)
|
266 |
+
|
267 |
+
def forward(self, feat, noise=None):
|
268 |
+
if noise is None:
|
269 |
+
batch, _, height, width = feat.shape
|
270 |
+
noise = torch.randn(batch, 1, height, width).to(feat.device)
|
271 |
+
|
272 |
+
return feat + self.weight * noise
|
273 |
+
|
274 |
+
|
275 |
+
class CCBN(nn.Module):
|
276 |
+
''' conditional batchnorm '''
|
277 |
+
def __init__(self, output_size, input_size, which_linear, eps=1e-5, momentum=0.1):
|
278 |
+
super().__init__()
|
279 |
+
self.output_size, self.input_size = output_size, input_size
|
280 |
+
|
281 |
+
# Prepare gain and bias layers
|
282 |
+
self.gain = which_linear(input_size, output_size)
|
283 |
+
self.bias = which_linear(input_size, output_size)
|
284 |
+
|
285 |
+
# epsilon to avoid dividing by 0
|
286 |
+
self.eps = eps
|
287 |
+
# Momentum
|
288 |
+
self.momentum = momentum
|
289 |
+
|
290 |
+
self.register_buffer('stored_mean', torch.zeros(output_size))
|
291 |
+
self.register_buffer('stored_var', torch.ones(output_size))
|
292 |
+
|
293 |
+
def forward(self, x, y):
|
294 |
+
# Calculate class-conditional gains and biases
|
295 |
+
gain = (1 + self.gain(y)).view(y.size(0), -1, 1, 1)
|
296 |
+
bias = self.bias(y).view(y.size(0), -1, 1, 1)
|
297 |
+
out = F.batch_norm(x, self.stored_mean, self.stored_var, None, None,
|
298 |
+
self.training, 0.1, self.eps)
|
299 |
+
return out * gain + bias
|
300 |
+
|
301 |
+
|
302 |
+
class Interpolate(nn.Module):
|
303 |
+
"""Interpolation module."""
|
304 |
+
|
305 |
+
def __init__(self, size, mode='bilinear', align_corners=False):
|
306 |
+
"""Init.
|
307 |
+
Args:
|
308 |
+
scale_factor (float): scaling
|
309 |
+
mode (str): interpolation mode
|
310 |
+
"""
|
311 |
+
super(Interpolate, self).__init__()
|
312 |
+
|
313 |
+
self.interp = nn.functional.interpolate
|
314 |
+
self.size = size
|
315 |
+
self.mode = mode
|
316 |
+
self.align_corners = align_corners
|
317 |
+
|
318 |
+
def forward(self, x):
|
319 |
+
"""Forward pass.
|
320 |
+
Args:
|
321 |
+
x (tensor): input
|
322 |
+
Returns:
|
323 |
+
tensor: interpolated data
|
324 |
+
"""
|
325 |
+
|
326 |
+
x = self.interp(
|
327 |
+
x,
|
328 |
+
size=self.size,
|
329 |
+
mode=self.mode,
|
330 |
+
align_corners=self.align_corners,
|
331 |
+
)
|
332 |
+
|
333 |
+
return x
|
334 |
+
|
335 |
+
|
336 |
+
|
337 |
+
#from pg_modules.projector import F_RandomProj
|
338 |
+
|
339 |
+
import torch
|
340 |
+
import torch.nn as nn
|
341 |
+
import timm
|
342 |
+
#from pg_modules.blocks import FeatureFusionBlock
|
343 |
+
|
344 |
+
|
345 |
+
def _make_scratch_ccm(scratch, in_channels, cout, expand=False):
|
346 |
+
# shapes
|
347 |
+
out_channels = [cout, cout*2, cout*4, cout*8] if expand else [cout]*4
|
348 |
+
|
349 |
+
scratch.layer0_ccm = nn.Conv2d(in_channels[0], out_channels[0], kernel_size=1, stride=1, padding=0, bias=True)
|
350 |
+
scratch.layer1_ccm = nn.Conv2d(in_channels[1], out_channels[1], kernel_size=1, stride=1, padding=0, bias=True)
|
351 |
+
scratch.layer2_ccm = nn.Conv2d(in_channels[2], out_channels[2], kernel_size=1, stride=1, padding=0, bias=True)
|
352 |
+
scratch.layer3_ccm = nn.Conv2d(in_channels[3], out_channels[3], kernel_size=1, stride=1, padding=0, bias=True)
|
353 |
+
|
354 |
+
scratch.CHANNELS = out_channels
|
355 |
+
|
356 |
+
return scratch
|
357 |
+
|
358 |
+
|
359 |
+
def _make_scratch_csm(scratch, in_channels, cout, expand):
|
360 |
+
scratch.layer3_csm = FeatureFusionBlock(in_channels[3], nn.ReLU(False), expand=expand, lowest=True)
|
361 |
+
scratch.layer2_csm = FeatureFusionBlock(in_channels[2], nn.ReLU(False), expand=expand)
|
362 |
+
scratch.layer1_csm = FeatureFusionBlock(in_channels[1], nn.ReLU(False), expand=expand)
|
363 |
+
scratch.layer0_csm = FeatureFusionBlock(in_channels[0], nn.ReLU(False))
|
364 |
+
|
365 |
+
# last refinenet does not expand to save channels in higher dimensions
|
366 |
+
scratch.CHANNELS = [cout, cout, cout*2, cout*4] if expand else [cout]*4
|
367 |
+
|
368 |
+
return scratch
|
369 |
+
|
370 |
+
|
371 |
+
def _make_efficientnet(model):
|
372 |
+
pretrained = nn.Module()
|
373 |
+
pretrained.layer0 = nn.Sequential(model.conv_stem, model.bn1, model.act1, *model.blocks[0:2])
|
374 |
+
pretrained.layer1 = nn.Sequential(*model.blocks[2:3])
|
375 |
+
pretrained.layer2 = nn.Sequential(*model.blocks[3:5])
|
376 |
+
pretrained.layer3 = nn.Sequential(*model.blocks[5:9])
|
377 |
+
return pretrained
|
378 |
+
|
379 |
+
|
380 |
+
def calc_channels(pretrained, inp_res=224):
|
381 |
+
channels = []
|
382 |
+
tmp = torch.zeros(1, 3, inp_res, inp_res)
|
383 |
+
|
384 |
+
# forward pass
|
385 |
+
tmp = pretrained.layer0(tmp)
|
386 |
+
channels.append(tmp.shape[1])
|
387 |
+
tmp = pretrained.layer1(tmp)
|
388 |
+
channels.append(tmp.shape[1])
|
389 |
+
tmp = pretrained.layer2(tmp)
|
390 |
+
channels.append(tmp.shape[1])
|
391 |
+
tmp = pretrained.layer3(tmp)
|
392 |
+
channels.append(tmp.shape[1])
|
393 |
+
|
394 |
+
return channels
|
395 |
+
|
396 |
+
|
397 |
+
def _make_projector(im_res, cout, proj_type, expand=False):
|
398 |
+
assert proj_type in [0, 1, 2], "Invalid projection type"
|
399 |
+
|
400 |
+
### Build pretrained feature network
|
401 |
+
model = timm.create_model('tf_efficientnet_lite0', pretrained=True)
|
402 |
+
pretrained = _make_efficientnet(model)
|
403 |
+
|
404 |
+
# determine resolution of feature maps, this is later used to calculate the number
|
405 |
+
# of down blocks in the discriminators. Interestingly, the best results are achieved
|
406 |
+
# by fixing this to 256, ie., we use the same number of down blocks per discriminator
|
407 |
+
# independent of the dataset resolution
|
408 |
+
im_res = 256
|
409 |
+
pretrained.RESOLUTIONS = [im_res//4, im_res//8, im_res//16, im_res//32]
|
410 |
+
pretrained.CHANNELS = calc_channels(pretrained)
|
411 |
+
|
412 |
+
if proj_type == 0: return pretrained, None
|
413 |
+
|
414 |
+
### Build CCM
|
415 |
+
scratch = nn.Module()
|
416 |
+
scratch = _make_scratch_ccm(scratch, in_channels=pretrained.CHANNELS, cout=cout, expand=expand)
|
417 |
+
pretrained.CHANNELS = scratch.CHANNELS
|
418 |
+
|
419 |
+
if proj_type == 1: return pretrained, scratch
|
420 |
+
|
421 |
+
### build CSM
|
422 |
+
scratch = _make_scratch_csm(scratch, in_channels=scratch.CHANNELS, cout=cout, expand=expand)
|
423 |
+
|
424 |
+
# CSM upsamples x2 so the feature map resolution doubles
|
425 |
+
pretrained.RESOLUTIONS = [res*2 for res in pretrained.RESOLUTIONS]
|
426 |
+
pretrained.CHANNELS = scratch.CHANNELS
|
427 |
+
|
428 |
+
return pretrained, scratch
|
429 |
+
|
430 |
+
|
431 |
+
class F_RandomProj(nn.Module):
|
432 |
+
def __init__(
|
433 |
+
self,
|
434 |
+
im_res=256,
|
435 |
+
cout=64,
|
436 |
+
expand=True,
|
437 |
+
proj_type=2, # 0 = no projection, 1 = cross channel mixing, 2 = cross scale mixing
|
438 |
+
**kwargs,
|
439 |
+
):
|
440 |
+
super().__init__()
|
441 |
+
self.proj_type = proj_type
|
442 |
+
self.cout = cout
|
443 |
+
self.expand = expand
|
444 |
+
|
445 |
+
# build pretrained feature network and random decoder (scratch)
|
446 |
+
self.pretrained, self.scratch = _make_projector(im_res=im_res, cout=self.cout, proj_type=self.proj_type, expand=self.expand)
|
447 |
+
self.CHANNELS = self.pretrained.CHANNELS
|
448 |
+
self.RESOLUTIONS = self.pretrained.RESOLUTIONS
|
449 |
+
|
450 |
+
def forward(self, x):
|
451 |
+
# predict feature maps
|
452 |
+
out0 = self.pretrained.layer0(x)
|
453 |
+
out1 = self.pretrained.layer1(out0)
|
454 |
+
out2 = self.pretrained.layer2(out1)
|
455 |
+
out3 = self.pretrained.layer3(out2)
|
456 |
+
|
457 |
+
# start enumerating at the lowest layer (this is where we put the first discriminator)
|
458 |
+
out = {
|
459 |
+
'0': out0,
|
460 |
+
'1': out1,
|
461 |
+
'2': out2,
|
462 |
+
'3': out3,
|
463 |
+
}
|
464 |
+
|
465 |
+
if self.proj_type == 0: return out
|
466 |
+
|
467 |
+
out0_channel_mixed = self.scratch.layer0_ccm(out['0'])
|
468 |
+
out1_channel_mixed = self.scratch.layer1_ccm(out['1'])
|
469 |
+
out2_channel_mixed = self.scratch.layer2_ccm(out['2'])
|
470 |
+
out3_channel_mixed = self.scratch.layer3_ccm(out['3'])
|
471 |
+
|
472 |
+
out = {
|
473 |
+
'0': out0_channel_mixed,
|
474 |
+
'1': out1_channel_mixed,
|
475 |
+
'2': out2_channel_mixed,
|
476 |
+
'3': out3_channel_mixed,
|
477 |
+
}
|
478 |
+
|
479 |
+
if self.proj_type == 1: return out
|
480 |
+
|
481 |
+
# from bottom to top
|
482 |
+
out3_scale_mixed = self.scratch.layer3_csm(out3_channel_mixed)
|
483 |
+
out2_scale_mixed = self.scratch.layer2_csm(out3_scale_mixed, out2_channel_mixed)
|
484 |
+
out1_scale_mixed = self.scratch.layer1_csm(out2_scale_mixed, out1_channel_mixed)
|
485 |
+
out0_scale_mixed = self.scratch.layer0_csm(out1_scale_mixed, out0_channel_mixed)
|
486 |
+
|
487 |
+
out = {
|
488 |
+
'0': out0_scale_mixed,
|
489 |
+
'1': out1_scale_mixed,
|
490 |
+
'2': out2_scale_mixed,
|
491 |
+
'3': out3_scale_mixed,
|
492 |
+
}
|
493 |
+
|
494 |
+
return out
|
495 |
+
|
496 |
+
|
497 |
+
#from pg_modules.diffaug import DiffAugment
|
498 |
+
# Differentiable Augmentation for Data-Efficient GAN Training
|
499 |
+
# Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han
|
500 |
+
# https://arxiv.org/pdf/2006.10738
|
501 |
+
|
502 |
+
import torch
|
503 |
+
import torch.nn.functional as F
|
504 |
+
|
505 |
+
|
506 |
+
def DiffAugment(x, policy='', channels_first=True):
|
507 |
+
if policy:
|
508 |
+
if not channels_first:
|
509 |
+
x = x.permute(0, 3, 1, 2)
|
510 |
+
for p in policy.split(','):
|
511 |
+
for f in AUGMENT_FNS[p]:
|
512 |
+
x = f(x)
|
513 |
+
if not channels_first:
|
514 |
+
x = x.permute(0, 2, 3, 1)
|
515 |
+
x = x.contiguous()
|
516 |
+
return x
|
517 |
+
|
518 |
+
|
519 |
+
def rand_brightness(x):
|
520 |
+
x = x + (torch.rand(x.size(0), 1, 1, 1, dtype=x.dtype, device=x.device) - 0.5)
|
521 |
+
return x
|
522 |
+
|
523 |
+
|
524 |
+
def rand_saturation(x):
|
525 |
+
x_mean = x.mean(dim=1, keepdim=True)
|
526 |
+
x = (x - x_mean) * (torch.rand(x.size(0), 1, 1, 1, dtype=x.dtype, device=x.device) * 2) + x_mean
|
527 |
+
return x
|
528 |
+
|
529 |
+
|
530 |
+
def rand_contrast(x):
|
531 |
+
x_mean = x.mean(dim=[1, 2, 3], keepdim=True)
|
532 |
+
x = (x - x_mean) * (torch.rand(x.size(0), 1, 1, 1, dtype=x.dtype, device=x.device) + 0.5) + x_mean
|
533 |
+
return x
|
534 |
+
|
535 |
+
|
536 |
+
def rand_translation(x, ratio=0.125):
|
537 |
+
shift_x, shift_y = int(x.size(2) * ratio + 0.5), int(x.size(3) * ratio + 0.5)
|
538 |
+
translation_x = torch.randint(-shift_x, shift_x + 1, size=[x.size(0), 1, 1], device=x.device)
|
539 |
+
translation_y = torch.randint(-shift_y, shift_y + 1, size=[x.size(0), 1, 1], device=x.device)
|
540 |
+
grid_batch, grid_x, grid_y = torch.meshgrid(
|
541 |
+
torch.arange(x.size(0), dtype=torch.long, device=x.device),
|
542 |
+
torch.arange(x.size(2), dtype=torch.long, device=x.device),
|
543 |
+
torch.arange(x.size(3), dtype=torch.long, device=x.device),
|
544 |
+
)
|
545 |
+
grid_x = torch.clamp(grid_x + translation_x + 1, 0, x.size(2) + 1)
|
546 |
+
grid_y = torch.clamp(grid_y + translation_y + 1, 0, x.size(3) + 1)
|
547 |
+
x_pad = F.pad(x, [1, 1, 1, 1, 0, 0, 0, 0])
|
548 |
+
x = x_pad.permute(0, 2, 3, 1).contiguous()[grid_batch, grid_x, grid_y].permute(0, 3, 1, 2)
|
549 |
+
return x
|
550 |
+
|
551 |
+
|
552 |
+
def rand_cutout(x, ratio=0.2):
|
553 |
+
cutout_size = int(x.size(2) * ratio + 0.5), int(x.size(3) * ratio + 0.5)
|
554 |
+
offset_x = torch.randint(0, x.size(2) + (1 - cutout_size[0] % 2), size=[x.size(0), 1, 1], device=x.device)
|
555 |
+
offset_y = torch.randint(0, x.size(3) + (1 - cutout_size[1] % 2), size=[x.size(0), 1, 1], device=x.device)
|
556 |
+
grid_batch, grid_x, grid_y = torch.meshgrid(
|
557 |
+
torch.arange(x.size(0), dtype=torch.long, device=x.device),
|
558 |
+
torch.arange(cutout_size[0], dtype=torch.long, device=x.device),
|
559 |
+
torch.arange(cutout_size[1], dtype=torch.long, device=x.device),
|
560 |
+
)
|
561 |
+
grid_x = torch.clamp(grid_x + offset_x - cutout_size[0] // 2, min=0, max=x.size(2) - 1)
|
562 |
+
grid_y = torch.clamp(grid_y + offset_y - cutout_size[1] // 2, min=0, max=x.size(3) - 1)
|
563 |
+
mask = torch.ones(x.size(0), x.size(2), x.size(3), dtype=x.dtype, device=x.device)
|
564 |
+
mask[grid_batch, grid_x, grid_y] = 0
|
565 |
+
x = x * mask.unsqueeze(1)
|
566 |
+
return x
|
567 |
+
|
568 |
+
|
569 |
+
AUGMENT_FNS = {
|
570 |
+
'color': [rand_brightness, rand_saturation, rand_contrast],
|
571 |
+
'translation': [rand_translation],
|
572 |
+
'cutout': [rand_cutout],
|
573 |
+
}
|
574 |
+
|
575 |
+
|
576 |
+
|
577 |
+
class SingleDisc(nn.Module):
|
578 |
+
def __init__(self, nc=None, ndf=None, start_sz=256, end_sz=8, head=None, separable=False, patch=False):
|
579 |
+
super().__init__()
|
580 |
+
channel_dict = {4: 512, 8: 512, 16: 256, 32: 128, 64: 64, 128: 64,
|
581 |
+
256: 32, 512: 16, 1024: 8}
|
582 |
+
|
583 |
+
# interpolate for start sz that are not powers of two
|
584 |
+
if start_sz not in channel_dict.keys():
|
585 |
+
sizes = np.array(list(channel_dict.keys()))
|
586 |
+
start_sz = sizes[np.argmin(abs(sizes - start_sz))]
|
587 |
+
self.start_sz = start_sz
|
588 |
+
|
589 |
+
# if given ndf, allocate all layers with the same ndf
|
590 |
+
if ndf is None:
|
591 |
+
nfc = channel_dict
|
592 |
+
else:
|
593 |
+
nfc = {k: ndf for k, v in channel_dict.items()}
|
594 |
+
|
595 |
+
# for feature map discriminators with nfc not in channel_dict
|
596 |
+
# this is the case for the pretrained backbone (midas.pretrained)
|
597 |
+
if nc is not None and head is None:
|
598 |
+
nfc[start_sz] = nc
|
599 |
+
|
600 |
+
layers = []
|
601 |
+
|
602 |
+
# Head if the initial input is the full modality
|
603 |
+
if head:
|
604 |
+
layers += [conv2d(nc, nfc[256], 3, 1, 1, bias=False),
|
605 |
+
nn.LeakyReLU(0.2, inplace=True)]
|
606 |
+
|
607 |
+
# Down Blocks
|
608 |
+
DB = partial(DownBlockPatch, separable=separable) if patch else partial(DownBlock, separable=separable)
|
609 |
+
while start_sz > end_sz:
|
610 |
+
layers.append(DB(nfc[start_sz], nfc[start_sz//2]))
|
611 |
+
start_sz = start_sz // 2
|
612 |
+
|
613 |
+
layers.append(conv2d(nfc[end_sz], 1, 4, 1, 0, bias=False))
|
614 |
+
self.main = nn.Sequential(*layers)
|
615 |
+
|
616 |
+
def forward(self, x, c):
|
617 |
+
return self.main(x)
|
618 |
+
|
619 |
+
|
620 |
+
class SingleDiscCond(nn.Module):
|
621 |
+
def __init__(self, nc=None, ndf=None, start_sz=256, end_sz=8, head=None, separable=False, patch=False, c_dim=1000, cmap_dim=64, embedding_dim=128, cond_size=128):
|
622 |
+
super().__init__()
|
623 |
+
self.cmap_dim = cmap_dim
|
624 |
+
self.cond_attn = CondAttnBlock(cmap_dim, cond_size, dim_head=64, heads=8, norm_context=False, cosine_sim_attn=False)
|
625 |
+
# midas channels
|
626 |
+
channel_dict = {4: 512, 8: 512, 16: 256, 32: 128, 64: 64, 128: 64,
|
627 |
+
256: 32, 512: 16, 1024: 8}
|
628 |
+
|
629 |
+
# interpolate for start sz that are not powers of two
|
630 |
+
if start_sz not in channel_dict.keys():
|
631 |
+
sizes = np.array(list(channel_dict.keys()))
|
632 |
+
start_sz = sizes[np.argmin(abs(sizes - start_sz))]
|
633 |
+
self.start_sz = start_sz
|
634 |
+
|
635 |
+
# if given ndf, allocate all layers with the same ndf
|
636 |
+
if ndf is None:
|
637 |
+
nfc = channel_dict
|
638 |
+
else:
|
639 |
+
nfc = {k: ndf for k, v in channel_dict.items()}
|
640 |
+
|
641 |
+
# for feature map discriminators with nfc not in channel_dict
|
642 |
+
# this is the case for the pretrained backbone (midas.pretrained)
|
643 |
+
if nc is not None and head is None:
|
644 |
+
nfc[start_sz] = nc
|
645 |
+
|
646 |
+
layers = []
|
647 |
+
|
648 |
+
# Head if the initial input is the full modality
|
649 |
+
if head:
|
650 |
+
layers += [conv2d(nc, nfc[256], 3, 1, 1, bias=False),
|
651 |
+
nn.LeakyReLU(0.2, inplace=True)]
|
652 |
+
|
653 |
+
# Down Blocks
|
654 |
+
DB = partial(DownBlockPatch, separable=separable) if patch else partial(DownBlock, separable=separable)
|
655 |
+
while start_sz > end_sz:
|
656 |
+
layers.append(DB(nfc[start_sz], nfc[start_sz//2]))
|
657 |
+
start_sz = start_sz // 2
|
658 |
+
self.main = nn.Sequential(*layers)
|
659 |
+
|
660 |
+
# additions for conditioning on class information
|
661 |
+
self.cls = conv2d(nfc[end_sz], self.cmap_dim, 4, 1, 0, bias=False)
|
662 |
+
#self.embed = nn.Embedding(num_embeddings=c_dim, embedding_dim=embedding_dim)
|
663 |
+
#self.embed_proj = nn.Sequential(
|
664 |
+
# nn.Linear(self.embed.embedding_dim, self.cmap_dim),
|
665 |
+
# nn.LeakyReLU(0.2, inplace=True),
|
666 |
+
#)
|
667 |
+
|
668 |
+
def forward(self, x, c):
|
669 |
+
h = self.main(x)
|
670 |
+
out = self.cls(h)
|
671 |
+
cond_pooled, cond, cond_mask = c
|
672 |
+
#print("COND", out.shape, cond.shape, cond_mask.shape, self.cond_sie)
|
673 |
+
cmap = self.cond_attn(out, cond, cond_mask)
|
674 |
+
# conditioning via projection
|
675 |
+
#cmap = self.embed_proj(self.embed(c)).unsqueeze(-1).unsqueeze(-1)
|
676 |
+
#cmap = 1
|
677 |
+
out = (out * cmap).sum(dim=1, keepdim=True) * (1 / np.sqrt(self.cmap_dim))
|
678 |
+
return out
|
679 |
+
|
680 |
+
|
681 |
+
class MultiScaleD(nn.Module):
|
682 |
+
def __init__(
|
683 |
+
self,
|
684 |
+
channels,
|
685 |
+
resolutions,
|
686 |
+
num_discs=1,
|
687 |
+
proj_type=2, # 0 = no projection, 1 = cross channel mixing, 2 = cross scale mixing
|
688 |
+
cond=1,
|
689 |
+
separable=False,
|
690 |
+
patch=False,
|
691 |
+
cond_size=128,
|
692 |
+
**kwargs,
|
693 |
+
):
|
694 |
+
super().__init__()
|
695 |
+
|
696 |
+
assert num_discs in [1, 2, 3, 4]
|
697 |
+
|
698 |
+
# the first disc is on the lowest level of the backbone
|
699 |
+
self.disc_in_channels = channels[:num_discs]
|
700 |
+
self.disc_in_res = resolutions[:num_discs]
|
701 |
+
|
702 |
+
Disc = SingleDiscCond if cond else SingleDisc
|
703 |
+
mini_discs = []
|
704 |
+
for i, (cin, res) in enumerate(zip(self.disc_in_channels, self.disc_in_res)):
|
705 |
+
start_sz = res if not patch else 16
|
706 |
+
mini_discs += [str(i), Disc(nc=cin, start_sz=start_sz, end_sz=8, separable=separable, patch=patch, cond_size=cond_size)],
|
707 |
+
self.mini_discs = nn.ModuleDict(mini_discs)
|
708 |
+
|
709 |
+
def forward(self, features, c):
|
710 |
+
all_logits = []
|
711 |
+
for k, disc in self.mini_discs.items():
|
712 |
+
all_logits.append(disc(features[k], c).view(features[k].size(0), -1))
|
713 |
+
|
714 |
+
all_logits = torch.cat(all_logits, dim=1)
|
715 |
+
return all_logits
|
716 |
+
|
717 |
+
|
718 |
+
class ProjectedDiscriminator(torch.nn.Module):
|
719 |
+
def __init__(
|
720 |
+
self,
|
721 |
+
diffaug=False,
|
722 |
+
interp224=False,
|
723 |
+
t_emb_dim = 128,
|
724 |
+
out_dim=64,
|
725 |
+
backbone_kwargs={},
|
726 |
+
act=torch.nn.LeakyReLU(0.2),
|
727 |
+
num_discs=1,
|
728 |
+
**kwargs
|
729 |
+
):
|
730 |
+
super().__init__()
|
731 |
+
self.diffaug = diffaug
|
732 |
+
self.act = act
|
733 |
+
self.interp224 = interp224
|
734 |
+
self.num_discs = num_discs
|
735 |
+
self.feature_network = F_RandomProj(**backbone_kwargs)
|
736 |
+
self.discriminator = MultiScaleD(
|
737 |
+
channels=[c*2+out_dim for c in self.feature_network.CHANNELS],
|
738 |
+
resolutions=self.feature_network.RESOLUTIONS,
|
739 |
+
**backbone_kwargs,
|
740 |
+
)
|
741 |
+
self.t_embed = torch.nn.ModuleList([TimestepEmbedding(
|
742 |
+
embedding_dim=t_emb_dim,
|
743 |
+
hidden_dim=t_emb_dim,
|
744 |
+
output_dim=out_dim,
|
745 |
+
act=act,
|
746 |
+
) for _ in range(num_discs)])
|
747 |
+
|
748 |
+
|
749 |
+
def train(self, mode=True):
|
750 |
+
self.feature_network = self.feature_network.train(False)
|
751 |
+
self.discriminator = self.discriminator.train(mode)
|
752 |
+
return self
|
753 |
+
|
754 |
+
def eval(self):
|
755 |
+
return self.train(False)
|
756 |
+
|
757 |
+
def forward(self, x, t, xprev, cond=None):
|
758 |
+
#t_embed = self.t_embed(t)
|
759 |
+
#t_embed = self.act(t_embed)
|
760 |
+
|
761 |
+
if self.diffaug:
|
762 |
+
x = DiffAugment(x, policy='color,translation,cutout')
|
763 |
+
|
764 |
+
if self.interp224:
|
765 |
+
x = F.interpolate(x, 256, mode='bilinear', align_corners=False)
|
766 |
+
|
767 |
+
features1 = self.feature_network(x)
|
768 |
+
features2 = self.feature_network(xprev)
|
769 |
+
features = {}
|
770 |
+
for k in features1.keys():
|
771 |
+
if int(k) >= self.num_discs:
|
772 |
+
continue
|
773 |
+
tcat = self.t_embed[int(k)](t)
|
774 |
+
#print(tcat.shape)
|
775 |
+
h, w = features1[k].shape[2:]
|
776 |
+
tcat = tcat.view(tcat.shape[0], tcat.shape[1], 1, 1).repeat(1,1, h, w)
|
777 |
+
#print(x.shape, xprev.shape, features1[k].shape, features2[k].shape, tcat.shape)
|
778 |
+
features[k] = torch.cat((features1[k], features2[k], tcat), dim=1)
|
779 |
+
#print(features[k].shape)
|
780 |
+
logits = self.discriminator(features, cond)
|
781 |
+
|
782 |
+
return logits
|
783 |
+
|
scripts/init.sh
CHANGED
@@ -1,2 +1,14 @@
|
|
1 |
-
|
2 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
ml purge
|
2 |
+
ml use $OTHERSTAGES
|
3 |
+
ml Stages/2022
|
4 |
+
ml GCC/11.2.0
|
5 |
+
ml OpenMPI/4.1.2
|
6 |
+
ml CUDA/11.5
|
7 |
+
ml cuDNN/8.3.1.22-CUDA-11.5
|
8 |
+
ml NCCL/2.12.7-1-CUDA-11.5
|
9 |
+
ml PyTorch/1.11-CUDA-11.5
|
10 |
+
ml Horovod/0.24
|
11 |
+
ml torchvision/0.12.0
|
12 |
+
source /p/home/jusers/cherti1/jureca/ccstdl/code/feed_forward_vqgan_clip/envs/jureca_2022/bin/activate
|
13 |
+
export HOROVOD_CACHE_CAPACITY=4096
|
14 |
+
export CUDA_VISIBLE_DEVICES=0,1,2,3
|
scripts/run_jurecadc_conda.sh
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/bash -x
|
2 |
+
#SBATCH --account=zam
|
3 |
+
#SBATCH --nodes=1
|
4 |
+
#SBATCH --ntasks-per-node=4
|
5 |
+
#SBATCH --cpus-per-task=24
|
6 |
+
#SBATCH --time=06:00:00
|
7 |
+
#SBATCH --gres=gpu:4
|
8 |
+
#SBATCH --partition=dc-gpu
|
9 |
+
ml CUDA
|
10 |
+
source /p/project/laionize/miniconda/bin/activate
|
11 |
+
conda activate ddgan
|
12 |
+
#source scripts/init_2022.sh
|
13 |
+
#source scripts/init_2020.sh
|
14 |
+
#source scripts/init.sh
|
15 |
+
export CUDA_VISIBLE_DEVICES=0,1,2,3
|
16 |
+
echo "Job id: $SLURM_JOB_ID"
|
17 |
+
export TOKENIZERS_PARALLELISM=false
|
18 |
+
#export NCCL_ASYNC_ERROR_HANDLING=1
|
19 |
+
export NCCL_IB_TIMEOUT=50
|
20 |
+
export UCX_RC_TIMEOUT=4s
|
21 |
+
export NCCL_IB_RETRY_CNT=10
|
22 |
+
export TORCH_DISTRIBUTED_DEBUG=INFO
|
23 |
+
srun python -u $*
|
scripts/run_juwelsbooster_conda.sh
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/bash -x
|
2 |
+
#SBATCH --account=laionize
|
3 |
+
#SBATCH --nodes=1
|
4 |
+
#SBATCH --ntasks-per-node=4
|
5 |
+
#SBATCH --cpus-per-task=24
|
6 |
+
#SBATCH --time=06:00:00
|
7 |
+
#SBATCH --gres=gpu:4
|
8 |
+
#SBATCH --partition=booster
|
9 |
+
ml CUDA
|
10 |
+
source /p/project/laionize/miniconda/bin/activate
|
11 |
+
conda activate ddgan
|
12 |
+
export CUDA_VISIBLE_DEVICES=0,1,2,3
|
13 |
+
echo "Job id: $SLURM_JOB_ID"
|
14 |
+
export TOKENIZERS_PARALLELISM=false
|
15 |
+
#export NCCL_ASYNC_ERROR_HANDLING=1
|
16 |
+
export NCCL_IB_TIMEOUT=50
|
17 |
+
export UCX_RC_TIMEOUT=4s
|
18 |
+
export NCCL_IB_RETRY_CNT=10
|
19 |
+
srun python -u $*
|
test.py
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from score_sde.models.projected_discriminator import ProjectedDiscriminator
|
2 |
+
import torch
|
3 |
+
discr = ProjectedDiscriminator(num_discs=4, backbone_kwargs={"cond_size": 768})
|
4 |
+
x = torch.randn(1,3,224,224)
|
5 |
+
t = torch.randint(0, 1, size=(1,))
|
6 |
+
cond = (None, torch.randn(1,77, 768), torch.ones(1,77, dtype=torch.bool))
|
7 |
+
y = discr(x, t, x, cond=cond)
|
8 |
+
print(y.shape)
|
test_ddgan.py
CHANGED
@@ -384,15 +384,20 @@ def sample_and_test(args):
|
|
384 |
for epoch in epochs:
|
385 |
args.epoch_id = epoch
|
386 |
path = './saved_info/dd_gan/{}/{}/netG_{}.pth'.format(args.dataset, args.exp, args.epoch_id)
|
387 |
-
|
388 |
if not os.path.exists(path):
|
389 |
continue
|
|
|
|
|
390 |
print(path)
|
391 |
|
392 |
#if not os.path.exists(next_path):
|
393 |
# print(f"STOP at {epoch}")
|
394 |
# break
|
395 |
-
|
|
|
|
|
|
|
396 |
suffix = '_' + args.eval_name if args.eval_name else ""
|
397 |
dest = './saved_info/dd_gan/{}/{}/eval_{}{}.json'.format(args.dataset, args.exp, args.epoch_id, suffix)
|
398 |
next_dest = './saved_info/dd_gan/{}/{}/eval_{}{}.json'.format(args.dataset, args.exp, args.epoch_id+1, suffix)
|
|
|
384 |
for epoch in epochs:
|
385 |
args.epoch_id = epoch
|
386 |
path = './saved_info/dd_gan/{}/{}/netG_{}.pth'.format(args.dataset, args.exp, args.epoch_id)
|
387 |
+
next_next_path = './saved_info/dd_gan/{}/{}/netG_{}.pth'.format(args.dataset, args.exp, args.epoch_id+2)
|
388 |
if not os.path.exists(path):
|
389 |
continue
|
390 |
+
if not os.path.exists(next_next_path):
|
391 |
+
break
|
392 |
print(path)
|
393 |
|
394 |
#if not os.path.exists(next_path):
|
395 |
# print(f"STOP at {epoch}")
|
396 |
# break
|
397 |
+
try:
|
398 |
+
ckpt = torch.load(path, map_location=device)
|
399 |
+
except Exception:
|
400 |
+
continue
|
401 |
suffix = '_' + args.eval_name if args.eval_name else ""
|
402 |
dest = './saved_info/dd_gan/{}/{}/eval_{}{}.json'.format(args.dataset, args.exp, args.epoch_id, suffix)
|
403 |
next_dest = './saved_info/dd_gan/{}/{}/eval_{}{}.json'.format(args.dataset, args.exp, args.epoch_id+1, suffix)
|
train_ddgan.py
CHANGED
@@ -210,6 +210,7 @@ def get_autocast(precision):
|
|
210 |
|
211 |
def train(rank, gpu, args):
|
212 |
from score_sde.models.discriminator import Discriminator_small, Discriminator_large, CondAttnDiscriminator, SmallCondAttnDiscriminator
|
|
|
213 |
from score_sde.models.ncsnpp_generator_adagn import NCSNpp
|
214 |
from EMA import EMA
|
215 |
|
@@ -281,6 +282,12 @@ def train(rank, gpu, args):
|
|
281 |
transforms.ToTensor(),
|
282 |
transforms.Normalize((0.5,0.5,0.5), (0.5,0.5,0.5))
|
283 |
])
|
|
|
|
|
|
|
|
|
|
|
|
|
284 |
shards = glob(os.path.join(args.dataset_root, "*.tar")) if os.path.isdir(args.dataset_root) else args.dataset_root
|
285 |
pipeline = [ResampledShards2(shards)]
|
286 |
pipeline.extend([
|
@@ -295,7 +302,7 @@ def train(rank, gpu, args):
|
|
295 |
pipeline.extend([
|
296 |
wds.select(filter_no_caption),
|
297 |
wds.decode("pilrgb", handler=log_and_continue),
|
298 |
-
wds.rename(image="jpg;png"),
|
299 |
wds.map_dict(image=train_transform),
|
300 |
wds.to_tuple("image","txt"),
|
301 |
wds.batched(batch_size, partial=False),
|
@@ -361,7 +368,13 @@ def train(rank, gpu, args):
|
|
361 |
t_emb_dim = args.t_emb_dim,
|
362 |
cond_size=text_encoder.output_size,
|
363 |
act=nn.LeakyReLU(0.2)).to(device)
|
364 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
365 |
broadcast_params(netG.parameters())
|
366 |
broadcast_params(netD.parameters())
|
367 |
|
@@ -387,7 +400,10 @@ def train(rank, gpu, args):
|
|
387 |
netD = nn.parallel.DistributedDataParallel(netD, device_ids=[gpu])
|
388 |
else:
|
389 |
netG = nn.parallel.DistributedDataParallel(netG, device_ids=[gpu])
|
390 |
-
netD = nn.parallel.DistributedDataParallel(netD, device_ids=[gpu])
|
|
|
|
|
|
|
391 |
|
392 |
if args.grad_checkpointing:
|
393 |
from fairscale.nn.checkpoint.checkpoint_activations import checkpoint_wrapper
|
@@ -430,7 +446,7 @@ def train(rank, gpu, args):
|
|
430 |
.format(checkpoint['epoch']))
|
431 |
else:
|
432 |
global_step, epoch, init_epoch = 0, 0, 0
|
433 |
-
use_cond_attn_discr = args.discr_type in ("large_cond_attn", "small_cond_attn", "large_attn_pool")
|
434 |
for epoch in range(init_epoch, args.num_epoch+1):
|
435 |
if args.dataset == "wds":
|
436 |
os.environ["WDS_EPOCH"] = str(epoch)
|
|
|
210 |
|
211 |
def train(rank, gpu, args):
|
212 |
from score_sde.models.discriminator import Discriminator_small, Discriminator_large, CondAttnDiscriminator, SmallCondAttnDiscriminator
|
213 |
+
from score_sde.models.projected_discriminator import ProjectedDiscriminator
|
214 |
from score_sde.models.ncsnpp_generator_adagn import NCSNpp
|
215 |
from EMA import EMA
|
216 |
|
|
|
282 |
transforms.ToTensor(),
|
283 |
transforms.Normalize((0.5,0.5,0.5), (0.5,0.5,0.5))
|
284 |
])
|
285 |
+
elif args.preprocessing == "simple_random_crop":
|
286 |
+
train_transform = transforms.Compose([
|
287 |
+
transforms.RandomCrop(args.image_size, interpolation=3),
|
288 |
+
transforms.ToTensor(),
|
289 |
+
transforms.Normalize((0.5,0.5,0.5), (0.5,0.5,0.5))
|
290 |
+
])
|
291 |
shards = glob(os.path.join(args.dataset_root, "*.tar")) if os.path.isdir(args.dataset_root) else args.dataset_root
|
292 |
pipeline = [ResampledShards2(shards)]
|
293 |
pipeline.extend([
|
|
|
302 |
pipeline.extend([
|
303 |
wds.select(filter_no_caption),
|
304 |
wds.decode("pilrgb", handler=log_and_continue),
|
305 |
+
wds.rename(image="jpg;png;webp"),
|
306 |
wds.map_dict(image=train_transform),
|
307 |
wds.to_tuple("image","txt"),
|
308 |
wds.batched(batch_size, partial=False),
|
|
|
368 |
t_emb_dim = args.t_emb_dim,
|
369 |
cond_size=text_encoder.output_size,
|
370 |
act=nn.LeakyReLU(0.2)).to(device)
|
371 |
+
elif args.discr_type == "projected_gan":
|
372 |
+
netD = ProjectedDiscriminator(
|
373 |
+
num_discs=4,
|
374 |
+
backbone_kwargs={"cond_size": text_encoder.output_size}
|
375 |
+
)
|
376 |
+
netD = netD.to(device)
|
377 |
+
|
378 |
broadcast_params(netG.parameters())
|
379 |
broadcast_params(netD.parameters())
|
380 |
|
|
|
400 |
netD = nn.parallel.DistributedDataParallel(netD, device_ids=[gpu])
|
401 |
else:
|
402 |
netG = nn.parallel.DistributedDataParallel(netG, device_ids=[gpu])
|
403 |
+
netD = nn.parallel.DistributedDataParallel(netD, device_ids=[gpu], find_unused_parameters=args.discr_type=="projected_gan")
|
404 |
+
#if args.discr_type == "projected_gan":
|
405 |
+
# netD._set_static_graph()
|
406 |
+
|
407 |
|
408 |
if args.grad_checkpointing:
|
409 |
from fairscale.nn.checkpoint.checkpoint_activations import checkpoint_wrapper
|
|
|
446 |
.format(checkpoint['epoch']))
|
447 |
else:
|
448 |
global_step, epoch, init_epoch = 0, 0, 0
|
449 |
+
use_cond_attn_discr = args.discr_type in ("large_cond_attn", "small_cond_attn", "large_attn_pool", "projected_gan")
|
450 |
for epoch in range(init_epoch, args.num_epoch+1):
|
451 |
if args.dataset == "wds":
|
452 |
os.environ["WDS_EPOCH"] = str(epoch)
|