-
Notifications
You must be signed in to change notification settings - Fork 0
/
AEModels.py
210 lines (183 loc) · 10.4 KB
/
AEModels.py
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
207
208
#Source code for 'Domain Transformer: Predicting Samples of Unseen, Future Domains' by Johannes Schneider, IJCNN, 2022, https://arxiv.org/abs/2106.06057; Github; https://github.com/JohnTailor/DoTra
#Licence: Use it however you like, but cite the paper :-)
#Autoencoder models and training
import numpy as np
import pickle,os,copy
import torch.optim as optim
import torch.cuda.amp as tca
import torch
from torch import nn
from torch.nn import functional as F
from torch.autograd import Variable
import torch.nn.functional as F
from aecyc.OnlineRep import imgutil
class AEDisc(nn.Module):
def __init__(self, cfg, input_size=(1, 32, 32)):
super(AEDisc, self).__init__()
output_size = 1
self.input_size = input_size
self.channel_mult = int(64 * cfg["netSi"])
bn = lambda x: nn.BatchNorm2d(x) if cfg["aeganbn"] else nn.Identity()
bn1d = lambda x: nn.BatchNorm1d(x) if cfg["aeganbn"] else nn.Identity()
slope = 0.2
self.conv = nn.Sequential(*[nn.Conv2d(in_channels=input_size[0], out_channels=self.channel_mult * 1, kernel_size=4, stride=2, padding=1), bn(self.channel_mult * 1), nn.LeakyReLU(slope, True),
nn.Conv2d(self.channel_mult * 1, self.channel_mult * 2, 4, 2, 1), bn(self.channel_mult * 2), nn.LeakyReLU(slope, inplace=True),
nn.Conv2d(self.channel_mult * 2, self.channel_mult * 4, 4, 2, 1), bn(self.channel_mult * 4), nn.LeakyReLU(slope, inplace=True),
nn.Conv2d(self.channel_mult * 4, self.channel_mult * 8, 4, 2, 1), bn(self.channel_mult * 8), nn.LeakyReLU(slope, inplace=True),
nn.Conv2d(self.channel_mult * 8, self.channel_mult * 8, 3, 2, 1), bn(self.channel_mult * 8), nn.LeakyReLU(slope, inplace=True)])
# self.flat_fts = self.get_flat_fts(self.conv)
self.nin = self.channel_mult * 8
self.linear = nn.Sequential(nn.Linear(self.nin, output_size) )
def forward(self, x):
for il, l in enumerate(self.conv):
x = l(x)
x = torch.flatten(x, start_dim=1) # x.view(-1, self.flat_fts)
return self.linear(x)
class CNN_Encoder(nn.Module):
def __init__(self, cfg, input_size=(1, 32, 32)):
super(CNN_Encoder, self).__init__()
output_size=cfg["aez"]
self.input_size = input_size
self.channel_mult = int(64*(cfg["netSi"]+0.25*int("F" in cfg["ds"][0])))
bn = lambda x: nn.BatchNorm2d(x) if cfg["aebn"] else nn.Identity()
bn1d = lambda x: nn.BatchNorm1d(x) if cfg["aebn"] else nn.Identity()
slope=0.05
self.conv = nn.Sequential(*[nn.Conv2d(in_channels=input_size[0],out_channels=self.channel_mult*1,kernel_size=4,stride=2,padding=1), bn(self.channel_mult*1),nn.LeakyReLU(slope, True),
nn.Conv2d(self.channel_mult*1, self.channel_mult*2, 4, 2, 1), bn(self.channel_mult*2), nn.LeakyReLU(slope, inplace=True),
nn.Conv2d(self.channel_mult*2, self.channel_mult*4, 4, 2, 1), bn(self.channel_mult*4), nn.LeakyReLU(slope, inplace=True),
nn.Conv2d(self.channel_mult*4, self.channel_mult*8, 4, 2, 1),bn(self.channel_mult*8), nn.LeakyReLU(slope, inplace=True),
nn.Conv2d(self.channel_mult*8, self.channel_mult*8, 3, 2, 1), bn(self.channel_mult*8), nn.LeakyReLU(slope, inplace=True)])
#self.flat_fts = self.get_flat_fts(self.conv)
self.nin=self.channel_mult*8
self.linear = nn.Sequential(nn.Linear(self.nin, output_size),bn1d(output_size),nn.LeakyReLU(slope),)
def forward(self, x):
for il,l in enumerate(self.conv):
x = l(x)
x = torch.flatten(x,start_dim=1)#x.view(-1, self.flat_fts)
return self.linear(x)
class CNN_Decoder(nn.Module):
def __init__(self, cfg):
super(CNN_Decoder, self).__init__()
self.input_dim = cfg["aez"] #cfg["aecfg"]["esize"]
self.channel_mult = int(64*(cfg["netSi"]+0.25*int("F" in cfg["ds"][0])))
self.fc_output_dim = self.channel_mult*16#self.input_dim#int(64*cfg["aecfg"]["netSi"]) #cfg["aecfg"]["esize"]#128#256
self.fc = nn.Sequential(nn.Linear(self.input_dim, self.fc_output_dim),nn.BatchNorm1d(self.fc_output_dim),nn.ReLU(True))
bn = lambda x: nn.BatchNorm2d(x) if cfg["aebn"] else nn.Identity()
slope=0.05
self.deconv = nn.Sequential(*[nn.ConvTranspose2d(self.fc_output_dim,self.channel_mult * 8, 4, 2,1, bias=False), bn(self.channel_mult * 8), nn.LeakyReLU(slope),
nn.ConvTranspose2d(self.channel_mult * 8,self.channel_mult * 4, 4, 2,1, bias = False), bn(self.channel_mult * 4), nn.LeakyReLU(slope),
nn.ConvTranspose2d(self.channel_mult * 4,self.channel_mult * 2, 4, 2, 1, bias = False), bn(self.channel_mult * 2), nn.LeakyReLU(slope),
nn.ConvTranspose2d(self.channel_mult * 2,self.channel_mult * 1, 4, 2, 1, bias = False), bn(self.channel_mult * 1), nn.LeakyReLU(slope),
nn.ConvTranspose2d(self.channel_mult * 1,cfg["imCh"], 4, 2, 1, bias = False)])
def forward(self, x):
x = self.fc(x)
x = x.view(-1, self.fc_output_dim, 1, 1)
for l in self.deconv: x=l(x)
return F.tanh(x)
class AENetwork(nn.Module):
def __init__(self, cfg):
super(AENetwork, self).__init__()
self.encoder = CNN_Encoder(cfg,input_size=(cfg["imCh"],32,32))
self.decoder = CNN_Decoder(cfg)
def enc(self, x): return self.encoder(x)
def dec(self, z): return self.decoder(z)
def forward(self, x):
z = self.enc(x)#.view(-1, 784))
return self.dec(z),z
def getAEDat(netAE,dataset,encoded=True):
netAE.eval()
aencx=[]
aency = []
def nor(x): return (x - np.mean(x, axis=(0, 2, 3), keepdims=True)) / (np.std(x, axis=(0, 2, 3), keepdims=True) + 1e-7)
with tca.autocast():
with torch.no_grad():
for i, data in enumerate(dataset):
x = data[0].cuda()
selfx,code = netAE(x)
tosave=code if encoded else selfx
aencx.append(tosave.detach().cpu().numpy())
aency.append(np.copy(data[1].cpu().numpy()))
return np.concatenate(aencx,axis=0),np.concatenate(aency,axis=0)
def getAEDatIter(netAE,trdata,tedata,encoded=True,cds=None):
aetrX, aetrY=getAEDat(netAE, trdata, encoded=encoded)
aetr_iter = cds(aetrX, aetrY, True, False)
aeteX, aeteY = getAEDat(netAE, tedata, encoded=encoded)
aete_iter = cds(aeteX, aeteY, False, False)
return aetr_iter,aete_iter
def runAE(cfg,dataset,tedataset,sname,cget,picname):
getM =getAEModel
netAE, acfg = getM(cfg, dataset, sname, cget,picname)
trds, teds = getAEDat(netAE, dataset), getAEDat(netAE, tedataset)
#imgutil.makeAndStore(trds[:64], trds[:64], cfg["bFolder"] + "samples/", "AE" + picname + fs(cfg) + ".png")
return trds,teds,acfg,netAE
def decay(ccf,epoch,optimizerCl):
if ccf["opt"][0] == "S" and (epoch + 1) % (ccf["opt"][1] // 3+ccf["opt"][1]//10+2 ) == 0:
for p in optimizerCl.param_groups: p['lr'] *= 0.1
print(" D", np.round(optimizerCl.param_groups[0]['lr'],5))
def getAEModel(cfg, train_dataset, sname, cget,picname=""): #Co,val_datasetMa,resFolder
ccf=cfg["aecfg"]
netAE = AENetwork(cfg).cuda()
if cfg["aeGAN"][0]:
netD=AEDisc(cfg).cuda()
optimizerD = optim.Adam(netD.parameters(), lr=cfg["aeGAN"][1]["lr"], betas=(0.5, 0.999))
optimizerG = optim.Adam(netAE.parameters(), lr=cfg["aeGAN"][1]["lr"], betas=(0.5, 0.999))
criterion = nn.BCEWithLogitsLoss()
real_label, fake_label = 1., 0.
gloss,drloss,dfloss=0,0,0
if ccf["opt"][0] == "S": optimizerCl = optim.SGD(netAE.parameters(), lr=ccf["opt"][1], momentum=0.8, weight_decay=ccf["opt"][2])
elif ccf["opt"][0] == "A": optimizerCl = optim.Adam(netAE.parameters(), ccf["opt"][1], weight_decay=ccf["opt"][2])
else: "Error opt not found"
closs, trep, loss = 0, cfg["epA"], nn.MSELoss()#nn.CrossEntropyLoss()
print("Train AE")
scaler = tca.GradScaler()
ulo = lambda x,t,e: 0.97*x+0.03*t.item() if e>1 else 0.85*x+0.15*t.item()
torch.backends.cudnn.benchmark = True
for epoch in range(trep):
netAE.train()
for i, data in enumerate(train_dataset):
with tca.autocast():
optimizerCl.zero_grad()
x=data[0].cuda()
outAE,lo=netAE(x)
errD_real = loss(torch.flatten(outAE,1),torch.flatten(x,1))
scaler.scale(errD_real).backward()
scaler.step(optimizerCl)
scaler.update()
if cfg["aeGAN"][0]:
## Train with all-real batch
netD.zero_grad()
b_size = x.size(0)
label = torch.full((b_size,), real_label, dtype=torch.float).cuda()
outreal = netD(x).view(-1)
errD_real = criterion(outreal, label)
scaler.scale(errD_real).backward()
## Train with all-fake batch
label.fill_(fake_label)
outfake = netD(outAE.detach()).view(-1)
errD_fake = criterion(outfake, label)
scaler.scale(errD_fake).backward()
scaler.step(optimizerD)
# (2) Update G network: maximize log(D(G(z)))
optimizerG.zero_grad()
label.fill_(real_label) # fake labels are real for generator cost
outAE, _ = netAE(x)
outfake = netD(outAE).view(-1)
errG = criterion(outfake, label)
scaler.scale(errG).backward()
scaler.step(optimizerG)
scaler.update()
gloss = ulo(gloss, errG, epoch)
drloss = ulo(drloss, errD_real, epoch)
dfloss = ulo(dfloss, errD_fake, epoch)
closs = ulo(closs,errD_real,epoch)
decay(ccf,epoch,optimizerCl)
netAE.eval()
if (epoch % 2 == 0 and epoch<=10) or (epoch % 10==0 and epoch>10):
print(epoch, "AE", np.round(np.array([closs]+([gloss,drloss,dfloss] if cfg["aeGAN"][0] else [])), 5), cfg["pr"])
if np.isnan(closs):
print("Failed!!!")
return None,None
lcfg = {"AELo": closs}
if cfg["aeGAN"][0]: lcfg={**lcfg,**{"glo":gloss,"drlo":drloss,"dflo":dfloss}}
netAE.eval()
return netAE, lcfg