-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrainerClasses.py
205 lines (165 loc) · 9.28 KB
/
trainerClasses.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
import numpy as np
from torch import nn
import torch.nn.functional as F
import time
from torch.utils.data import DataLoader
import wandb
import torch
def admmInputGenerator(genData, U_, S_, V_, imgSize):
compData = F.linear(genData, U_.T) # compressed data
lsqrInp = F.linear(compData / (S_ + 1e-4), V_).reshape(imgSize)
return compData, lsqrInp
def transformDataset(data, imgSizes, rescaleVals, randVals = None):
dims = len(imgSizes)
reScaleMin, reScaleMax = rescaleVals
imgSize = data.shape
if dims == 2:
n1, n2 = imgSizes
data -= data.reshape(imgSize[0], imgSize[1], -1).min(dim = 2).values[:,:,None,None]
data /= data.reshape(imgSize[0], imgSize[1], -1).max(dim = 2).values[:,:,None,None]
if (n1 < data.shape[2]) or (n2 < data.shape[3]):
if randVals is None:
rand1 = torch.randint(low = 0, high = data.shape[2] - n1, size = (1,))
rand2 = torch.randint(low = 0, high = data.shape[3] - n2, size = (1,))
else:
rand1 = randVals[0]
rand2 = randVals[1]
else:
rand1 = 0
rand2 = 0
data = data[:, :, rand1:rand1 + n1, rand2:rand2 + n2]
else:
n1, n2, n3 = imgSizes # n1 is 16
data -= data.reshape(imgSize[0], imgSize[1], -1).min(dim = 2).values[:,:,None,None,None]
data /= data.reshape(imgSize[0], imgSize[1], -1).max(dim = 2).values[:,:,None,None,None]
diffS = np.array(imgSizes) - np.array(data.shape[2:])
diffS *= (diffS > 0)
data = F.pad(data, (diffS[2], diffS[2], diffS[1], diffS[1], diffS[0], diffS[0])) # new data size is
if (n1 < data.shape[2]) or (n2 < data.shape[3]) or (n3 < data.shape[3]):
if randVals is None:
rand1 = torch.randint(low = 0, high = data.shape[2] - n1, size = (1,))
rand2 = torch.randint(low = 0, high = data.shape[3] - n2, size = (1,))
rand3 = torch.randint(low = 0, high = data.shape[4] - n3, size = (1,))
else:
rand1 = randVals[0]
rand2 = randVals[1]
rand3 = randVals[2]
else:
rand1 = 0
rand2 = 0
rand3 = 0
data = data[:, :, rand1:rand1 + n1, rand2:rand2 + n2, rand3:rand3 + n3]
if reScaleMax > 0:
randScale = torch.rand((imgSize[0], 1, 1, 1), device = data.device) * reScaleMax + reScaleMin
if dims == 3:
randScale = randScale.reshape(imgSize[0], 1, 1, 1, 1)
else:
randScale = 1
data *= randScale
return data
def returnNsTerm(fixedNoiseStdFlag = False,shape=0, maxNoiseStd=0.05, minNoiseStd=0, useDev = torch.device("cpu")):
if fixedNoiseStdFlag:
noiseStd = maxNoiseStd
return torch.randn(shape, device = useDev) * noiseStd
else:
noiseStd = (torch.rand(shape, device = useDev)*(maxNoiseStd-minNoiseStd)+minNoiseStd)
return torch.randn(shape, device = useDev) * noiseStd
def trainDenoiser(model, epoch_nb, loss, optimizer, scheduler, trainDataset, valDataset, batch_size_train, batch_size_val, rescaleVals = [1, 1], saveModelEpoch=0, valEpoch=0, saveDirectory='', maxNoiseStd = 0.1, optionalMessage="",
wandbFlag=False, fixedNoiseStdFlag = False, minNoiseStd =0, dims = 2):
trainLosses = torch.zeros(epoch_nb)
trainNrmses = torch.zeros(epoch_nb)
trainPsnrs = torch.zeros(epoch_nb)
valLosses = list()
valNrmses = list()
valPsnrs = list()
trainLoader = DataLoader(trainDataset, batch_size_train, shuffle=True)
valLoader = DataLoader(valDataset, valDataset.__len__(), shuffle=False)
for epoch in range(1,1+int(epoch_nb)):
tempLosses = list()
model.train()
tempNrmseNumeratorSquare = 0
tempNrmseDenumeratorSquare = 0
tempNumel = 0
tempTime = time.time()
if saveModelEpoch > 0:
if (epoch % saveModelEpoch == 0):
torch.save(model.state_dict(), saveDirectory+r"/"+ optionalMessage +"epoch"+ str(epoch)+ ".pth")
for idx, data in enumerate(trainLoader, 0):
data = data.float().cuda()
data = transformDataset(data, [*data.shape[2:]], rescaleVals)
noiseTerm = returnNsTerm(fixedNoiseStdFlag = fixedNoiseStdFlag, shape=data.shape, \
maxNoiseStd=maxNoiseStd, minNoiseStd=minNoiseStd, useDev = data.device)
noisyInp = data + noiseTerm
modelOut = model(noisyInp)
model.zero_grad()
model_loss = loss(modelOut, data)
model_loss.backward()
optimizer.step()
if dims == 3:
percent = 10
if idx%int(trainDataset.__len__()/batch_size_train/percent)==0:
print('Epoch {0:d} | {1:d}% | batch nrmse: {2:.5f}'.format(epoch,percent*idx//int(trainDataset.__len__()/batch_size_train/percent),(float(torch.norm(modelOut-data)))/(float(torch.norm(data))))) # myflag
with torch.no_grad():
tempLosses.append(float(model_loss))
tempNrmseNumeratorSquare += (float(torch.norm(modelOut-data)))**2
tempNrmseDenumeratorSquare += (float(torch.norm(data)))**2
tempNumel += modelOut.numel()
# back to epoch
model.eval()
scheduler.step()
trainLosses[epoch-1] = sum(tempLosses)/len(tempLosses)
trainNrmses[epoch-1] = (tempNrmseNumeratorSquare/tempNrmseDenumeratorSquare)**(1/2)
trainPsnrs[epoch-1] = 20 * \
torch.log10(1 / (tempNrmseDenumeratorSquare**(1/2) * #Should we correct 1 -> valGround.max()
trainNrmses[epoch-1] / (tempNumel) ** (1/2)))
epochTime = time.time() - tempTime
if wandbFlag:
wandb.log({"train_loss": trainLosses[epoch-1], "train_nrmse": trainNrmses[epoch-1], "train_psnr": trainPsnrs[epoch-1]})
print("Epoch: {0}, Train Loss = {1:.6f}, Train nRMSE = {2:.6f}, Train pSNR = {3:.6f}, time elapsed = {4:.6f}".format(epoch,
trainLosses[epoch-1], trainNrmses[epoch-1], trainPsnrs[epoch-1], epochTime))
if valEpoch>0:
if epoch % valEpoch == 0:
with torch.no_grad():
model.eval()
valInp = next(iter(valLoader))
valGround = valInp.clone()
valGround = transformDataset(valGround, [*valInp.shape[2:]], rescaleVals)
noiseTerm = returnNsTerm(fixedNoiseStdFlag = fixedNoiseStdFlag,shape=valInp.shape,\
maxNoiseStd=maxNoiseStd, minNoiseStd=minNoiseStd, useDev = valGround.device)
valInpVal = valGround + noiseTerm
valOut = torch.zeros_like(valGround)
deviceVal = valGround.device #'cuda' if valOut.is_cuda else 'cpu'
iii = 0
while(iii < valInpVal.shape[0]-(valInpVal.shape[0] % batch_size_val)):
valInpC = valInpVal[iii:iii+batch_size_val].float().cuda()
valOut[iii:iii+batch_size_val] = model(valInpC).to(deviceVal)
iii += batch_size_val
valInpC = valInpVal[iii:].float().cuda()
valOut[iii:] = model(valInpC).to(deviceVal)
valLoss = float(nn.L1Loss()(valGround, valOut))
valNrmse = float(torch.norm(valGround-valOut)/torch.norm(valGround))
valPSNR= float(20 * \
torch.log10(1 / (torch.norm(valGround) * #Should we correct 1 -> valGround.max()
valNrmse / (valOut.numel()) ** (1/2))))
valPSNR_avg = (20 *
torch.log10(1 / (torch.norm(valGround.reshape(valGround.shape[0],-1)-valOut.reshape(valGround.shape[0],-1), dim = (1)).squeeze() / (valOut[0,0].numel()) ** (1/2))))
valLosses.append(valLoss)
valNrmses.append(valNrmse)
valPsnrs.append(valPSNR)
if wandbFlag:
wandb.log({"valid_nRMSE": valNrmse,
"ref_nRMSE": torch.norm(valInp-valGround)/torch.norm(valGround),
'valid_pSNR': valPSNR,
'valid_pSNRavg': valPSNR_avg.mean(0),
'valid_pSNRstd': valPSNR_avg.std(0),
'valid_loss': valLoss})
print("---Epoch: {0}, Val Loss = {1:.6f}, Val nRMSE = {2:.6f}, Val pSNR = {3:.6f}".format(epoch, valLoss, valNrmse, valPSNR))
if wandbFlag:
wandb.log({"valid_nRMSE": valNrmse,
'valid_pSNR': valPSNR,
'valid_pSNRavg': valPSNR_avg.mean(0),
'valid_pSNRstd': valPSNR_avg.std(0),
'valid_loss': valLoss})
print("---Epoch: {0}, Val Loss = {1:.6f}, Val nRMSE = {2:.6f}, Val pSNR = {3:.6f}".format(epoch, valLoss, valNrmse, valPSNR))
torch.save(model.state_dict(), saveDirectory+r"/"+ optionalMessage +"epoch"+ str(epoch)+ "END.pth")
return model, [trainLosses.numpy(), trainNrmses.numpy(), trainPsnrs.numpy()], [np.array(valLosses), np.array(valNrmses), np.array(valPsnrs)]