forked from mfahes/LPALM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
164 lines (132 loc) · 6.77 KB
/
train.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
import torch
import torch.nn as nn
import numpy as np
from ISTA_LLT import ISTA_LLT
from LISTA_LeCun import LISTA_LeCun
from LISTA_CP import LISTA_CP
from LPALM import LPALM
from param_PALM import param_PALM
import matplotlib.pyplot as plt
from scipy.io import savemat
import os
def train_non_blind(train_loader, val_loader, num_epochs=10, T=10, alpha = 10, mode = None,L_shared = True, theta_shared=True,We_shared=True,G_shared=True,W_shared=True,noise_A=None,realistic_train=False):
criterion = nn.MSELoss()
layers = np.arange(T)
A = next(iter(train_loader))[1]
if mode == 'LC':
model = LISTA_LeCun(T, alpha= alpha, A=A,theta_shared=theta_shared,We_shared=We_shared,G_shared= G_shared)
elif mode == 'CP':
model = LISTA_CP(T,alpha=alpha,A=A,theta_shared=theta_shared, W_shared= W_shared)
elif mode == 'LLT':
model = ISTA_LLT(T,alpha=alpha,A=A,L_shared=L_shared, theta_shared= theta_shared)
else:
print("The entered mode doesn't correspond to a model")
total_params = [p.numel() for p in model.parameters()]
print(total_params)
optimizer = torch.optim.Adam(list(model.parameters()), lr=0.0001, betas=(0.9, 0.999))
train_total_loss = []
val_total_loss = []
list_err_val_layer = [[0 for i in range(T)] for j in range(num_epochs)]
for epoch in range(num_epochs):
train_total = 0
model.train()
for i, (X,A,S) in enumerate(train_loader):
optimizer.zero_grad()
if mode == 'LC':
S_pred, S_s = model(X)
else:
S_pred, S_s = model(X,A,noise_A=noise_A)
train_loss = torch.numel(S) * criterion(S.float(), S_pred.float()) / torch.norm(S.float())**2
train_loss.backward()
optimizer.step()
train_total += train_loss.item()
train_total /= i+1
train_total_loss.append(train_total)
model.eval()
with torch.no_grad():
val_total = 0
for i, (X,A,S) in enumerate(val_loader):
if mode == 'LC':
S_pred, S_s_val = model(X)
else:
S_pred,S_s_val = model(X,A,noise_A=noise_A)
val_loss = torch.numel(S) * criterion(S.float(), S_pred.float()) / torch.norm(S.float())**2
val_total += val_loss.item()
for j,s in enumerate(S_s_val):
list_err_val_layer[epoch][j] += torch.numel(S) *criterion(S.float(), s.float()).item() / torch.norm(S.float())**2
list_err_val_layer[epoch] = [x/(i+1) for x in list_err_val_layer[epoch]]
val_total /= i+1
val_total_loss.append(val_total)
print('Epoch-{0} lr: {1}'.format(epoch, optimizer.param_groups[0]['lr']))
if epoch % 5 == 0:
print()
print("epoch:{} | training loss:{:.5f} | validation loss:{:.5f} ".format(epoch, train_total,val_total))
dirName = 'models_'+ mode + '_realistic_t' if realistic_train else 'models_'+ mode +'_synthetic'
if noise_A: dirName = dirName+ str(noise_A)
try:
os.mkdir(dirName)
print("Directory " , dirName , " Created ")
except FileExistsError:
print("Directory " , dirName , " already exists")
torch.save(model, dirName+'/'+ mode +str(epoch)+'.pth')
return train_total_loss, val_total_loss,model
def train_blind(train_loader, val_loader, num_epochs=10, T=10,learn_L_S = False ,L_S_shared=False, LISTA_S = False, W_X_S_shared= False, W_S_S_shared= False,theta_shared= False, LISTA_CP_S = True, W_CP_S_shared= False, ISTA_LLT_S = False, learn_L_A = True , L_A_shared= False, LISTA_A = False , W_A_A_shared = False, W_X_A_shared = False, LISTA_CP_A = False, W_CP_A_shared = False,non_update_A = False,loss_funct='supervised_A_S',realistic_train=False):
torch.autograd.set_detect_anomaly(True)
criterion = nn.MSELoss()
A = next(iter(train_loader))[1]
S = next(iter(train_loader))[2]
model = param_PALM(T, S=S , A=A,learn_L_S = learn_L_S ,L_S_shared=L_S_shared, LISTA_S = LISTA_S, W_X_S_shared= W_X_S_shared , W_S_S_shared= W_S_S_shared ,theta_shared= theta_shared, LISTA_CP_S = LISTA_CP_S, W_CP_S_shared= W_CP_S_shared , ISTA_LLT_S = ISTA_LLT_S, learn_L_A = learn_L_A , L_A_shared= L_A_shared , LISTA_A = LISTA_A , W_A_A_shared = W_A_A_shared, W_X_A_shared = W_X_A_shared , LISTA_CP_A = LISTA_CP_A, W_CP_A_shared = W_CP_A_shared, non_update_A = non_update_A)
total_params = [p.numel() for p in model.parameters()]
print(total_params)
optimizer = torch.optim.Adam(list(model.parameters()), lr=0.0001, betas=(0.9, 0.999))
train_total_loss = []
val_total_loss = []
list_err_val_S_layer = [[0 for i in range(T)] for j in range(num_epochs)]
list_err_val_A_layer = [[0 for i in range(T)] for j in range(num_epochs)]
for epoch in range(num_epochs):
train_total = 0
model.train()
for i, (X,A,S) in enumerate(train_loader):
optimizer.zero_grad()
S_pred, A_pred, S_s_train, A_s_train = model(X)
train_loss=0
if loss_funct == 'unsupervised':
train_loss = torch.numel(X) * criterion(X.float(), torch.bmm(A_pred.float(),S_pred.float())) / torch.norm(X.float())**2
elif loss_funct == 'supervised_A':
train_loss = torch.numel(A) * criterion(A.float(), A_pred.float()) / torch.norm(A.float())**2
elif loss_funct == 'supervised_S':
train_loss = torch.numel(S) * criterion(S.float(), S_pred.float()) / torch.norm(S.float())**2
elif loss_funct == 'supervised_A_S':
train_loss = torch.numel(S) * criterion(S.float(), S_pred.float()) / torch.norm(S.float())**2 + torch.numel(A) * criterion(A.float(),A_pred.float()) / torch.norm(A.float())**2
train_loss.backward()
optimizer.step()
train_total += train_loss.item()
train_total /= i+1
train_total_loss.append(train_total)
model.eval()
with torch.no_grad():
val_total = 0
for i, (X,A,S) in enumerate(val_loader):
S_pred, A_pred, S_s_val, A_s_val = model(X)
val_loss = torch.numel(S) * criterion(S.float(), S_pred.float()) / torch.norm(S.float())**2 + torch.numel(A) * criterion(A.float(),A_pred.float()) / torch.norm(A.float())**2
val_total += val_loss.item()
for j,s in enumerate(S_s_val):
list_err_val_S_layer[epoch][j] += torch.numel(S) * criterion(S.float(), s.float()).item() / torch.norm(S.float())**2
for j,a in enumerate(A_s_val):
list_err_val_A_layer[epoch][j] += torch.numel(A) * criterion(A.float(), a.float()).item() / torch.norm(A.float())**2
list_err_val_S_layer[epoch] = [x/(i+1) for x in list_err_val_S_layer[epoch]]
list_err_val_A_layer[epoch] = [x/(i+1) for x in list_err_val_A_layer[epoch]]
val_total /= i+1
val_total_loss.append(val_total)
print('Epoch-{0} lr: {1}'.format(epoch, optimizer.param_groups[0]['lr']))
if epoch % 5 == 0:
print()
print("epoch:{} | training loss:{:.5f} | validation loss:{:.5f} ".format(epoch, train_total,val_total))
dirName = 'models_LPALM_realistic_t' if realistic_train else 'models_LPALM_synthetic'
try:
os.mkdir(dirName)
print("Directory " , dirName , " Created ")
except FileExistsError:
print("Directory " , dirName , " already exists")
torch.save(model, dirName+'/LPALM'+str(epoch)+'.pth')
return train_total_loss, val_total_loss,model