-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtestit.py
220 lines (180 loc) · 7.29 KB
/
testit.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
209
210
211
212
213
214
215
216
217
218
219
220
import torch
import proximal_gradient.proximalGradient as pg
class OneLayerNet(torch.nn.Module):
def __init__(self, D_in, H, D_out):
"""
The constructor creates one linear layer and assigns it a name.
"""
super(OneLayerNet, self).__init__()
self.linear1 = torch.nn.Linear(D_in, D_out)
self.linear1.name = "linear1"
def forward(self, x):
"""
Simple forward step
"""
y_pred = self.linear1(x)
# Uncomment for verbose debugging
#print("linear1:", self.linear1)
#for param in self.linear1.parameters():
# print("param:", param)
# print("param.grad:", param.grad)
##print("linear1.grad:", self.linear1.grad)
##print("linear1.grad:", self.linear1.data)
return y_pred
def build_model():
# Values for the network size
N, D_in, H, D_out = 4, 3, 4, 2
#N, D_in, H, D_out = 4, 3, 10, 5
# Create random Tensors to hold inputs and outputs
x = torch.zeros(N, D_in)
y = torch.ones(N, D_out)
print("x.requires_grad")
print(x.requires_grad)
# Construct our model by instantiating the class defined above
model = OneLayerNet(D_in, H, D_out)
print("model:", model)
criterion = torch.nn.MSELoss(size_average=False)
#optimizer = torch.optim.SGD(model.parameters(), lr=1e-4)
optimizer = torch.optim.SGD(model.parameters(), lr=1e-1)
return (x,y,model,criterion,optimizer)
def test_l1(network):
x, y, model, criterion, optimizer = network
for t in range(10):
# Forward pass: Compute predicted y by passing x to the model
y_pred = model(x)
# Compute and print loss
cross_entropy_loss = criterion(y_pred, y)
loss = cross_entropy_loss
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
# print("model.linear1.weight.grad:", model.linear1.weight.grad)
# print("model.linear1.bias.grad:", model.linear1.bias.grad)
# print("model.linear1.weight before:", model.linear1.weight)
# print("model.linear1.bias before:", model.linear1.bias)
optimizer.step()
# print("model.linear1.weight after:", model.linear1.weight)
# print("model.linear1.bias after:", model.linear1.bias)
# print("model.linear1.weight.norm():", model.linear1.weight.norm())
#L1...
print("weight before:", model.linear1.weight)
pg.l1(model.linear1.weight, model.linear1.bias, reg=0.1)
print("weight after:", model.linear1.weight)
def test_l21(network):
print("Testing l21")
x, y, model, criterion, optimizer = network
for t in range(10):
# Forward pass: Compute predicted y by passing x to the model
y_pred = model(x)
# Compute and print loss
cross_entropy_loss = criterion(y_pred, y)
loss = cross_entropy_loss
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
#print("weight before:", model.linear1.weight)
#print("bias before:", model.linear1.bias)
#pg.l21(model.linear1.weight, reg=0.1) # Use defaults
pg.l21(model.linear1.weight, model.linear1.bias, reg=0.1)
#pg.l21(model.linear1, reg=0.01) # Test different learning rates
#pg.l21(model.linear1, reg=0.1)
#pg.l21_slow(model.linear1.weight, reg=0.1) # Slow version to double check accuracy
#print("weight after:", model.linear1.weight)
#print("bias after:", model.linear1.bias)
def test_l2(network):
print("Testing l2")
x, y, model, criterion, optimizer = network
for t in range(10):
# Forward pass: Compute predicted y by passing x to the model
y_pred = model(x)
# Compute and print loss
cross_entropy_loss = criterion(y_pred, y)
loss = cross_entropy_loss
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Seeing about l2...
#print("L2:", model.linear1.weight.norm(2))
#print("weight before:", model.linear1.weight)
pg.l2(model.linear1.weight, model.linear1.bias, reg=0.1)
#print("weight after:", model.linear1.weight)
def test_linf1(network):
x, y, model, criterion, optimizer = network
for t in range(500):
# Forward pass: Compute predicted y by passing x to the model
y_pred = model(x)
# Compute and print loss
cross_entropy_loss = criterion(y_pred, y)
loss = cross_entropy_loss
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
#Linf1...
print("weight before:", model.linear1.weight)
print("bias before:", model.linear1.bias)
pg.linf1(model.linear1.weight, model.linear1.bias, reg=0.1)
print("weight after:", model.linear1.weight)
print("bias after:", model.linear1.bias)
def test_linf(network):
x, y, model, criterion, optimizer = network
for t in range(10):
# Forward pass: Compute predicted y by passing x to the model
y_pred = model(x)
# Compute and print loss
cross_entropy_loss = criterion(y_pred, y)
loss = cross_entropy_loss
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
#Linf
print("weight before:", model.linear1.weight)
pg.linf(model.linear1.weight, model.linear1.bias, reg=0.1)
print("weight after:", model.linear1.weight)
def test_elasticnet(network):
print("Testing elasticnet")
x, y, model, criterion, optimizer = network
for t in range(10):
# Forward pass: Compute predicted y by passing x to the model
y_pred = model(x)
# Compute and print loss
cross_entropy_loss = criterion(y_pred, y)
loss = cross_entropy_loss
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
#Elastic Net
#print("weight before:", model.linear1.weight)
pg.elasticnet(model.linear1.weight, model.linear1.bias, reg=0.1)
#print("weight after:", model.linear1.weight)
def test_logbarrier(network):
print("Testing logbarrier")
x, y, model, criterion, optimizer = network
for t in range(10):
# Forward pass: Compute predicted y by passing x to the model
y_pred = model(x)
# Compute and print loss
cross_entropy_loss = criterion(y_pred, y)
loss = cross_entropy_loss
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
#Log Barrier...
#print("weight before:", model.linear1.weight)
pg.logbarrier(model.linear1.weight, model.linear1.bias, reg=0.1)
#print("weight after:", model.linear1.weight)
def main():
network = build_model()
test_l1(network)
test_linf1(network)
test_elasticnet(network)
test_logbarrier(network)
test_l2(network)
test_l21(network)
if __name__ == "__main__":
main()