-
Notifications
You must be signed in to change notification settings - Fork 0
/
learner_v0.py
227 lines (196 loc) · 8.07 KB
/
learner_v0.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
221
222
223
224
225
226
227
import torch
from torch import nn
from torch.nn import functional as F
import numpy as np
class Learner(nn.Module):
"""
Base learner?
"""
def __init__(self, config):
"""
:param config: network config file, type:list of (string, list)
:param imgc: 1 or 3
:param imgsz: 28 or 84
"""
super(Learner, self).__init__()
self.config = config
# this dict contains all tensors needed to be optimized
self.vars = nn.ParameterList()
# running_mean and running_var
self.vars_bn = nn.ParameterList() # batch norm
for i, (name, param) in enumerate(self.config):
if name is 'conv2d':
# [ch_out, ch_in, kernelsz, kernelsz]
w = nn.Parameter(torch.ones(*param[:4]))
# gain=1 according to cbfin's implementation
torch.nn.init.kaiming_normal_(w)
self.vars.append(w) #
# [ch_out] # bias
self.vars.append(nn.Parameter(torch.zeros(param[0])))
elif name is 'convt2d':
# [ch_in, ch_out, kernelsz, kernelsz, stride, padding]
w = nn.Parameter(torch.ones(*param[:4]))
# gain=1 according to cbfin's implementation
torch.nn.init.kaiming_normal_(w)
self.vars.append(w)
# [ch_in, ch_out]# bias
self.vars.append(nn.Parameter(torch.zeros(param[1])))
elif name is 'linear':
# [ch_out, ch_in]
w = nn.Parameter(torch.ones(*param))
# gain=1 according to cbfinn's implementation
torch.nn.init.kaiming_normal_(w)
self.vars.append(w)
# [ch_out] # bias
self.vars.append(nn.Parameter(torch.zeros(param[0])))
elif name is 'bn':
'''
w and b are learnable parameters each with size c of the input
'''
# [ch_out]
w = nn.Parameter(torch.ones(param[0]))
self.vars.append(w)
# [ch_out]
self.vars.append(nn.Parameter(torch.zeros(param[0])))
# must set requires_grad=False
running_mean = nn.Parameter(torch.zeros(param[0]),
requires_grad=False)
running_var = nn.Parameter(torch.ones(param[0]),
requires_grad=False)
self.vars_bn.extend([running_mean, running_var])
elif name in ['tanh', 'relu', 'upsample', 'avg_pool2d',
'max_pool2d', 'flatten', 'reshape', 'leakyrelu',
'sigmoid']:
continue
else:
raise NotImplementedError
def extra_repr(self):
info = ''
for name, param in self.config:
if name is 'conv2d':
tmp = (r'conv2d:(ch_in:%d, ch_out:%d, k:%dx%d, stride:%d,'
r'padding:%d)'%(param[1], param[0], param[2],
param[3], param[4], param[5],)
info += tmp + '\n'
elif name is 'convt2d':
tmp = (r'convTranspose2d:(ch_in:%d, ch_out:%d, k:%dx%d,'
r' stride:%d, padding:%d)')%(param[0], param[1],
param[2], param[3], param[4], param[5],)
info += tmp + '\n'
elif name is 'linear':
tmp = 'linear:(in:%d, out:%d)'%(param[1], param[0])
info += tmp + '\n'
elif name is 'leakyrelu':
tmp = 'leakyrelu:(slope:%f)'%(param[0])
info += tmp + '\n'
elif name is 'avg_pool2d':
tmp = 'avg_pool2d:(k:%d, stride:%d, padding:%d)'%(param[0],
param[1], param[2])
info += tmp + '\n'
elif name is 'max_pool2d':
tmp = 'max_pool2d:(k:%d, stride:%d, padding:%d)'%(param[0],
param[1], param[2])
info += tmp + '\n'
# below are implemented in forward
elif name in ['flatten', 'tanh', 'relu', 'upsample', 'reshape',
'sigmoid', 'use_logits', 'bn']:
tmp = name + ':' + str(tuple(param))
info += tmp + '\n'
else:
raise NotImplementedError
return info
def forward(self, x, vars=None, bn_training=True):
"""
This function can be called by finetunning, however, in finetunning,
we dont wish to update running_mean/running_var.
Thought weights/bias of bn is updated, it has been separated
by fast_weights.
Indeed, to not update running_mean/running_var, we need set
update_bn_statistics=False
but weight/bias will be updated and not dirty initial theta parameters
via fast_weigths.
:param x: [b, 1, 28, 28]
:param vars:
:param bn_training: set False to not update
:return: x, loss, likelihood, kld
"""
if vars is None:
vars = self.vars
idx = 0
bn_idx = 0
for name, param in self.config:
if name is 'conv2d':
w, b = vars[idx], vars[idx + 1]
# remember to keep synchrozied of forward_encoder and
# forward_decoder!
x = F.conv2d(x, w, b, stride=param[4], padding=param[5])
idx += 2
# print(name, param, '\tout:', x.shape)
elif name is 'convt2d':
w, b = vars[idx], vars[idx + 1]
# remember to keep synchrozied of forward_encoder and
# forward_decoder!
x = F.conv_transpose2d(x, w, b, stride=param[4],
padding=param[5])
idx += 2
# print(name, param, '\tout:', x.shape)
elif name is 'linear':
w, b = vars[idx], vars[idx + 1]
x = F.linear(x, w, b)
idx += 2
# print('forward:', idx, x.norm().item())
elif name is 'bn':
w, b = vars[idx], vars[idx + 1]
running_mean, running_var = \
self.vars_bn[bn_idx], self.vars_bn[bn_idx+1]
x = F.batch_norm(x, running_mean, running_var, weight=w,
bias=b, training=bn_training)
idx += 2
bn_idx += 2
elif name is 'flatten':
# print(x.shape)
x = x.view(x.size(0), -1)
elif name is 'reshape':
# [b, 8] => [b, 2, 2, 2]
x = x.view(x.size(0), *param)
elif name is 'relu':
x = F.relu(x, inplace=param[0])
elif name is 'leakyrelu':
x = F.leaky_relu(x, negative_slope=param[0], inplace=param[1])
elif name is 'tanh':
x = F.tanh(x)
elif name is 'sigmoid':
x = torch.sigmoid(x)
elif name is 'upsample':
x = F.upsample_nearest(x, scale_factor=param[0])
elif name is 'max_pool2d':
x = F.max_pool2d(x, param[0], param[1], param[2])
elif name is 'avg_pool2d':
x = F.avg_pool2d(x, param[0], param[1], param[2])
else:
raise NotImplementedError
# make sure variable is used properly
assert idx == len(vars)
assert bn_idx == len(self.vars_bn)
return x
def zero_grad(self, vars=None):
"""
:param vars:
:return:
"""
with torch.no_grad():
if vars is None:
for p in self.vars:
if p.grad is not None:
p.grad.zero_()
else:
for p in vars:
if p.grad is not None:
p.grad.zero_()
def parameters(self):
"""
override this function since initial parameters will return with a
generator.
:return:
"""
return self.vars