-
Notifications
You must be signed in to change notification settings - Fork 3
Expand file tree
/
Copy pathtrain.py
More file actions
259 lines (203 loc) · 8.07 KB
/
train.py
File metadata and controls
259 lines (203 loc) · 8.07 KB
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
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
'''
Helper files for training.
'''
from torch.autograd import Function, gradcheck
from torch.utils.data import DataLoader, Dataset
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
import pickle
import os
from main import sample
def load_data(path, num_train, num_test):
'''
Loads dataset from `path` split into Pytorch train and test of
given sizes. Train set is taken from the front while
test set is taken from behind.
:param path: path to .p file containing data.
'''
f = open(path, 'rb')
all_data = pickle.load(f)['samples']
ndata_all = all_data.size()[0]
assert num_train+num_test <= ndata_all
train_data = all_data[:num_train]
test_data = all_data[(ndata_all-num_test):]
return train_data, test_data
def load_log_ll(path, num_train, num_test):
f = open(path, 'rb')
all_log_ll = pickle.load(f)['log_ll']
ndata_all = all_log_ll.numel()
assert num_train+num_test <= ndata_all
train_log_ll = all_log_ll[:num_train]
test_log_ll = all_log_ll[(ndata_all-num_test):]
return train_log_ll, test_log_ll
def get_optim(name, net, args):
if name == 'SGD':
optimizer = optim.SGD(net.parameters(), args['lr'], args['momentum'])
elif name == 'Adam':
# TODO: add in more. Note: we do not use this in the paper.
optimizer = optim.Adam(net.parameters(), args['lr'])
elif name == 'RMSprop':
# TODO: add in more. Note: we do not use this in the paper.
optimizer = optim.RMSprop(net.parameters(), args['lr'])
return optimizer
def expt(train_data, val_data,
net,
optim_name,
optim_args,
identifier,
num_epochs=1000,
batch_size=100,
chkpt_freq=50,
):
os.mkdir('./checkpoints/%s' % identifier)
os.mkdir('./sample_figs/%s' % identifier)
train_loader = DataLoader(
train_data, batch_size=batch_size, shuffle=True)
# IMPORTANT: for this experiment, we did *not* perform hyperparameter tuning.
# Hence, the `validation loss' here is essentially `test` loss.
val_loader = DataLoader(
val_data, batch_size=1000000, shuffle=True)
optimizer = get_optim(optim_name, net, optim_args)
train_loss_per_epoch = []
for epoch in range(num_epochs):
loss_per_minibatch = []
for i, data in enumerate(train_loader, 0):
optimizer.zero_grad()
d = torch.tensor(data, requires_grad=True)
p = net(d, mode='pdf')
logloss = -torch.sum(torch.log(p))
reg_loss = logloss
reg_loss.backward()
scalar_loss = (reg_loss/p.numel()).detach().numpy().item()
loss_per_minibatch.append(scalar_loss)
optimizer.step()
train_loss_per_epoch.append(np.mean(loss_per_minibatch))
print('Training loss at epoch %s: %s' %
(epoch, train_loss_per_epoch[-1]))
if epoch % chkpt_freq == 0:
print('Checkpointing')
torch.save({
'epoch': epoch,
'model_state_dict': net.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': logloss,
}, './checkpoints/%s/epoch%s' % (identifier, epoch))
"""
if args.dims == 2:
print('Scatter sampling')
samples = sample(net, 2, 1000)
plt.scatter(samples[:, 0], samples[:, 1])
plt.savefig('./sample_figs/%s/epoch%s.png' %
(identifier, epoch))
plt.clf()
else:
print('Not doign scatter plot, dims > 2')
"""
print('Evaluating validation/test loss.')
for j, val_data in enumerate(val_loader, 0):
net.zero_grad()
val_p = net(val_data, mode='pdf')
val_loss = -torch.mean(torch.log(val_p))
print('Average validation/test loss %s' % val_loss)
def make_ranged_data(data, width, seed):
'''
Transforms each coordinate (x, y) to
the range ([x-e1, x+e2], [y-e3, y+e4])
where e1, e2, e3, e4 are drawn uniformly from [0, width].
The final ranges are snapped to [0, 1].
'''
assert data.shape[1] == 2
np.random.seed(seed)
epsilon_lower = np.random.random_sample((data.shape)) * width
epsilon_upper = np.random.random_sample((data.shape)) * width
epsilon_lower = torch.from_numpy(epsilon_lower)
epsilon_upper = torch.from_numpy(epsilon_upper)
bounds_lower = torch.max(torch.zeros_like(data), data - epsilon_lower)
bounds_upper = torch.min(torch.ones_like(data), data + epsilon_upper)
return bounds_lower, bounds_upper
def expt_cdf_noisy(train_data, val_data,
net,
optim_name,
optim_args,
identifier,
width,
seed,
num_epochs=1000,
batch_size=100,
chkpt_freq=50,
):
'''
Add in uncertainty in all points
'''
os.mkdir('./checkpoints/%s' % identifier)
os.mkdir('./sample_figs/%s' % identifier)
train_bounds_lower, train_bounds_upper = make_ranged_data(
train_data, width, seed)
val_bounds_lower, val_bounds_upper = make_ranged_data(
val_data, width, seed)
train_bounds = torch.cat(
[train_bounds_lower, train_bounds_upper], dim=1)
val_bounds = torch.cat(
[val_bounds_lower, val_bounds_upper], dim=1)
train_loader = DataLoader(
train_bounds, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(
val_data, batch_size=1000000, shuffle=True)
optimizer = get_optim(optim_name, net, optim_args)
train_loss_per_epoch = []
for epoch in range(num_epochs):
loss_per_minibatch = []
for i, data in enumerate(train_loader, 0):
optimizer.zero_grad()
d = torch.tensor(data, requires_grad=True)
dsize = d.shape[0]
big = data[:, 2:]
small = data[:, 0:2]
cross1 = torch.cat(
[data[:, 0:1], data[:, 3:4]], dim=1)
cross2 = torch.cat(
[data[:, 2:3], data[:, 1:2]], dim=1)
joint = torch.cat([big, small, cross1, cross2], dim=0)
P_raw = net(torch.tensor(joint, requires_grad=True), mode='cdf')
P_big = P_raw[:dsize]
P_small = P_raw[dsize:(2*dsize)]
P_cross1 = P_raw[(2*dsize):(3*dsize)]
P_cross2 = P_raw[(3*dsize):(4*dsize)]
P = P_big + P_small - P_cross1 - P_cross2
logloss = -torch.sum(torch.log(P))
reg_loss = logloss
reg_loss.backward()
scalar_loss = (reg_loss/P.numel()).detach().numpy().item()
loss_per_minibatch.append(scalar_loss)
optimizer.step()
train_loss_per_epoch.append(np.mean(loss_per_minibatch))
print('Training loss at epoch %s: %s' %
(epoch, train_loss_per_epoch[-1]))
if epoch % chkpt_freq == 0:
print('Checkpointing')
torch.save({
'epoch': epoch,
'model_state_dict': net.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': logloss,
}, './checkpoints/%s/epoch%s' % (identifier, epoch))
"""
if args.dims == 2:
print('Scatter sampling')
samples = sample(net, 2, 1000)
plt.scatter(samples[:, 0], samples[:, 1])
plt.savefig('./sample_figs/%s/epoch%s.png' %
(identifier, epoch))
plt.clf()
else:
print('Not doign scatter plot, dims > 2')
"""
print('Evaluating validation loss')
for j, val_data in enumerate(val_loader, 0):
net.zero_grad()
val_p = net(val_data, mode='pdf')
val_loss = -torch.mean(torch.log(val_p))
print('Average validation loss %s' % val_loss)