-
Notifications
You must be signed in to change notification settings - Fork 5
/
train_BNN.py
290 lines (250 loc) · 12.1 KB
/
train_BNN.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
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
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
from __future__ import print_function
from __future__ import division
import torch
import time
import numpy as np
import torch.utils.data
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from test_tube import Experiment
from src.utils import mkdir, cprint
def train_VI_classification(net, name, save_dir, batch_size, nb_epochs, trainset, valset, cuda,
flat_ims=False, nb_its_dev=1, early_stop=None, load_path=None, save_freq=20,
stop_criteria='test_ELBO', tags=None, show=False):
exp = Experiment(name=name,
debug=False,
save_dir=save_dir,
autosave=True)
if load_path is not None:
net.load(load_path)
exp_version = exp.version
media_dir = exp.get_media_path(name, exp_version)
models_dir = exp.get_data_path(name, exp_version) + '/models'
mkdir(models_dir)
exp.tag({
'n_layers': net.model.n_layers,
'batch_size': batch_size,
'init_lr': net.lr,
'lr_schedule': net.schedule,
'nb_epochs': nb_epochs,
'early_stop': early_stop,
'stop_criteria': stop_criteria,
'nb_its_dev': nb_its_dev,
'model_loaded': load_path,
'cuda': cuda,
})
if net.model.__class__.__name__ == 'arq_uncert_conv2d_resnet':
exp.tag({
'outer_width': net.model.outer_width,
'inner_width': net.model.inner_width
})
else:
exp.tag({'width': net.model.width})
exp.tag({
'prob_model': net.model.prob_model.name,
'prob_model_summary': net.model.prob_model.summary
})
if tags is not None:
exp.tag(tags)
if cuda:
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True, pin_memory=True,
num_workers=3)
valloader = torch.utils.data.DataLoader(valset, batch_size=batch_size, shuffle=False, pin_memory=True,
num_workers=3)
else:
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True, pin_memory=False,
num_workers=3)
valloader = torch.utils.data.DataLoader(valset, batch_size=batch_size, shuffle=False, pin_memory=False,
num_workers=3)
## ---------------------------------------------------------------------------------------------------------------------
# net dims
cprint('c', '\nNetwork:')
epoch = 0
## ---------------------------------------------------------------------------------------------------------------------
# train
cprint('c', '\nTrain:')
print(' init cost variables:')
mloglike_train = np.zeros(nb_epochs)
KL_train = np.zeros(nb_epochs)
ELBO_train = np.zeros(nb_epochs)
ELBO_test = np.zeros(nb_epochs)
err_train = np.zeros(nb_epochs)
mloglike_dev = np.zeros(nb_epochs)
err_dev = np.zeros(nb_epochs)
best_epoch = 0
best_train_ELBO = -np.inf
best_test_ELBO = -np.inf
best_dev_ll = -np.inf
tic0 = time.time()
for i in range(epoch, nb_epochs):
net.set_mode_train(True)
tic = time.time()
nb_samples = 0
for x, y in trainloader:
if flat_ims:
x = x.view(x.shape[0], -1)
KL, minus_loglike, err = net.fit(x, y)
err_train[i] += err
mloglike_train[i] += minus_loglike/len(trainloader)
KL_train[i] += KL/len(trainloader)
nb_samples += len(x)
# mloglike_train[i] *= nb_samples
# KL_train[i] *= nb_samples
ELBO_train[i] = (-KL_train[i] - mloglike_train[i]) * nb_samples
err_train[i] /= nb_samples
toc = time.time()
# ---- print
print("it %d/%d, sample minus loglike = %f, sample KL = %.10f, err = %f, ELBO = %f" % \
(i, nb_epochs, mloglike_train[i], KL_train[i], err_train[i], ELBO_train[i]), end="")
exp.log({'epoch': i, 'MLL': mloglike_train[i], 'KLD': KL_train[i], 'err': err_train[i], 'ELBO': ELBO_train[i]})
cprint('r', ' time: %f seconds\n' % (toc - tic))
net.update_lr(i, 0.1)
# ---- dev
if i % nb_its_dev == 0:
tic = time.time()
nb_samples = 0
for j, (x, y) in enumerate(valloader):
if flat_ims:
x = x.view(x.shape[0], -1)
minus_loglike, err = net.eval(x, y)
mloglike_dev[i] += minus_loglike / len(valloader)
err_dev[i] += err
nb_samples += len(x)
ELBO_test[i] = (-KL_train[i] - mloglike_dev[i]) * nb_samples
ELBO_test[i] = (-KL_train[i] - mloglike_dev[i]) * nb_samples
err_dev[i] /= nb_samples
toc = time.time()
cprint('g', ' sample minus loglike = %f, err = %f, ELBO = %f\n' % (mloglike_dev[i], err_dev[i], ELBO_test[i]), end="")
cprint('g', ' (prev best it = %i, sample minus loglike = %f, ELBO = %f)\n' % (best_epoch, best_dev_ll, best_test_ELBO), end="")
cprint('g', ' time: %f seconds\n' % (toc - tic))
exp.log({'epoch': i, 'MLL_val': mloglike_dev[i], 'err_val': err_dev[i], 'ELBO_val': ELBO_test[i]})
if stop_criteria == 'test_LL' and -mloglike_dev[i] > best_dev_ll:
best_dev_ll = -mloglike_dev[i]
best_epoch = i
cprint('b', 'best test loglike: %d' % best_dev_ll)
net.save(models_dir + '/theta_best.dat')
probs = net.model.prob_model.get_q_probs().data.cpu().numpy()
cuttoff = np.max(probs)*0.95
exp.tag({"q_vec": net.model.get_q_vector().cpu().detach().numpy(),
"q_probs": net.model.prob_model.get_q_probs().cpu().detach().numpy(),
"expected_depth": np.sum(probs * np.arange(net.model.n_layers + 1)),
"95th_depth": np.argmax(probs > cuttoff), "best_epoch": best_epoch,
"best_dev_ll": best_dev_ll})
if stop_criteria == 'test_ELBO' and ELBO_test[i] > best_test_ELBO:
best_test_ELBO = ELBO_test[i]
best_epoch = i
cprint('b', 'best test ELBO: %d' % best_test_ELBO)
net.save(models_dir + '/theta_best.dat')
probs = net.model.prob_model.get_q_probs().data.cpu().numpy()
cuttoff = np.max(probs)*0.95
exp.tag({"q_vec": net.model.get_q_vector().cpu().detach().numpy(),
"q_probs": net.model.prob_model.get_q_probs().cpu().detach().numpy(),
"expected_depth": np.sum(probs * np.arange(net.model.n_layers + 1)),
"95th_depth": np.argmax(probs > cuttoff),
"best_epoch": best_epoch,
"best_test_ELBO": best_test_ELBO})
if stop_criteria == 'train_ELBO' and ELBO_train[i] > best_train_ELBO:
best_train_ELBO = ELBO_train[i]
best_epoch = i
cprint('b', 'best train ELBO: %d' % best_train_ELBO)
net.save(models_dir + '/theta_best.dat')
probs = net.model.prob_model.get_q_probs().data.cpu().numpy()
cuttoff = np.max(probs)*0.95
exp.tag({"q_vec": net.model.get_q_vector().cpu().detach().numpy(),
"q_probs": net.model.prob_model.get_q_probs().cpu().detach().numpy(),
"expected_depth": np.sum(probs * np.arange(net.model.n_layers + 1)),
"95th_depth": np.argmax(probs > cuttoff),
"best_epoch": best_epoch,
"best_train_ELBO": best_train_ELBO})
if save_freq is not None and i % save_freq == 0:
exp.tag({
"final_q_vec": net.model.get_q_vector().cpu().detach().numpy(),
"final_q_probs": net.model.prob_model.get_q_probs().cpu().detach().numpy(),
"final_expected_depth": np.sum(net.model.prob_model.get_q_probs().data.cpu().numpy() * np.arange(net.model.n_layers + 1))
})
net.save(models_dir + '/theta_last.dat')
if early_stop is not None and (i - best_epoch) > early_stop:
exp.tag({"early_stop_epoch": i})
cprint('r', ' stopped early!\n')
break
toc0 = time.time()
runtime_per_it = (toc0 - tic0) / float(i + 1)
cprint('r', ' average time: %f seconds\n' % runtime_per_it)
## ---------------------------------------------------------------------------------------------------------------------
# fig cost vs its
textsize = 15
marker = 5
plt.figure(dpi=100)
fig, ax1 = plt.subplots()
ax1.plot(range(0, i, nb_its_dev), np.clip(mloglike_dev[:i:nb_its_dev], a_min=-5, a_max=5), 'b-')
ax1.plot(np.clip(mloglike_train[:i], a_min=-5, a_max=5), 'r--')
ax1.set_ylabel('Cross Entropy')
plt.xlabel('epoch')
plt.grid(b=True, which='major', color='k', linestyle='-')
plt.grid(b=True, which='minor', color='k', linestyle='--')
lgd = plt.legend(['test', 'train'], markerscale=marker, prop={'size': textsize, 'weight': 'normal'})
ax = plt.gca()
plt.title('classification costs')
for item in ([ax.title, ax.xaxis.label, ax.yaxis.label] +
ax.get_xticklabels() + ax.get_yticklabels()):
item.set_fontsize(textsize)
item.set_weight('normal')
plt.savefig(media_dir + '/cost.png', bbox_extra_artists=(lgd,), bbox_inches='tight')
if show:
plt.show()
plt.figure(dpi=100)
fig, ax1 = plt.subplots()
ax1.plot(range(0, i), KL_train[:i], 'b-')
ax1.set_ylabel('KL')
plt.xlabel('epoch')
plt.grid(b=True, which='major', color='k', linestyle='-')
plt.grid(b=True, which='minor', color='k', linestyle='--')
lgd = plt.legend(['KL'], markerscale=marker, prop={'size': textsize, 'weight': 'normal'})
ax = plt.gca()
plt.title('KL divideed by number of samples')
for item in ([ax.title, ax.xaxis.label, ax.yaxis.label] +
ax.get_xticklabels() + ax.get_yticklabels()):
item.set_fontsize(textsize)
item.set_weight('normal')
plt.savefig(media_dir + '/KL.png', bbox_extra_artists=(lgd,), bbox_inches='tight')
if show:
plt.show()
plt.figure(dpi=100)
fig, ax1 = plt.subplots()
ax1.plot(range(0, i), ELBO_train[:i], 'b-')
ax1.set_ylabel('nats')
plt.xlabel('epoch')
plt.grid(b=True, which='major', color='k', linestyle='-')
plt.grid(b=True, which='minor', color='k', linestyle='--')
lgd = plt.legend(['ELBO'], markerscale=marker, prop={'size': textsize, 'weight': 'normal'})
ax = plt.gca()
plt.title('ELBO')
for item in ([ax.title, ax.xaxis.label, ax.yaxis.label] +
ax.get_xticklabels() + ax.get_yticklabels()):
item.set_fontsize(textsize)
item.set_weight('normal')
plt.savefig(media_dir + '/ELBO.png', bbox_extra_artists=(lgd,), bbox_inches='tight')
if show:
plt.show()
plt.figure(dpi=100)
fig, ax2 = plt.subplots()
ax2.set_ylabel('% error')
ax2.semilogy(range(0, i, nb_its_dev), err_dev[:i:nb_its_dev], 'b-')
ax2.semilogy(err_train[:i], 'r--')
ax2.set_ylim(top=1, bottom=1e-3)
plt.xlabel('epoch')
plt.grid(b=True, which='major', color='k', linestyle='-')
plt.grid(b=True, which='minor', color='k', linestyle='--')
ax2.get_yaxis().set_minor_formatter(matplotlib.ticker.ScalarFormatter())
ax2.get_yaxis().set_major_formatter(matplotlib.ticker.ScalarFormatter())
lgd = plt.legend(['test error', 'train error'], markerscale=marker, prop={'size': textsize, 'weight': 'normal'})
ax = plt.gca()
for item in ([ax.title, ax.xaxis.label, ax.yaxis.label] +
ax.get_xticklabels() + ax.get_yticklabels()):
item.set_fontsize(textsize)
item.set_weight('normal')
plt.savefig(media_dir + '/err.png', bbox_extra_artists=(lgd,), box_inches='tight')
if show:
plt.show()
return exp, mloglike_train, KL_train, ELBO_train, err_train, mloglike_dev, err_dev