-
Notifications
You must be signed in to change notification settings - Fork 303
/
train_KFLaplace_MNIST.py
246 lines (190 loc) · 9.01 KB
/
train_KFLaplace_MNIST.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
from __future__ import division, print_function
import time
import torch.utils.data
from torchvision import transforms, datasets
import argparse
import matplotlib
from src.KF_Laplace.model import *
from src.KF_Laplace.hessian_operations import chol_scale_invert_kron_factor
matplotlib.use('Agg')
import matplotlib.pyplot as plt
parser = argparse.ArgumentParser(description='Train Bayesian Neural Net on MNIST with MC Dropout Variational Inference')
parser.add_argument('--weight_decay', type=float, nargs='?', action='store', default=0.01,
help='Specify the precision of an isotropic Gaussian prior. Default: 0.01')
parser.add_argument('--hessian_diag_sig', type=float, nargs='?', action='store', default=0.15,
help='Specify Gaussian prior std for a diagonal term that is added to the approximate hessian. Default: 0.15.')
parser.add_argument('--epochs', type=int, nargs='?', action='store', default=10,
help='How many epochs to train. Default: 10.')
parser.add_argument('--lr', type=float, nargs='?', action='store', default=1e-3,
help='learning rate. Default: 1e-3.')
parser.add_argument('--models_dir', type=str, nargs='?', action='store', default='KFLaplace_models',
help='Where to save learnt weights, train vectors and Hessian params. Default: \'KFLaplace_models\'.')
parser.add_argument('--results_dir', type=str, nargs='?', action='store', default='KFLaplace_results',
help='Where to save learnt training plots. Default: \'KFLaplace_results\'.')
args = parser.parse_args()
# Where to save models weights
models_dir = args.models_dir
# Where to save plots and error, accuracy vectors
results_dir = args.results_dir
mkdir(models_dir)
mkdir(results_dir)
# ------------------------------------------------------------------------------------------------------
# train config
NTrainPointsMNIST = 60000
batch_size = 128
nb_epochs = args.epochs
log_interval = 1
# ------------------------------------------------------------------------------------------------------
# dataset
cprint('c', '\nData:')
# load data
# data augmentation
transform_train = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=(0.1307,), std=(0.3081,))
])
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=(0.1307,), std=(0.3081,))
])
use_cuda = torch.cuda.is_available()
trainset = datasets.MNIST(root='../data', train=True, download=True, transform=transform_train)
valset = datasets.MNIST(root='../data', train=False, download=True, transform=transform_test)
if use_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:')
lr = args.lr
prior_sig = np.sqrt(1/args.weight_decay)
########################################################################################
net = KBayes_Net(lr=lr, channels_in=1, side_in=28, cuda=use_cuda, classes=10, n_hid=1200, batch_size=batch_size, prior_sig=prior_sig)
## ---------------------------------------------------------------------------------------------------------------------
# train
epoch = 0
cprint('c', '\nTrain:')
print(' init cost variables:')
pred_cost_train = np.zeros(nb_epochs)
err_train = np.zeros(nb_epochs)
cost_dev = np.zeros(nb_epochs)
err_dev = np.zeros(nb_epochs)
best_err = np.inf
nb_its_dev = 1
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:
cost_pred, err = net.fit(x, y)
err_train[i] += err
pred_cost_train[i] += cost_pred
nb_samples += len(x)
pred_cost_train[i] /= nb_samples
err_train[i] /= nb_samples
toc = time.time()
net.epoch = i
# ---- print
print("it %d/%d, Jtr_pred = %f, err = %f, " % (i, nb_epochs, pred_cost_train[i], err_train[i]), end="")
cprint('r', ' time: %f seconds\n' % (toc - tic))
# ---- dev
if i % nb_its_dev == 0:
net.set_mode_train(False)
nb_samples = 0
for j, (x, y) in enumerate(valloader):
cost, err, probs = net.eval(x, y)
cost_dev[i] += cost
err_dev[i] += err
nb_samples += len(x)
cost_dev[i] /= nb_samples
err_dev[i] /= nb_samples
cprint('g', ' Jdev = %f, err = %f\n' % (cost_dev[i], err_dev[i]))
if err_dev[i] < best_err:
best_err = err_dev[i]
cprint('b', 'best test error')
net.save(models_dir+'/theta_best.dat')
toc0 = time.time()
runtime_per_it = (toc0 - tic0) / float(nb_epochs)
cprint('r', ' average time: %f seconds\n' % runtime_per_it)
net.save(models_dir+'/theta_last.dat')
## ---------------------------------------------------------------------------------------------------------------------
# results
cprint('c', '\nRESULTS:')
nb_parameters = net.get_nb_parameters()
best_cost_dev = np.min(cost_dev)
best_cost_train = np.min(pred_cost_train)
err_dev_min = err_dev[::nb_its_dev].min()
print(' cost_dev: %f (cost_train %f)' % (best_cost_dev, best_cost_train))
print(' err_dev: %f' % (err_dev_min))
print(' nb_parameters: %d (%s)' % (nb_parameters, humansize(nb_parameters)))
print(' time_per_it: %fs\n' % (runtime_per_it))
## Save results for plots
np.save(results_dir + '/cost_train.npy', pred_cost_train)
np.save(results_dir + '/cost_dev.npy', cost_dev)
np.save(results_dir + '/err_train.npy', err_train)
np.save(results_dir + '/err_dev.npy', err_dev)
## Time to do Laplace Approximation.
print('MAP configuration reached: Calculating block diagonal Hessian.')
# Get hessian factors
EQ1, EHH1, MAP1, EQ2, EHH2, MAP2, EQ3, EHH3, MAP3 = net.get_K_laplace_params(trainloader)
h_params = [EQ1, EHH1, MAP1, EQ2, EHH2, MAP2, EQ3, EHH3, MAP3]
save_object(h_params, models_dir+'/block_hessian_params.pkl')
print('Hessian Parameters Saved')
data_scale = np.sqrt(len(trainset))
prior_sig = args.hessian_diag_sig
prior_prec = 1/prior_sig**2
prior_scale = np.sqrt(prior_prec)
# Scale and invert factors
# upper_Qinv, lower_HHinv
print('Scaling and inverting Hessian factors')
scale_inv_EQ1 = chol_scale_invert_kron_factor(EQ1, prior_scale, data_scale, upper=True)
scale_inv_EHH1 = chol_scale_invert_kron_factor(EHH1, prior_scale, data_scale, upper=False)
scale_inv_EQ2 = chol_scale_invert_kron_factor(EQ2, prior_scale, data_scale, upper=True)
scale_inv_EHH2 = chol_scale_invert_kron_factor(EHH2, prior_scale, data_scale, upper=False)
scale_inv_EQ3 = chol_scale_invert_kron_factor(EQ3, prior_scale, data_scale, upper=True)
scale_inv_EHH3 = chol_scale_invert_kron_factor(EHH3, prior_scale, data_scale, upper=False)
## ---------------------------------------------------------------------------------------------------------------------
# fig cost vs its
textsize = 15
marker = 5
plt.figure(dpi=100)
fig, ax1 = plt.subplots()
ax1.plot(range(0, nb_epochs, nb_its_dev), cost_dev[::nb_its_dev], 'b-')
ax1.plot(pred_cost_train, '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 error', 'train error'], 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(results_dir + '/cost.png', bbox_extra_artists=(lgd,), bbox_inches='tight')
plt.figure(dpi=100)
fig2, ax2 = plt.subplots()
ax2.set_ylabel('% error')
ax2.semilogy(range(0, nb_epochs, nb_its_dev), 100 * err_dev[::nb_its_dev], 'b-')
ax2.semilogy(100 * err_train, 'r--')
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(results_dir + '/err.png', bbox_extra_artists=(lgd,), box_inches='tight')