-
Notifications
You must be signed in to change notification settings - Fork 622
/
Copy pathtrain_cifar_feature_matching.py
188 lines (165 loc) · 8.95 KB
/
train_cifar_feature_matching.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
import argparse
import time
import numpy as np
import theano as th
import theano.tensor as T
from theano.sandbox.rng_mrg import MRG_RandomStreams
import lasagne
import lasagne.layers as ll
from lasagne.init import Normal
from lasagne.layers import dnn
import nn
import sys
import plotting
import cifar10_data
# settings
parser = argparse.ArgumentParser()
parser.add_argument('--seed', default=1)
parser.add_argument('--seed_data', default=1)
parser.add_argument('--count', default=400)
parser.add_argument('--batch_size', default=100)
parser.add_argument('--unlabeled_weight', type=float, default=1.)
parser.add_argument('--learning_rate', type=float, default=0.0003)
parser.add_argument('--data_dir', type=str, default='/home/tim/data/cifar-10-python')
args = parser.parse_args()
print(args)
# fixed random seeds
rng_data = np.random.RandomState(args.seed_data)
rng = np.random.RandomState(args.seed)
theano_rng = MRG_RandomStreams(rng.randint(2 ** 15))
lasagne.random.set_rng(np.random.RandomState(rng.randint(2 ** 15)))
# load CIFAR-10
trainx, trainy = cifar10_data.load(args.data_dir, subset='train')
trainx_unl = trainx.copy()
trainx_unl2 = trainx.copy()
testx, testy = cifar10_data.load(args.data_dir, subset='test')
nr_batches_train = int(trainx.shape[0]/args.batch_size)
nr_batches_test = int(testx.shape[0]/args.batch_size)
# specify generative model
noise_dim = (args.batch_size, 100)
noise = theano_rng.uniform(size=noise_dim)
gen_layers = [ll.InputLayer(shape=noise_dim, input_var=noise)]
gen_layers.append(nn.batch_norm(ll.DenseLayer(gen_layers[-1], num_units=4*4*512, W=Normal(0.05), nonlinearity=nn.relu), g=None))
gen_layers.append(ll.ReshapeLayer(gen_layers[-1], (args.batch_size,512,4,4)))
gen_layers.append(nn.batch_norm(nn.Deconv2DLayer(gen_layers[-1], (args.batch_size,256,8,8), (5,5), W=Normal(0.05), nonlinearity=nn.relu), g=None)) # 4 -> 8
gen_layers.append(nn.batch_norm(nn.Deconv2DLayer(gen_layers[-1], (args.batch_size,128,16,16), (5,5), W=Normal(0.05), nonlinearity=nn.relu), g=None)) # 8 -> 16
gen_layers.append(nn.weight_norm(nn.Deconv2DLayer(gen_layers[-1], (args.batch_size,3,32,32), (5,5), W=Normal(0.05), nonlinearity=T.tanh), train_g=True, init_stdv=0.1)) # 16 -> 32
gen_dat = ll.get_output(gen_layers[-1])
# specify discriminative model
disc_layers = [ll.InputLayer(shape=(None, 3, 32, 32))]
disc_layers.append(ll.DropoutLayer(disc_layers[-1], p=0.2))
disc_layers.append(nn.weight_norm(dnn.Conv2DDNNLayer(disc_layers[-1], 96, (3,3), pad=1, W=Normal(0.05), nonlinearity=nn.lrelu)))
disc_layers.append(nn.weight_norm(dnn.Conv2DDNNLayer(disc_layers[-1], 96, (3,3), pad=1, W=Normal(0.05), nonlinearity=nn.lrelu)))
disc_layers.append(nn.weight_norm(dnn.Conv2DDNNLayer(disc_layers[-1], 96, (3,3), pad=1, stride=2, W=Normal(0.05), nonlinearity=nn.lrelu)))
disc_layers.append(ll.DropoutLayer(disc_layers[-1], p=0.5))
disc_layers.append(nn.weight_norm(dnn.Conv2DDNNLayer(disc_layers[-1], 192, (3,3), pad=1, W=Normal(0.05), nonlinearity=nn.lrelu)))
disc_layers.append(nn.weight_norm(dnn.Conv2DDNNLayer(disc_layers[-1], 192, (3,3), pad=1, W=Normal(0.05), nonlinearity=nn.lrelu)))
disc_layers.append(nn.weight_norm(dnn.Conv2DDNNLayer(disc_layers[-1], 192, (3,3), pad=1, stride=2, W=Normal(0.05), nonlinearity=nn.lrelu)))
disc_layers.append(ll.DropoutLayer(disc_layers[-1], p=0.5))
disc_layers.append(nn.weight_norm(dnn.Conv2DDNNLayer(disc_layers[-1], 192, (3,3), pad=0, W=Normal(0.05), nonlinearity=nn.lrelu)))
disc_layers.append(nn.weight_norm(ll.NINLayer(disc_layers[-1], num_units=192, W=Normal(0.05), nonlinearity=nn.lrelu)))
disc_layers.append(nn.weight_norm(ll.NINLayer(disc_layers[-1], num_units=192, W=Normal(0.05), nonlinearity=nn.lrelu)))
disc_layers.append(ll.GlobalPoolLayer(disc_layers[-1]))
disc_layers.append(nn.weight_norm(ll.DenseLayer(disc_layers[-1], num_units=10, W=Normal(0.05), nonlinearity=None), train_g=True, init_stdv=0.1))
disc_params = ll.get_all_params(disc_layers, trainable=True)
# costs
labels = T.ivector()
x_lab = T.tensor4()
x_unl = T.tensor4()
temp = ll.get_output(gen_layers[-1], deterministic=False, init=True)
temp = ll.get_output(disc_layers[-1], x_lab, deterministic=False, init=True)
init_updates = [u for l in gen_layers+disc_layers for u in getattr(l,'init_updates',[])]
output_before_softmax_lab = ll.get_output(disc_layers[-1], x_lab, deterministic=False)
output_before_softmax_unl = ll.get_output(disc_layers[-1], x_unl, deterministic=False)
output_before_softmax_gen = ll.get_output(disc_layers[-1], gen_dat, deterministic=False)
l_lab = output_before_softmax_lab[T.arange(args.batch_size),labels]
l_unl = nn.log_sum_exp(output_before_softmax_unl)
l_gen = nn.log_sum_exp(output_before_softmax_gen)
loss_lab = -T.mean(l_lab) + T.mean(T.mean(nn.log_sum_exp(output_before_softmax_lab)))
loss_unl = -0.5*T.mean(l_unl) + 0.5*T.mean(T.nnet.softplus(l_unl)) + 0.5*T.mean(T.nnet.softplus(l_gen))
train_err = T.mean(T.neq(T.argmax(output_before_softmax_lab,axis=1),labels))
# test error
output_before_softmax = ll.get_output(disc_layers[-1], x_lab, deterministic=True)
test_err = T.mean(T.neq(T.argmax(output_before_softmax,axis=1),labels))
# Theano functions for training the disc net
lr = T.scalar()
disc_params = ll.get_all_params(disc_layers, trainable=True)
disc_param_updates = nn.adam_updates(disc_params, loss_lab + args.unlabeled_weight*loss_unl, lr=lr, mom1=0.5)
disc_param_avg = [th.shared(np.cast[th.config.floatX](0.*p.get_value())) for p in disc_params]
disc_avg_updates = [(a,a+0.0001*(p-a)) for p,a in zip(disc_params,disc_param_avg)]
disc_avg_givens = [(p,a) for p,a in zip(disc_params,disc_param_avg)]
init_param = th.function(inputs=[x_lab], outputs=None, updates=init_updates) # data based initialization
train_batch_disc = th.function(inputs=[x_lab,labels,x_unl,lr], outputs=[loss_lab, loss_unl, train_err], updates=disc_param_updates+disc_avg_updates)
test_batch = th.function(inputs=[x_lab,labels], outputs=test_err, givens=disc_avg_givens)
samplefun = th.function(inputs=[],outputs=gen_dat)
# Theano functions for training the gen net
output_unl = ll.get_output(disc_layers[-2], x_unl, deterministic=False)
output_gen = ll.get_output(disc_layers[-2], gen_dat, deterministic=False)
m1 = T.mean(output_unl,axis=0)
m2 = T.mean(output_gen,axis=0)
loss_gen = T.mean(abs(m1-m2)) # feature matching loss
gen_params = ll.get_all_params(gen_layers, trainable=True)
gen_param_updates = nn.adam_updates(gen_params, loss_gen, lr=lr, mom1=0.5)
train_batch_gen = th.function(inputs=[x_unl,lr], outputs=None, updates=gen_param_updates)
# select labeled data
inds = rng_data.permutation(trainx.shape[0])
trainx = trainx[inds]
trainy = trainy[inds]
txs = []
tys = []
for j in range(10):
txs.append(trainx[trainy==j][:args.count])
tys.append(trainy[trainy==j][:args.count])
txs = np.concatenate(txs, axis=0)
tys = np.concatenate(tys, axis=0)
# //////////// perform training //////////////
for epoch in range(1200):
begin = time.time()
lr = np.cast[th.config.floatX](args.learning_rate * np.minimum(3. - epoch/400., 1.))
# construct randomly permuted minibatches
trainx = []
trainy = []
for t in range(int(np.ceil(trainx_unl.shape[0]/float(txs.shape[0])))):
inds = rng.permutation(txs.shape[0])
trainx.append(txs[inds])
trainy.append(tys[inds])
trainx = np.concatenate(trainx, axis=0)
trainy = np.concatenate(trainy, axis=0)
trainx_unl = trainx_unl[rng.permutation(trainx_unl.shape[0])]
trainx_unl2 = trainx_unl2[rng.permutation(trainx_unl2.shape[0])]
if epoch==0:
print(trainx.shape)
init_param(trainx[:500]) # data based initialization
# train
loss_lab = 0.
loss_unl = 0.
train_err = 0.
for t in range(nr_batches_train):
ran_from = t*args.batch_size
ran_to = (t+1)*args.batch_size
ll, lu, te = train_batch_disc(trainx[ran_from:ran_to],trainy[ran_from:ran_to],
trainx_unl[ran_from:ran_to],lr)
loss_lab += ll
loss_unl += lu
train_err += te
train_batch_gen(trainx_unl2[t*args.batch_size:(t+1)*args.batch_size],lr)
loss_lab /= nr_batches_train
loss_unl /= nr_batches_train
train_err /= nr_batches_train
# test
test_err = 0.
for t in range(nr_batches_test):
test_err += test_batch(testx[t*args.batch_size:(t+1)*args.batch_size],testy[t*args.batch_size:(t+1)*args.batch_size])
test_err /= nr_batches_test
# report
print("Iteration %d, time = %ds, loss_lab = %.4f, loss_unl = %.4f, train err = %.4f, test err = %.4f" % (epoch, time.time()-begin, loss_lab, loss_unl, train_err, test_err))
sys.stdout.flush()
# generate samples from the model
sample_x = samplefun()
img_bhwc = np.transpose(sample_x[:100,], (0, 2, 3, 1))
img_tile = plotting.img_tile(img_bhwc, aspect_ratio=1.0, border_color=1.0, stretch=True)
img = plotting.plot_img(img_tile, title='CIFAR10 samples')
plotting.plt.savefig("cifar_sample_feature_match.png")
# save params
#np.savez('disc_params.npz', *[p.get_value() for p in disc_params])
#np.savez('gen_params.npz', *[p.get_value() for p in gen_params])