-
Notifications
You must be signed in to change notification settings - Fork 1
/
test_vae_mnist.py
314 lines (270 loc) · 12.7 KB
/
test_vae_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
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
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
# test vae on mnist
import theano
import theano.tensor as T
import numpy as np
from vae import Qsampler, VAEModel
from samples_save import ImagesSamplesSave
from blocks.initialization import Constant, NdarrayInitialization, Sparse, Orthogonal
from blocks.bricks import MLP, Tanh, Rectifier
from blocks.bricks.cost import BinaryCrossEntropy, CategoricalCrossEntropy
from blocks.graph import ComputationGraph
from blocks.extensions import FinishAfter, Printing
from blocks.extensions.monitoring import DataStreamMonitoring, TrainingDataMonitoring
from blocks.model import Model
from fuel.streams import DataStream
from fuel.schemes import SequentialScheme
from blocks.main_loop import MainLoop
from blocks.algorithms import Momentum, RMSProp, Scale
from fuel.transformers import Flatten
from blocks.algorithms import GradientDescent
from contextlib import closing
floatX = theano.config.floatX
from fuel.datasets import MNIST
def test_vae():
activation = Rectifier()
full_weights_init = Orthogonal()
weights_init = full_weights_init
layers = [784, 400, 20]
encoder_layers = layers[:-1]
encoder_mlp = MLP([activation] * (len(encoder_layers)-1),
encoder_layers,
name="MLP_enc", biases_init=Constant(0.), weights_init=weights_init)
enc_dim = encoder_layers[-1]
z_dim = layers[-1]
#sampler = Qlinear(input_dim=enc_dim, output_dim=z_dim, biases_init=Constant(0.), weights_init=full_weights_init)
sampler = Qsampler(input_dim=enc_dim, output_dim=z_dim, biases_init=Constant(0.), weights_init=full_weights_init)
decoder_layers = layers[:] ## includes z_dim as first layer
decoder_layers.reverse()
decoder_mlp = MLP([activation] * (len(decoder_layers)-2) + [Rectifier()],
decoder_layers,
name="MLP_dec", biases_init=Constant(0.), weights_init=weights_init)
vae = VAEModel(encoder_mlp, sampler, decoder_mlp)
vae.initialize()
x = T.matrix('features')
batch_size = 124
x_recons, kl_terms = vae.reconstruct(x)
recons_term = BinaryCrossEntropy().apply(x, T.clip(x_recons, 1e-5, 1 - 1e-5))
recons_term.name = "recons_term"
cost = recons_term + kl_terms.mean()
cost.name = "cost"
cg = ComputationGraph(cost)
temp = cg.parameters
for t, i in zip(temp, range(len(temp))):
t.name = t.name+str(i)+"vae_mnist"
step_rule = RMSProp(0.001, 0.95)
train_set = MNIST('train')
train_set.sources = ("features", )
test_set = MNIST("test")
test_set.sources = ("features", )
data_stream = Flatten(DataStream.default_stream(
train_set, iteration_scheme=SequentialScheme(train_set.num_examples, batch_size)))
data_stream_monitoring = Flatten(DataStream.default_stream(
train_set, iteration_scheme=SequentialScheme(train_set.num_examples, batch_size)))
data_stream_test = Flatten(DataStream.default_stream(
test_set, iteration_scheme=SequentialScheme(test_set.num_examples, batch_size)))
algorithm = GradientDescent(cost=cost, params=cg.parameters,
step_rule=step_rule)
monitor_train = TrainingDataMonitoring(
variables=[cost], prefix="train", every_n_batches=10)
monitor_valid = DataStreamMonitoring(
variables=[cost], data_stream=data_stream_test, prefix="valid", every_n_batches=10)
# drawing_samples = ImagesSamplesSave("../data_mnist", vae, (28, 28), every_n_epochs=1)
extensions = [ monitor_train,
monitor_valid,
FinishAfter(after_n_batches=1500),
Printing(every_n_batches=10)
]
main_loop = MainLoop(data_stream=data_stream,
algorithm=algorithm, model = Model(cost),
extensions=extensions)
main_loop.run()
from blocks.serialization import dump
with closing(open('../data_mnist/model_0', 'w')) as f:
dump(vae, f)
if __name__ == '__main__':
test_vae()
"""
def main(name, model, epochs, batch_size, learning_rate, bokeh, layers, gamma,
rectifier, predict, dropout, qlinear, sparse):
runname = "vae%s-L%s%s%s%s-l%s-g%s-b%d" % (name, layers,
'r' if rectifier else '',
'd' if dropout else '',
'l' if qlinear else '',
shnum(learning_rate), shnum(gamma), batch_size//100)
if rectifier:
activation = Rectifier()
full_weights_init = Orthogonal()
else:
activation = Tanh()
full_weights_init = Orthogonal()
if sparse:
runname += '-s%d'%sparse
weights_init = Sparse(num_init=sparse, weights_init=full_weights_init)
else:
weights_init = full_weights_init
layers = map(int,layers.split(','))
encoder_layers = layers[:-1]
encoder_mlp = MLP([activation] * (len(encoder_layers)-1),
encoder_layers,
name="MLP_enc", biases_init=Constant(0.), weights_init=weights_init)
enc_dim = encoder_layers[-1]
z_dim = layers[-1]
if qlinear:
sampler = Qlinear(input_dim=enc_dim, output_dim=z_dim, biases_init=Constant(0.), weights_init=full_weights_init)
else:
sampler = Qsampler(input_dim=enc_dim, output_dim=z_dim, biases_init=Constant(0.), weights_init=full_weights_init)
decoder_layers = layers[:] ## includes z_dim as first layer
decoder_layers.reverse()
decoder_mlp = MLP([activation] * (len(decoder_layers)-2) + [Sigmoid()],
decoder_layers,
name="MLP_dec", biases_init=Constant(0.), weights_init=weights_init)
vae = VAEModel(encoder_mlp, sampler, decoder_mlp)
vae.initialize()
x = tensor.matrix('features')
if predict:
mean_z, enc = vae.mean_z(x)
# cg = ComputationGraph([mean_z, enc])
newmodel = Model([mean_z,enc])
else:
x_recons, kl_terms = vae.reconstruct(x)
recons_term = BinaryCrossEntropy().apply(x, x_recons)
recons_term.name = "recons_term"
cost = recons_term + kl_terms.mean()
cg = ComputationGraph([cost])
if gamma > 0:
weights = VariableFilter(roles=[WEIGHT])(cg.variables)
cost += gamma * blocks.theano_expressions.l2_norm(weights)
cost.name = "nll_bound"
newmodel = Model(cost)
if dropout:
weights = [v for k,v in newmodel.get_params().iteritems()
if k.find('MLP')>=0 and k.endswith('.W') and not k.endswith('MLP_enc/linear_0.W')]
cg = apply_dropout(cg,weights,0.5)
target_cost = cg.outputs[0]
else:
target_cost = cost
if name == 'mnist':
if predict:
train_ds = MNIST("train")
else:
train_ds = MNIST("train", sources=['features'])
test_ds = MNIST("test")
else:
datasource_dir = os.path.join(fuel.config.data_path, name)
datasource_fname = os.path.join(datasource_dir , name+'.hdf5')
if predict:
train_ds = H5PYDataset(datasource_fname, which_set='train')
else:
train_ds = H5PYDataset(datasource_fname, which_set='train', sources=['features'])
test_ds = H5PYDataset(datasource_fname, which_set='test')
train_s = DataStream(train_ds,
iteration_scheme=SequentialScheme(
train_ds.num_examples, batch_size))
test_s = DataStream(test_ds,
iteration_scheme=SequentialScheme(
test_ds.num_examples, batch_size))
if predict:
from itertools import chain
fprop = newmodel.get_theano_function()
allpdata = None
alledata = None
f = train_s.sources.index('features')
assert f == test_s.sources.index('features')
sources = test_s.sources
alllabels = dict((s,[]) for s in sources if s != 'features')
for data in chain(train_s.get_epoch_iterator(), test_s.get_epoch_iterator()):
for s,d in zip(sources,data):
if s != 'features':
alllabels[s].extend(list(d))
pdata, edata = fprop(data[f])
if allpdata is None:
allpdata = pdata
else:
allpdata = np.vstack((allpdata, pdata))
if alledata is None:
alledata = edata
else:
alledata = np.vstack((alledata, edata))
print 'Saving',allpdata.shape,'intermidiate layer, for all training and test examples, to',name+'_z.npy'
np.save(name+'_z', allpdata)
print 'Saving',alledata.shape,'last encoder layer to',name+'_e.npy'
np.save(name+'_e', alledata)
print 'Saving additional labels/targets:',','.join(alllabels.keys()),
print ' of size',','.join(map(lambda x: str(len(x)),alllabels.values())),
print 'to',name+'_labels.pkl'
with open(name+'_labels.pkl','wb') as fp:
pickle.dump(alllabels, fp, -1)
else:
algorithm = GradientDescent(
cost=target_cost, params=cg.parameters,
step_rule=Adam(learning_rate) # Scale(learning_rate=learning_rate)
)
extensions = []
if model:
extensions.append(LoadFromDump(model))
extensions += [Timing(),
FinishAfter(after_n_epochs=epochs),
DataStreamMonitoring(
[recons_term, cost],
test_s,
prefix="test"),
TrainingDataMonitoring(
[cost,
aggregation.mean(algorithm.total_gradient_norm)],
prefix="train",
after_epoch=True),
Dump(runname, every_n_epochs=10),
Printing()]
if bokeh:
extensions.append(Plot(
'Auto',
channels=[
['test_recons_term','test_nll_bound','train_nll_bound'
],
['train_total_gradient_norm']]))
main_loop = MainLoop(
algorithm,
train_s,
model=newmodel,
extensions=extensions)
main_loop.run()
if __name__ == "__main__":
logging.basicConfig(level=logging.INFO)
parser = ArgumentParser("An example of training a Variational-Autoencoder.")
parser.add_argument("--name", default="mnist",
help="name of hdf5 data set")
parser.add_argument("--model",
help="start model to read")
parser.add_argument("--epochs", type=int, default=1000,
help="Number of training epochs to do.")
parser.add_argument("--bs", "--batch-size", type=int, dest="batch_size",
default=500, help="Size of each mini-batch")
parser.add_argument("--lr", "--learning-rate", type=float, dest="learning_rate",
default=1e-3, help="Learning rate")
parser.add_argument("--bokeh", action='store_true', default=False,
help="Set if you want to use Bokeh ")
parser.add_argument("--layers",
default="784,100,20", help="number of units in each layer of the encoder"
" (use 784, on first layer, for mnist.)"
" The last number (e.g. 20) is the dimension of the intermidiate layer."
" The decoder has the same layers as the encoder but in reverse"
" (e.g. 100, 784)")
parser.add_argument("--gamma", type=float,
default=3e-4, help="L2 weight")
parser.add_argument("-r","--rectifier",action='store_true',default=False,
help="Use RELU activation on hidden (default Tanh)")
parser.add_argument("-p","--predict",action='store_true',default=False,
help="Generate prediction of the intermidate layer and last layer of the encoder"
" instead of training."
" You must supply a pre-trained model and define all parameters to be the same"
" as in training. ")
parser.add_argument("-d","--dropout",action='store_true',default=False,
help="Use dropout")
parser.add_argument("-l","--qlinear",action='store_true',default=False,
help="Perform a deterministic linear transformation instead of sampling"
" on the intermidiate layer")
parser.add_argument("-s","--sparse",type=int,
help="Use sparse weight initialization. Give the number of non zero weights")
args = parser.parse_args()
main(**vars(args))
"""