-
Notifications
You must be signed in to change notification settings - Fork 0
/
common_cause.py
287 lines (234 loc) · 10.9 KB
/
common_cause.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
## Code accompanying the paper Design Motifs for Generative Design
## Provided for research use only
from __future__ import division, absolute_import
import sys
import matplotlib
import argparse
import lasagne
from helpers.data import clip, write_model
from tqdm import tqdm
from helpers.mnist import load_mnist
from lasagne.updates import adam
import lasagne.layers as layers
from theano.tensor.shared_randomstreams import RandomStreams
import numpy as np
import theano
import theano.tensor as T
theano.config.compute_test_value = 'warn' # 'warn' runs test values
matplotlib.use('Agg')
tanh = lasagne.nonlinearities.tanh
sigmoid = lasagne.nonlinearities.sigmoid
linear = lasagne.nonlinearities.linear
relu = lasagne.nonlinearities.rectify
softmax = lasagne.nonlinearities.softmax
NETWORK_DIM = 1024
LABEL_DIM = 1
# LABEL_DIM = 10 # e.g. labelled MNIST
DATA_DIM = 784
SAVEPATH = './'
np.random.seed(42)
srng = RandomStreams(42)
shared = lambda X: theano.shared(np.asarray(X, dtype=theano.config.floatX))
def unpack_params(params, latent_dim):
return params[:, :latent_dim], params[:, latent_dim:]
def encoder_l(latent_dim, input_var=None):
# input is concatenation of MNIST digit and one-hot encoded label
input = layers.InputLayer(shape=(None, DATA_DIM+LABEL_DIM),
input_var=input_var)
h1 = lasagne.layers.DenseLayer(input, NETWORK_DIM, nonlinearity=tanh)
h2 = lasagne.layers.DenseLayer(h1, NETWORK_DIM, nonlinearity=tanh)
h3 = lasagne.layers.DenseLayer(h2, NETWORK_DIM, nonlinearity=tanh)
mu = lasagne.layers.DenseLayer(h3, latent_dim, nonlinearity=linear)
log_std = lasagne.layers.DenseLayer(h3, latent_dim, nonlinearity=linear)
return mu, log_std
def encoder_u(latent_dim, input_var=None):
# input is concatenation of MNIST digit and one-hot encoded label
input = layers.InputLayer(shape=(None, DATA_DIM), input_var=input_var)
h1 = lasagne.layers.DenseLayer(input, NETWORK_DIM, nonlinearity=tanh)
h2 = lasagne.layers.DenseLayer(h1, NETWORK_DIM, nonlinearity=tanh)
h3 = lasagne.layers.DenseLayer(h2, NETWORK_DIM, nonlinearity=tanh)
mu = lasagne.layers.DenseLayer(h3, latent_dim, nonlinearity=linear)
log_std = lasagne.layers.DenseLayer(h3, latent_dim, nonlinearity=linear)
return mu, log_std
def generator_x(n_hidden, input_var=None):
# parameterize p(x | z) network
input = layers.InputLayer(shape=(None, n_hidden), input_var=input_var)
h1 = lasagne.layers.DenseLayer(input, NETWORK_DIM, nonlinearity=tanh)
h2 = lasagne.layers.DenseLayer(h1, NETWORK_DIM, nonlinearity=tanh)
h3 = lasagne.layers.DenseLayer(h2, NETWORK_DIM, nonlinearity=tanh)
return lasagne.layers.DenseLayer(h3, DATA_DIM, nonlinearity=sigmoid)
def linear_regression(input_dim, input_var=None):
input = lasagne.layers.InputLayer(shape=(None, input_dim),
input_var=input_var)
return lasagne.layers.DenseLayer(input, 1, nonlinearity=linear)
def multiclass_logistic(input_dim, num_classes, input_var=None): # i.e., gen_t
# add L2 regularization
input = lasagne.layers.InputLayer(shape=(None, input_dim),
input_var=input_var)
return lasagne.layers.DenseLayer(input, num_classes, nonlinearity=softmax)
def sample_q(mu, log_std, s=42):
if "gpu" in theano.config.device:
rng = theano.sandbox.cuda.rng_curand.CURAND_RandomStreams(seed=s)
else:
rng = T.shared_randomstreams.RandomStreams(seed=s)
eps = rng.normal(mu.shape)
z = mu + T.exp(0.5 * log_std) * eps
return z
def lower_bound_l(enc, gen_x, gen_t, data, targets, scale_l):
X_and_t = T.concatenate([data, targets], axis=1)
mu, log_std = layers.get_output(enc, X_and_t)
zIxt = sample_q(mu, log_std)
targetsIz = layers.get_output(gen_t, zIxt) # ? x 1
xIz = layers.get_output(gen_x, zIxt)
log_ptIz = gaussian_log_likelihood(data, 0.03, targetsIz, 1)
log_pxIz = bernoulli_log_likelihood(xIz, data)
KL_qIIp = -0.5 * T.sum(1 + log_std - mu ** 2 - T.exp(log_std), axis=1)
loss = - T.mean(log_pxIz + log_ptIz - KL_qIIp) # N_l * mean(lower_bound_l)
recons_loss = T.mean(T.sum(T.square(data - xIz), axis=1))
return loss, recons_loss
def lower_bound_u(enc, gen, data, scale_u):
mu, log_std = layers.get_output(enc, data)
latents = sample_q(mu, log_std)
generated = layers.get_output(gen, latents)
log_pxIz = bernoulli_log_likelihood(generated, data)
KL_qIIp = -0.5 * T.sum(1 + log_std - mu ** 2 - T.exp(log_std), axis=1)
loss = -T.mean(log_pxIz - KL_qIIp)
recons_loss = T.mean(T.sum(T.square(data - generated), axis=1))
return loss, recons_loss
def main(step_size, batch_size, n_epochs, save, plot, latent_dim,
proportion_labelled):
# data, targets = load_mnist('train', label=True) # class labels
data, _ = load_mnist('train', label=True)
# brightness target function
targets = data.reshape(-1, 784).sum(axis=-1, keepdims=True)
n = data.shape[0]
n_valid = 10000
n_train = n - n_valid
n_batches = n_train//batch_size
bs_l = int(batch_size*proportion_labelled)
bs_u = int(batch_size*(1-proportion_labelled))
# calculate weightings for sup and unsup losses
scale_l = bs_l / batch_size
scale_u = bs_u / batch_size
assert(scale_l + scale_u == 1)
X_train = data[:n_train, :].reshape(-1, 784)
X_valid = data[n_train:, :].reshape(-1, 784)
t_train = targets[:n_train]
t_valid = targets[n_train:]
n_u = int(n_train * (1 - proportion_labelled))
scope = 'main/' # track scope for debugging
X_u = T.fmatrix(scope + 'X_unlabelled')
X_l = T.fmatrix(scope + 'X_labelled')
t_l = T.fmatrix(scope + 'targets')
X_u.tag.test_value = np.random.rand(100, DATA_DIM).astype(np.float32)
X_l.tag.test_value = np.random.rand(100, DATA_DIM).astype(np.float32)
t_l.tag.test_value = np.random.randint(0, 783, (100, 1)).astype(np.float32)
# construct unsupervised lower bound
# q(z | x) network
enc_zIx = encoder_u(latent_dim)
# p(x | z) network
gen_xIz = generator_x(latent_dim)
loss_u, recons_loss_u = lower_bound_u(enc_zIx, gen_xIz, X_u, scale_u)
# construct supervised lower bound
# p(t | z) network
enc_zIxt = encoder_l(latent_dim)
# gen_tIz = multiclass_logistic(latent_dim, num_classes)
gen_tIz = linear_regression(latent_dim) # input is ? x latent_dim
if proportion_labelled > 0:
loss_l, recons_loss_l = lower_bound_l(enc_zIxt, gen_xIz, gen_tIz,
X_l, t_l, scale_l)
else:
loss_l = 0
recons_loss_l = 0
# SSVAE lower bound
loss = (loss_u + loss_l) #/ batch_size
valid_loss = (loss_u + loss_l) #/ n_valid
recons_loss = (recons_loss_l + recons_loss_u) / batch_size
valid_loss = [loss, recons_loss]
# build symbolic sampler for learned generator
noise = T.fmatrix()
noise.tag.test_value = np.random.rand(100, latent_dim).astype(np.float32)
gen_im = layers.get_output(gen_xIz, noise)
gen_fun = theano.function([noise], gen_im)
# learn SSVAE parameters
idx = T.iscalar()
idx.tag.test_value = 0
params = layers.get_all_params(enc_zIx) +\
layers.get_all_params(gen_xIz)
if proportion_labelled > 0:
params += layers.get_all_params(enc_zIxt) +\
layers.get_all_params(gen_tIz)
# split train data into sup/unsup
optimize_params = adam(loss, params, learning_rate=step_size)
train_u = shared(X_train[:n_u, :])
train_l = shared(X_train[n_u:, :])
train_t = shared(t_train[n_u:])
train_dict = {X_u: train_u[idx*bs_u:(idx+1)*bs_u],
X_l: train_l[idx*bs_l:(idx+1)*bs_l],
t_l: train_t[idx*bs_l:(idx+1)*bs_l]}
# split validation data into sup/unsup
n_v_u = int(n_valid * (1 - proportion_labelled))
valid_u = shared(X_valid[:n_v_u, :])
valid_l = shared(X_valid[n_v_u:, :])
valid_t = shared(t_valid[n_v_u:])
valid_dict = {X_u: valid_u, X_l: valid_l,
t_l: valid_t}
# build function to forward propagate data and update weights in SSVAE
train = theano.function([idx], [loss, recons_loss], updates=optimize_params,
givens=train_dict, on_unused_input='warn')
validate = theano.function([], valid_loss, givens=valid_dict,
on_unused_input='warn')
losses = {"train": [], "valid": []}
train_loss = None
model_name='t:reg_bMNISTz'+str(latent_dim)+'bs'+\
str(batch_size)+'lr'+str(step_size)+\
'l%'+str(proportion_labelled) + 'joint'
for e in tqdm(range(n_epochs)):
for i in tqdm(range(n_batches)):
train_loss = train(i)
# sanity test
if np.isnan(train_loss[0]):
print ("NaN detected!")
sys.stdout.flush()
break
valid_loss = validate()
losses["train"].append(train_loss[0])
losses["valid"].append(valid_loss[0])
# write progress to STDOUT
if e%5 == 0:
print("epoch "+str(e)+" train_loss: " + str(train_loss[0])
+ " train_recon_loss: "+str(train_loss[1]))
sys.stdout.flush()
print(" valid_loss: "+str(valid_loss[0]) + " valid_recon_loss: "
+ str(valid_loss[1]))
sys.stdout.flush()
# checkpoint model parameters
if save and (e+1)%5 == 0:
write_model(gen_xIz, model_name + 'genxIz', e, SAVEPATH)
write_model(enc_zIx, model_name + 'enczIx', e, SAVEPATH)
write_model(enc_zIxt, model_name + 'enczIxt', e, SAVEPATH)
write_model(gen_tIz, model_name + 'gentIz', e, SAVEPATH)
def gaussian_log_likelihood(X_data, fixed_var, generated, input_dim):
return -T.sum(T.square(generated - X_data), [-1]) / (2 * fixed_var) \
- input_dim / 2 * np.log(2 * np.pi) \
- input_dim / 2 * np.log(fixed_var)
def bernoulli_log_likelihood(generated, input):
return -T.sum(T.nnet.binary_crossentropy(clip(generated), input),
axis=1)
def get_args(parser):
parser.add_argument("--step_size", default=0.0001, type=float)
parser.add_argument("--batch_size", default=50, type=int)
parser.add_argument("--n_epochs", default=100, type=int)
parser.add_argument("--latent_dim", default=2, type=int)
parser.add_argument("--save", default=1, type=int)
parser.add_argument("--plot", default=1, type=int)
parser.add_argument("--genx", default=None, type=str)
parser.add_argument("--gent", default=None, type=str)
parser.add_argument("--proportion_labelled", default=0.1, type=float)
return parser.parse_args()
if __name__ == "__main__":
args = get_args(argparse.ArgumentParser())
print("no pretrained files found")
sys.exit(-1)
main(args.step_size, args.batch_size, args.n_epochs, args.save, args.plot,
args.latent_dim, args.proportion_labelled)