-
Notifications
You must be signed in to change notification settings - Fork 14
/
generate.py
254 lines (225 loc) · 12.7 KB
/
generate.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
'''
This code
(1)generates data in various ways given a trained SGAN
(2)inferences latent code for test set images
(3)performs style transfer
Note: Due to the effect of Batch Normalization, it is better to generate batch_size_g (see train file, 200 for cifiar10) samples distributed equally across class in each batch.
'''
import gzip, os, cPickle, time, math, argparse, shutil, sys
import numpy as np
import theano, lasagne
import theano.tensor as T
import lasagne.layers as ll
import lasagne.nonlinearities as ln
from lasagne.layers import dnn
import nn
from lasagne.init import Normal
from theano.sandbox.rng_mrg import MRG_RandomStreams
from layers.merge import ConvConcatLayer, MLPConcatLayer
from layers.deconv import Deconv2DLayer
from components.shortcuts import convlayer, mlplayer
from components.objectives import categorical_crossentropy_ssl_separated, maximum_mean_discripancy, categorical_crossentropy, feature_matching
from utils.create_ssl_data import create_ssl_data, create_ssl_data_subset
from utils.others import get_nonlin_list, get_pad_list, bernoullisample, printarray_2D, array2file_2D
import utils.paramgraphics as paramgraphics
from utils.checkpoints import load_weights
from visualize import *
# global
parser = argparse.ArgumentParser()
parser.add_argument("-oldmodel", type=str, default=argparse.SUPPRESS)
parser.add_argument("-dataset", type=str, default='cifar10')
args = parser.parse_args()
filename_script=os.path.basename(os.path.realpath(__file__))
outfolder='imags/cifar10_sgan/'
if not os.path.exists(outfolder):
os.makedirs(outfolder)
# seeds
seed=1100
rng=np.random.RandomState(seed)
theano_rng=MRG_RandomStreams(rng.randint(2 ** 15))
lasagne.random.set_rng(np.random.RandomState(rng.randint(2 ** 15)))
# G
n_z=100
batch_size_g=100
num_x=50000
# data dependent
if args.dataset == 'svhn' or args.dataset == 'cifar10':
gen_final_non=ln.tanh
num_classes=10
dim_input=(32,32)
in_channels=3
colorImg=True
generation_scale=True
if args.dataset == 'svhn':
def rescale(mat):
return np.transpose(np.cast[theano.config.floatX]((-127.5 + mat)/127.5),(3,2,0,1))
import svhn_data
eval_x, eval_y = svhn_data.load('./svhn/','test')
eval_y = np.int32(eval_y)
eval_x = rescale(eval_x)
else:
def rescale(mat):
return np.cast[theano.config.floatX](mat)
import cifar10_data
eval_x, eval_y = cifar10_data.load('./cifar10/','test')
eval_y = np.int32(eval_y)
eval_x = rescale(eval_x)
elif args.dataset == 'mnist':
gen_final_non=ln.sigmoid
num_classes=10
dim_input=(28,28)
in_channels=1
colorImg=False
generation_scale=False
'''
models
'''
# symbols
sym_y_g = T.ivector()
sym_z_input = T.matrix()
sym_z_rand = theano_rng.uniform(size=(batch_size_g, n_z))
sym_z_shared = T.tile(theano_rng.uniform((batch_size_g/num_classes, n_z)), (num_classes, 1))
sym_x_u_i = T.tensor4()
# generator y2x: p_g(x, y) = p(y) p_g(x | y) where x = G(z, y), z follows p_g(z)
gen_in_z = ll.InputLayer(shape=(None, n_z))
gen_in_y = ll.InputLayer(shape=(None,))
gen_layers = [gen_in_z]
if args.dataset == 'svhn' or args.dataset == 'cifar10':
gen_layers.append(MLPConcatLayer([gen_layers[-1], gen_in_y], num_classes, name='gen-00'))
gen_layers.append(nn.batch_norm(ll.DenseLayer(gen_layers[-1], num_units=4*4*512, W=Normal(0.05), nonlinearity=nn.relu, name='gen-01'), g=None, name='gen-02'))
gen_layers.append(ll.ReshapeLayer(gen_layers[-1], (-1,512,4,4), name='gen-03'))
gen_layers.append(ConvConcatLayer([gen_layers[-1], gen_in_y], num_classes, name='gen-10'))
gen_layers.append(nn.batch_norm(nn.Deconv2DLayer(gen_layers[-1], (None,256,8,8), (5,5), W=Normal(0.05), nonlinearity=nn.relu, name='gen-11'), g=None, name='gen-12')) # 4 -> 8
gen_layers.append(ConvConcatLayer([gen_layers[-1], gen_in_y], num_classes, name='gen-20'))
gen_layers.append(nn.batch_norm(nn.Deconv2DLayer(gen_layers[-1], (None,128,16,16), (5,5), W=Normal(0.05), nonlinearity=nn.relu, name='gen-21'), g=None, name='gen-22')) # 8 -> 16
gen_layers.append(ConvConcatLayer([gen_layers[-1], gen_in_y], num_classes, name='gen-30'))
gen_layers.append(nn.weight_norm(nn.Deconv2DLayer(gen_layers[-1], (None,3,32,32), (5,5), W=Normal(0.05), nonlinearity=gen_final_non, name='gen-31'), train_g=True, init_stdv=0.1, name='gen-32')) # 16 -> 32
elif args.dataset == 'mnist':
gen_layers.append(MLPConcatLayer([gen_layers[-1], gen_in_y], num_classes, name='gen-1'))
gen_layers.append(ll.batch_norm(ll.DenseLayer(gen_layers[-1], num_units=500, nonlinearity=ln.softplus, name='gen-2'), name='gen-3'))
gen_layers.append(MLPConcatLayer([gen_layers[-1], gen_in_y], num_classes, name='gen-4'))
gen_layers.append(ll.batch_norm(ll.DenseLayer(gen_layers[-1], num_units=500, nonlinearity=ln.softplus, name='gen-5'), name='gen-6'))
gen_layers.append(MLPConcatLayer([gen_layers[-1], gen_in_y], num_classes, name='gen-7'))
gen_layers.append(nn.l2normalize(ll.DenseLayer(gen_layers[-1], num_units=28**2, nonlinearity=gen_final_non, name='gen-8')))
# inference module
inf_in_x = ll.InputLayer(shape=(None, in_channels) + dim_input)
inf_layers = [inf_in_x]
inf_layers.append(ll.batch_norm(dnn.Conv2DDNNLayer(inf_layers[-1], 64, (4,4), stride=2, pad=1, W=Normal(0.05), nonlinearity=nn.lrelu, name='inf-02'), name='inf-03'))
inf_layers.append(ll.batch_norm(dnn.Conv2DDNNLayer(inf_layers[-1], 128, (4,4), stride=2, pad=1, W=Normal(0.05), nonlinearity=nn.lrelu, name='inf-11'), name='inf-12'))
inf_layers.append(ll.batch_norm(dnn.Conv2DDNNLayer(inf_layers[-1], 256, (4,4), stride=2, pad=1, W=Normal(0.05), nonlinearity=nn.lrelu, name='inf-21'), name='inf-22'))
inf_layers.append(ll.batch_norm(dnn.Conv2DDNNLayer(inf_layers[-1], 512, (4,4), stride=2, pad=1, W=Normal(0.05), nonlinearity=nn.lrelu, name='inf-31'), name='inf-32'))
inf_layers.append(ll.DenseLayer(inf_layers[-1], num_units=n_z, W=Normal(0.05), nonlinearity=None, name='inf-4'))
# outputs
gen_out_x = ll.get_output(gen_layers[-1], {gen_in_y:sym_y_g, gen_in_z:sym_z_rand}, deterministic=False)
gen_out_x_shared = ll.get_output(gen_layers[-1], {gen_in_y:sym_y_g, gen_in_z:sym_z_shared}, deterministic=False)
gen_out_x_interpolation = ll.get_output(gen_layers[-1], {gen_in_y:sym_y_g, gen_in_z:sym_z_input}, deterministic=False)
generate = theano.function(inputs=[sym_y_g], outputs=gen_out_x)
generate_shared = theano.function(inputs=[sym_y_g], outputs=gen_out_x_shared)
generate_interpolation = theano.function(inputs=[sym_y_g, sym_z_input], outputs=gen_out_x_interpolation)
inf_z = ll.get_output(inf_layers[-1], {inf_in_x:sym_x_u_i}, deterministic=False)
inf = theano.function(inputs=[sym_x_u_i], outputs=inf_z)
'''
Load pretrained model
'''
load_weights(args.oldmodel, gen_layers+inf_layers)
# style transfer
# batch_size=100
# features = np.load("trans.npy").astype(np.float32)#np.zeros([100, n_z])#
# for ppp in range(10):
# images = np.zeros([10*32, 11*32 + 3, 3])
# eval_x = eval_x[rng.permutation(eval_x.shape[0])]
# for t in range(10):
# train_xu = eval_x[t*batch_size:(t+1)*batch_size]
# images[t*32:(t+1)*32, 0:32, :] = np.transpose(train_xu[0], [1, 2, 0])
# z_ = inf(train_xu)
# features[t*10:(t+1)*10] = z_[:10]
# #z_ = features[t*10:(t+1)*10]
# print(sum(z_[0]*z_[0]))
# sample_y = np.int32(np.repeat(np.arange(10), batch_size/10))
# images_ = generate_interpolation(sample_y, np.tile(z_[:10], [10, 1])) #sess.run(x_gen, feed_dict={y_in:sample_y, z_in:np.tile(z_[:10], [10, 1])})
# for j in range(10):
# images[t*32:(t+1)*32, 35+j*32:35+(j+1)*32, :] = np.transpose(images_[j*10], [1, 2, 0])
# scipy.misc.imsave(os.path.join(outfolder, 'tran-'+args.dataset+"-"+str(ppp)+'.png'), images)
#np.save("trans", features)
# generate images in various ways
images = [[] for i in range(10)]
for i in xrange(10):
sample_y = np.int32(np.repeat(np.arange(num_classes), batch_size_g/num_classes))
# inds = np.random.permutation(batch_size_g)
# sample_y = sample_y[inds]
x_gen_batch = generate(sample_y)
for j in range(10):
images[j].append(x_gen_batch[j*10:(j+1)*10])
#x_gen_batch = x_gen_batch.reshape((batch_size_g,-1))
#scipy.misc.imsave(os.path.join(outfolder, 'class-'+str(i)+'.png'), concat_multiple_images(np.transpose((x_gen_batch[:100]+1)/2, [0,2,3,1])))
for i in range(10):
tmp = np.concatenate(images[i], 0)
scipy.misc.imsave(os.path.join(outfolder, 'class-'+str(i)+'.png'), concat_multiple_images(np.transpose((tmp+1)/2, [0,2,3,1])))
# exit(0)
#interpolation on latent space (z) class conditionally
for i in xrange(10):
sample_y = np.int32(np.repeat(np.arange(num_classes), batch_size_g/num_classes))
orignial_z = np.repeat(rng.uniform(size=(num_classes,n_z)), batch_size_g/num_classes, axis=0)
target_z = np.repeat(rng.uniform(size=(num_classes,n_z)), batch_size_g/num_classes, axis=0)
alpha = np.tile(np.arange(batch_size_g/num_classes) * 1.0 / (batch_size_g/num_classes-1), num_classes)
alpha = alpha.reshape(-1,1)
z = np.float32((1-alpha)*orignial_z+alpha*target_z)
x_gen_batch = generate_interpolation(sample_y, z)
#x_gen_batch = x_gen_batch.reshape((batch_size_g,-1))
scipy.misc.imsave(os.path.join(outfolder, 'interpolation-'+str(i)+'.png'), concat_multiple_images(np.transpose((x_gen_batch[:100]+1)/2, [0,2,3,1])))
#image = paramgraphics.mat_to_img(x_gen_batch.T, dim_input, colorImg=colorImg, tile_shape=(num_classes, 2*num_classes), scale=generation_scale, save_path=os.path.join(outfolder, 'interpolation-'+str(i)+'.png'))
# class conditionally generation with shared z and fixed y
for i in xrange(10):
sample_y = np.int32(np.repeat(np.arange(num_classes), batch_size_g/num_classes))
x_gen_batch = generate_shared(sample_y)
#x_gen_batch = x_gen_batch.reshape((batch_size_g,-1))
scipy.misc.imsave(os.path.join(outfolder, 'shared-'+str(i)+'.png'), concat_multiple_images(np.transpose((x_gen_batch[:100]+1)/2, [0,2,3,1])))
#image = paramgraphics.mat_to_img(x_gen_batch.T, dim_input, colorImg=colorImg, tile_shape=(num_classes, 2*num_classes), scale=generation_scale, save_path=os.path.join(outfolder, 'shared-'+str(i)+'.png'))
# generation with randomly sampled z and y
for i in xrange(10):
sample_y = np.int32(np.repeat(np.arange(num_classes), batch_size_g/num_classes))
# inds = np.random.permutation(batch_size_g)
# sample_y = sample_y[inds]
x_gen_batch = generate(sample_y)
#x_gen_batch = x_gen_batch.reshape((batch_size_g,-1))
scipy.misc.imsave(os.path.join(outfolder, 'random-'+str(i)+'.png'), concat_multiple_images(np.transpose((x_gen_batch[:100]+1)/2, [0,2,3,1])))
#image = paramgraphics.mat_to_img(x_gen_batch.T, dim_input, colorImg=colorImg, tile_shape=(num_classes, 2*num_classes), scale=generation_scale, save_path=os.path.join(outfolder, 'random-'+str(i)+'.png'))
#randomly sample real images
for i in range(10):
bx = eval_x[rng.permutation(eval_x.shape[0])][:100]
scipy.misc.imsave(os.path.join(outfolder, 'sample-'+str(i)+'.png'), concat_multiple_images(np.transpose((bx+1)/2, [0,2,3,1])))
#randomly sample real images condition on class
for i in range(10):
inds = np.random.permutation(eval_x.shape[0])
eval_y = eval_y[inds]
eval_x = eval_x[inds]
bx = []
for j in range(10):
bx.append(eval_x[eval_y==j][:10])
#bx = eval_x[rng.permutation(eval_x.shape[0])][:100]
bx = np.concatenate(bx, 0)
#bx = np.reshape(np.transpose(np.reshape(bx, [10, 10, 3, 32, 32]), [1,0,2,3,4]), [100, 3, 32, 32])
scipy.misc.imsave(os.path.join(outfolder, 'sample-cond-'+str(i)+'.png'), concat_multiple_images(np.transpose((bx+1)/2, [0,2,3,1])))
# randomly generate 50000 images for inception score computation
gens = 50000
images = np.zeros([gens, 3, 32, 32])
labels = np.zeros([gens])
for t in range(int(gens/batch_size_g)):
#sample_y = np.argmax(np.random.multinomial(1, [1/10.]*10, size=batch_size_g), 1)#np.int32(np.repeat(np.arange(num_classes), batch_size_g/num_classes))
sample_y = np.int32(np.repeat(np.arange(num_classes), batch_size_g/num_classes))
inds = np.random.permutation(batch_size_g)
sample_y = sample_y[inds]
image_ = generate(sample_y)
images[t*batch_size_g:(t+1)*batch_size_g] = image_#np.reshape(image_, [batch_size_g, 28, 28, 1])
labels[t*batch_size_g:(t+1)*batch_size_g] = sample_y
print(labels[100:200])
np.savez('generations-sgan-inc-50000.npz', x=images[:gens], y=labels[:gens])
# inference latent code for test images
features = np.zeros([eval_x.shape[0], 100])
print(eval_y.shape)
labels = eval_y
for i in range(int(math.ceil(eval_x.shape[0]/100.0))):
inf_z_ = inf(eval_x[100*i:min(100*(i+1), eval_x.shape[0])])
features[100*i:min(100*(i+1), eval_x.shape[0])] = inf_z_
print(features.shape)
np.savez('features-sgan-cifar.npz', x=features, y=labels)