-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathdataset.py
232 lines (195 loc) · 9.16 KB
/
dataset.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
from six.moves import xrange
import numpy as np
import tensorflow as tf
import random
import os
from PIL import Image
from tqdm import tqdm
class Omniglot():
def __init__(self,
data_dir='datasets/omniglot',
image_size=(28,28),
seed=0):
def _full_path(dir):
return [ os.path.join(dir,f) for f in os.listdir(dir) ]
alphabets = [ (os.path.basename(f),f) for f in _full_path(os.path.join(data_dir,'images_background')) if os.path.isdir(f) ] + \
[ (os.path.basename(f),f) for f in _full_path(os.path.join(data_dir,'images_evaluation')) if os.path.isdir(f) ]
chars = [
(alphabet+'_' +os.path.basename(f),f)
for alphabet,path in alphabets
for f in _full_path(path)
if os.path.isdir(f)
]
old_state = random.getstate()
random.seed(seed)
random.shuffle(chars)
random.setstate(old_state)
chars = [ (name,[f for f in _full_path(path) if os.path.splitext(f)[1] == '.png']) for name,path in chars]
assert( len(alphabets) == 50)
assert( len(chars) == 1623)
for name,filenames in chars :
assert(len(filenames) == 20)
self.chars = chars
self.train_chars = chars[:1200]
self.valid_chars = chars[1200:]
def build_queue(self,task_num,n_way,k_shots,train=True,num_threads=1):
chars = self.train_chars if train else self.valid_chars
with tf.device('/cpu'):
# Load all images to memory for faster reading.
chars_cache_idx = []
ims = np.zeros((len(chars)*20,28,28,1),np.float32)
cnt = 0
for (name,files) in tqdm(chars):
idxes = []
for fnames in files :
resized = Image.open(fnames).resize((28,28),resample=Image.LANCZOS)
ims[cnt] = np.expand_dims(np.asarray(resized,np.float32),3)
idxes.append(cnt)
cnt += 1
chars_cache_idx.append((name,idxes))
ims = tf.convert_to_tensor(ims)
def _get_single_task(n_way,k_shots):
idxes = np.random.choice( len(chars), n_way, replace=False)
rots = np.random.choice( 4, n_way, replace=True )
names = [ chars_cache_idx[idx][0] for idx in idxes]
files = [ np.random.choice(chars_cache_idx[idx][1],k_shots*2,replace=False)
for idx in idxes ]
files = np.stack(files,axis=0)
rots = np.tile( rots.reshape(n_way,1),[1,k_shots] )
labels = np.tile( np.arange(0,n_way).reshape(n_way,1),[1,k_shots*2] )
x,x_prime = np.split(files, 2, axis=1)
y,y_prime = np.split(labels,2, axis=1)
rots = rots.reshape(-1)
x = x.reshape(-1)
x_prime = x_prime.reshape(-1)
y = y.reshape(-1)
y_prime = y_prime.reshape(-1)
return np.array(names,np.string_),rots,x,y,x_prime,y_prime
def _read_single_im(elems):
f_idx,rot = elems #file index, rotations
_t = ims[f_idx]
def _raise():
assert_op = tf.Assert(False,['Undefined Rotation'])
with tf.control_dependencies([assert_op]):
return _t
_t = tf.case({
tf.equal(rot,0): lambda : _t,
tf.equal(rot,1): lambda : tf.image.rot90(_t,1),
tf.equal(rot,2): lambda : tf.image.rot90(_t,2),
tf.equal(rot,3): lambda : tf.image.rot90(_t,3)},
default= _raise)
#_t = tf.cast(_t,tf.float32) / 255.0 # Omniglot is already [0-1] since images are BW
#_t = tf.subtract(_t, 0.5)
#_t = tf.multiply(_t, 2.0)
return tf.transpose(_t,[2,0,1])
task,rots,x,y,x_prime,y_prime = tf.py_func(_get_single_task,
[n_way,k_shots],
[tf.string,tf.int64,tf.int64,tf.int64,tf.int64,tf.int64],
stateful=True)
task = tf.reshape(task,(n_way,))
x = tf.reshape(x,(n_way*k_shots,))
y = tf.reshape(y,(n_way*k_shots,))
x_prime = tf.reshape(x_prime,(n_way*k_shots,))
y_prime = tf.reshape(y_prime,(n_way*k_shots,))
x = tf.map_fn(_read_single_im,[x,rots],dtype=tf.float32,back_prop=False,parallel_iterations=1)
x_prime = tf.map_fn(_read_single_im,[x_prime,rots],dtype=tf.float32,back_prop=False,parallel_iterations=1)
# Build task batch
tasks, x, y, x_prime, y_prime = tf.train.batch(
[task,x,y,x_prime,y_prime],
batch_size=task_num,
num_threads=num_threads,
capacity=10*task_num)
return tasks,x,y,x_prime,y_prime
class Sinusoid():
def __init__(self,
amp_range=(0.1,5.0),
phase_range=(0,np.pi),
input_range=(-5.0,5.0) ):
self.amp_range = amp_range
self.phase_range = phase_range
self.input_range = input_range
def _generate_random_task(self):
amp = np.random.uniform(*self.amp_range)
phase = np.random.uniform(*self.phase_range)
return amp, phase
def _generate_random_task_batch(self,amp,phase,batch_size):
x = np.random.uniform(*self.input_range, size=batch_size)
y = amp * np.sin(x-phase)
return x, y
def _generate_valid_task_batch(self,amp,phase,num_pts):
x = np.linspace(*self.input_range,num=num_pts)
y = amp * np.sin(x-phase)
return x, y
def build_queue(self,task_num,batch_size,train=True):
# Actually, this does not build queue since data can generated on the fly.
with tf.device('/cpu'):
def _gen(task_num,batch_size,train):
tasks = []
xs,ys = [], []
xs_prime, ys_prime = [], []
for num in xrange(task_num):
amp,phase = self._generate_random_task()
x,y = self._generate_random_task_batch(amp,phase,batch_size)
if(train):
x_prime,y_prime = self._generate_random_task_batch(amp,phase,batch_size)
else:
x_prime,y_prime = self._generate_valid_task_batch(amp,phase,batch_size*20)
tasks.append((amp,phase))
xs.append(x)
ys.append(y)
xs_prime.append(x_prime)
ys_prime.append(y_prime)
return np.array(tasks,np.float32), \
np.array(xs,np.float32), np.array(ys,np.float32), \
np.array(xs_prime,np.float32), np.array(ys_prime,np.float32)
tasks,x,y,x_prime,y_prime = tf.py_func(_gen,
[task_num,batch_size,train],
[tf.float32,tf.float32,tf.float32,tf.float32,tf.float32],
stateful=True)
tasks = tf.reshape(tasks,[task_num,2])
x = tf.reshape(x,[task_num,batch_size,1])
y = tf.reshape(y,[task_num,batch_size,1])
if( train ):
x_prime = tf.reshape(x_prime,[task_num,batch_size,1])
y_prime = tf.reshape(y_prime,[task_num,batch_size,1])
else :
x_prime = tf.reshape(x_prime,[task_num,batch_size*20,1])
y_prime = tf.reshape(y_prime,[task_num,batch_size*20,1])
return tasks,x,y,x_prime,y_prime
if __name__ == "__main__":
from tqdm import tqdm
omni = Omniglot()
gen_op = omni.build_queue(10,5,2)
x_op = tf.reshape(gen_op[1], [10*5*2,1,28,28])
x_prime_op = tf.reshape(gen_op[3], [10*5*2,1,28,28])
tf.summary.image('x',tf.transpose(x_op[:10],(0,2,3,1)),max_outputs=10)
tf.summary.image('x_prime',tf.transpose(x_prime_op[:10],(0,2,3,1)),max_outputs=10)
summary_op = tf.summary.merge_all()
#sin = Sinusoid()
#gen_op = sin.build_queue(10,10)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.graph.finalize()
sess.run(init_op)
summary_writer = tf.summary.FileWriter('./log_temp',sess.graph)
try:
# Start Queueing
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord,sess=sess)
for step in tqdm(xrange(100)):
#(tasks,x,y,x_prime,y_prime) = sess.run(gen_op)
(tasks,x,y,x_prime,y_prime), summary_str = sess.run([gen_op,summary_op])
summary_writer.add_summary(summary_str,step)
except Exception as e:
coord.request_stop(e)
finally :
coord.request_stop()
coord.join(threads)
print(tasks)
#print(x)
print(y)
#print(x_prime)
print(y_prime)