-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
213 lines (198 loc) · 8.45 KB
/
utils.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
import time
import sys
import os
from PIL import Image
import numpy as np
import lasagne as nn
import theano
import theano.tensor as T
import h5py
from fuel.datasets.hdf5 import H5PYDataset
from fuel.schemes import ShuffledScheme, SequentialScheme
from fuel.streams import DataStream
# runs training loop, expects data in DataH5PYStreamer format
# tr_transform and te_transform must return list or tuple, to allow
# for situations where the functions require 2+ inputs
def train_with_hdf5(data, num_epochs, train_fn, test_fn,
tr_transform = lambda x:x,
te_transform = lambda x:x,
verbose=True, train_shuffle=True,
save_best_params_to=None,
last_layer=None,
max_per_epoch=-1):
tr_stream = data.streamer(training=True, shuffled=train_shuffle)
te_stream = data.streamer(training=False, shuffled=False)
from tqdm import tqdm
ret = []
mve_params = None
mve = None
for epoch in range(num_epochs):
start = time.time()
tr_err, tr_batches = 0,0
for imb in tqdm(tr_stream.get_epoch_iterator(), total=data.ntrain/data.batch_size):
if imb[0].shape[0] != data.batch_size:
continue
imb = tr_transform(imb)
if not isinstance(imb, tuple):
imb = (imb,)
tr_err += train_fn(*imb)
tr_batches += 1
if max_per_epoch > 0 and tr_batches > max_per_epoch:
break
val_err, val_batches = 0,0
for imb in tqdm(te_stream.get_epoch_iterator(), total=data.ntest/data.batch_size):
if imb[0].shape[0] != data.batch_size:
continue
imb = te_transform(imb)
if not isinstance(imb, tuple):
imb = (imb,)
val_err += test_fn(*imb)
val_batches += 1
if max_per_epoch > 0 and val_batches > max_per_epoch:
break
val_err /= val_batches
tr_err /= tr_batches
if save_best_params_to is not None:
if mve is None or val_err < mve:
mve = val_err
mve_params = nn.layers.get_all_param_values(last_layer)
if verbose:
print('ep {}/{} - tl {:.5f} - vl {:.5f} - t {:.3f}s'.format(
epoch, num_epochs, tr_err, val_err, time.time()-start))
ret.append((tr_err, val_err))
if save_best_params_to is not None:
save_params(mve_params, save_best_params_to)
return ret
# goes from raw image array (usually uint8) to floatX, square=True crops to
# size of the short edge, center=True crops at center, otherwise crop is
# random
def raw_to_floatX(imb, pixel_shift=0.5, square=True, center=False, rng=None):
rng = rng if rng else np.random
w,h = imb.shape[2], imb.shape[3] # image size
x, y = 0,0 # offsets
if square:
if w > h:
if center:
x = (w-h)/2
else:
x = rng.randint(w-h)
w=h
elif h > w:
if center:
y = (h-w)/2
else:
y = rng.randint(h-w)
h=w
return nn.utils.floatX(imb)[:,:,x:x+w,y:y+h]/ 255. - pixel_shift
# creates and hdf5 file from a dataset given a split in the form {'train':(0,n)}, etc
# appears to save in unpredictable order, so order must be verified after creation
def save_hd5py(dataset_dict, destfile, indices_dict):
f = h5py.File(destfile, mode='w')
for name, dataset in dataset_dict.iteritems():
dat = f.create_dataset(name, dataset.shape, dtype=str(dataset.dtype))
dat[...] = dataset
split_dict = dict((k, dict((name, v) for name in dataset_dict.iterkeys()))
for k,v in indices_dict.iteritems())
f.attrs['split'] = H5PYDataset.create_split_array(split_dict)
f.flush()
f.close()
# for organizing an hdf5 file for streaming
class DataH5PyStreamer:
def __init__(self, h5filename, ntrain=None, nval=None, ntest=None, batch_size=128):
self.tr_data = H5PYDataset(h5filename, which_sets=('train',))
self.te_data = H5PYDataset(h5filename, which_sets=('test',))
self.ntrain = ntrain if ntrain is not None else self.tr_data.num_examples
self.ntest = ntest if ntest is not None else self.te_data.num_examples
self.batch_size = batch_size
def dataset(self, training=True):
return self.tr_data if training else self.te_data
def streamer(self, training=True, shuffled=False):
n = self.ntrain if training else self.ntest
sch = ShuffledScheme(examples=n, batch_size=self.batch_size) if shuffled else \
SequentialScheme(examples=n, batch_size=self.batch_size)
return DataStream(self.tr_data if training else self.te_data, \
iteration_scheme = sch)
# helper function for building vae's
def log_likelihood(tgt, mu, ls):
return T.sum(-(np.float32(0.5 * np.log(2 * np.pi)) + ls)
- 0.5 * T.sqr(tgt - mu) / T.exp(2 * ls))
# from the array used for testing, to the kind used in Image.fromarray(..)
def get_picture_array(X, index, shift=0.5):
ch, w, h = X.shape[1], X.shape[2], X.shape[3]
ret = ((X[index]+shift)*255.).reshape(ch,w,h).transpose(2,1,0).clip(0,255).astype(np.uint8)
if ch == 1:
ret=ret.reshape(h,w)
return ret
# returns an Image with X on top, Xpr on bottom, index as requeseted or random if -1
def get_image_pair(X, Xpr,index=-1,shift=0.5):
mode = 'RGB' if X.shape[1] == 3 else 'L'
index = np.random.randint(X.shape[0]) if index == -1 else index
original_image = Image.fromarray(get_picture_array(X, index,shift=shift),mode=mode)
new_size = (original_image.size[0], original_image.size[1]*2)
new_im = Image.new(mode, new_size)
new_im.paste(original_image, (0,0))
rec_image = Image.fromarray(get_picture_array(Xpr, index,shift=shift),mode=mode)
new_im.paste(rec_image, (0,original_image.size[1]))
return new_im
# gets array (in format used for storage) from an Image
def arr_from_img_storage(im):
w,h=im.size
arr=np.asarray(im.getdata(), dtype=np.uint8)
c = np.product(arr.size) / (w*h)
return arr.reshape(h,w,c).transpose(2,1,0)
# gets array (in format used for testing) from an Image
def arr_from_img(im,shift=0.5):
w,h=im.size
arr=np.asarray(im.getdata(), dtype=theano.config.floatX)
c = np.product(arr.size) / (w*h)
return arr.reshape((h,w,c)).transpose(2,1,0) / 255. - shift
# loads params in npz (if filename is a .npz) or pickle if not
def load_params(model, fn):
if 'npz' in fn:
with np.load(fn) as f:
param_values = [f['arr_%d' % i] for i in range(len(f.files))]
nn.layers.set_all_param_values(model, param_values)
else:
with open(fn, 'r') as re:
import pickle
nn.layers.set_all_param_values(model, pickle.load(re))
# saves params in npz (if filename is a .npz) or pickle if not
def save_params(model, fn):
if 'npz' in fn:
if isinstance(model, list):
param_vals = model
else:
param_vals = nn.layers.get_all_param_values(model)
np.savez(fn, *param_vals)
else:
with open(fn, 'w') as wr:
import pickle
pickle.dump(param_vals, wr)
# reset shared variable values of accumulators to recover from NaN
def reset_accs(updates, params):
for key in updates:
if not key in params:
v = key.get_value(borrow=True)
key.set_value(np.zeros(v.shape,dtype=v.dtype))
# build loss as in (Kingma, Welling 2014) Autoencoding Variational Bayes
def build_vae_loss(input_var, l_z_mu, l_z_ls, l_x_mu_list, l_x_ls_list, l_x_list, l_x,
deterministic, binary, L):
layer_outputs = nn.layers.get_output([l_z_mu, l_z_ls] + l_x_mu_list + l_x_ls_list
+ l_x_list + [l_x], deterministic=deterministic)
z_mu = layer_outputs[0]
z_ls = layer_outputs[1]
x_mu = [] if binary else layer_outputs[2:2+L]
x_ls = [] if binary else layer_outputs[2+L:2+2*L]
x_list = layer_outputs[2:2+L] if binary else layer_outputs[2+2*L:2+3*L]
x = layer_outputs[-1]
kl_div = 0.5 * T.sum(1 + 2*z_ls - T.sqr(z_mu) - T.exp(2 * z_ls))
if binary:
logpxz = sum(nn.objectives.binary_crossentropy(x, input_var).sum()
for x in x_list) * (-1./L)
prediction = x_list[0] if deterministic else x
else:
logpxz = sum(log_likelihood(input_var.flatten(2), mu, ls)
for mu, ls in zip(x_mu, x_ls))/L
prediction = x_mu[0] if deterministic else T.sum(x_mu, axis=0)/L
loss = -1 * (logpxz + kl_div)
return loss, prediction