-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
324 lines (279 loc) · 11.2 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
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
315
316
317
318
319
320
321
322
323
324
from torchvision import datasets, transforms
from torch.utils.data import Dataset, DataLoader
import numpy as np
from numpy import log, pi
import torch
import matplotlib.pyplot as plt
from torchvision.utils import make_grid
import time
import io
import tensorflow as tf
class MNISTDataset(Dataset):
"""
Noisable MNIST dataset. element-wise uniform noise will be added to each item
From Nice paper. 为了更符合高斯,但实际效果可能有限。
"""
def __init__(self, noise=0.1, train=True):
transform = transforms.Compose([
transforms.ToTensor(), # first, convert image to PyTorch tensor
transforms.Normalize((0.1307,), (0.3081,)) # normalize inputs
])
self.MNIST = datasets.MNIST(root='./data/mnist', train=train, transform=transform, download=True)
self.noise = noise
def __getitem__(self, index):
data, target = self.MNIST[index]
data = data + (torch.rand(data.size()) - 0.5) * self.noise
# Your transformations here (or set it in CIFAR10)
return data, target, index
def __len__(self):
return len(self.MNIST)
class AEMNISTDataset(Dataset):
"""
Noisable AutoEncoded-MNIST dataset.
1. element-wise uniform noise will be added to each image.
2. an additional feature vector generated by a pretrained AutoEncoder will be returned.
"""
def __init__(self, ae_model, N, noise=0.1, train=True, aex_file=None):
transform = transforms.Compose([
transforms.ToTensor(), # first, convert image to PyTorch tensor
transforms.Normalize((0.1307,), (0.3081,)) # normalize inputs
])
self.MNIST = datasets.MNIST(root='./data/mnist', train=train, transform=transform, download=True)
self.MNIST.data = self.MNIST.data[:N]
self.targets = np.asarray(self.MNIST.targets[:N])
self.noise = noise
if aex_file is None:
lst = []
for i in range(len(self.MNIST)):
xi, _ = self.MNIST[i]
xi = xi.view(-1)
if cfg['USE_CUDA']:
xi = xi.cuda()
xi = ae_model(xi)
xi = xi.detach().cpu().numpy().reshape(1, -1)
lst.append(xi)
xs0 = np.concatenate(lst, axis=0)
xs0 = torch.from_numpy(xs0)
# x = self.MNIST.data.type(torch.FloatTensor)
# xs0 = ae_model((x-0.1307)/0.3081).detach().cpu().numpy()
#self.ae_x = xs0
else:
xs0 = torch.load(aex_file)[:N]
_mu = torch.mean(xs0, 0)
_std = torch.std(xs0, 0)
self.ae_x = (xs0-_mu)/_std
def __getitem__(self, index):
data, target = self.MNIST[index]
data = data + (torch.rand(data.size()) - 0.5) * self.noise
# Your transformations here (or set it in CIFAR10)
return data, self.ae_x[index], target, index
def __len__(self):
return len(self.MNIST)
class InfiniteDataLoader(DataLoader):
"""
reload the dataset from start when meets an end.
"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# Initialize an iterator over the dataset.
self.dataset_iterator = super().__iter__()
def __iter__(self):
return self
def __next__(self):
try:
batch = next(self.dataset_iterator)
except StopIteration:
# Dataset exhausted, use a new fresh iterator.
self.dataset_iterator = super().__iter__()
batch = next(self.dataset_iterator)
return batch
def plot2tfimage(figure):
"""Converts the matplotlib plot specified by 'figure' to a PNG image and
returns it. The supplied figure is closed and inaccessible after this call."""
# Save the plot to a PNG in memory.
buf = io.BytesIO()
plt.savefig(buf, format='png')
# Closing the figure prevents it from being displayed directly inside
# the notebook.
plt.close(figure)
buf.seek(0)
# Convert PNG buffer to TF image
image = tf.image.decode_png(buf.getvalue(), channels=4)
# Add the batch dimension
image = tf.expand_dims(image, 0)
return image
def show_clusters(K, ks, dataset, xs0, mu_K, W=28, H=28, k_lst=None):
if k_lst == None:
k_lst = range(K)
for k in k_lst:
print('Cluster %d:' % k)
plt.imshow(np.asarray(mu_K[k]).reshape(W, H))
plt.title('Centroid')
plt.show()
cnt = min(25, len(np.where(ks == k)[0]))
idx = np.random.choice(np.where(ks == k)[0], cnt, replace=False)
imgx = make_grid(dataset.MNIST.data[idx].view(-1, 1, W, H).float(), nrow=5, normalize=True).detach().numpy()
imgz = make_grid(torch.from_numpy(xs0[idx]).view(-1, 1, W, H).float(), nrow=5, normalize=True).detach().numpy()
# imgx = 1.0*(imgx+np.min(imgx))/(np.max(imgx)-np.min(imgx))
# imgz = 1.0*(imgz+np.min(imgz))/(np.max(imgz)-np.min(imgz))
plt.subplot(1, 2, 1)
plt.imshow(np.transpose(imgx, (1, 2, 0)), interpolation='nearest')
plt.title('Origin space')
plt.subplot(1, 2, 2)
plt.imshow(np.transpose(imgz, (1, 2, 0)), interpolation='nearest')
plt.title('Transformed space')
plt.show()
def sample_clusters(nK_total, ks, dataset, xs0, W=28, H=28, fig_size=8, cluster_idx_lst=None):
N_CLASS = 15
N_CLASS_SAMPLE = 9
N_IMAGE_ROW = 3
aeW = 2
aeH = 5
# time consuming operation
if cluster_idx_lst is None:
cluster_idx_lst = list(range(min(nK_total, N_CLASS)))
figure = plt.figure()
plt.subplots(nrows=len(cluster_idx_lst), ncols=3, figsize=(fig_size, fig_size * len(cluster_idx_lst) / 3))
plt.tight_layout()
ll = len(cluster_idx_lst)
for _kid in range(ll):
# 1. show cluster mean vector
k = cluster_idx_lst[_kid]
plt.title('k:%d/%d |C_k|=%d' % (k, nK_total, len(np.where(ks == k)[0])))
# sample images idx from each cluster
cnt = min(N_CLASS_SAMPLE, len(np.where(ks == k)[0]))
idx = np.random.choice(np.where(ks == k)[0], cnt, replace=False)
# 2. plot raw image
imgx = make_grid(dataset.MNIST.data[idx].view(-1, 1, W, H).float(),
nrow=N_IMAGE_ROW, normalize=True
).detach().numpy()
plt.subplot(ll, 3, _kid * 3 + 2)
plt.imshow(np.transpose(imgx, (1, 2, 0)), interpolation='nearest')
plt.title('Origin space')
# 3. plot flow repr of image
image_z = torch.from_numpy(xs0[idx]).view(-1, 1, aeW, aeH).float()
imgz = make_grid(image_z, nrow=N_IMAGE_ROW, normalize=True).detach().numpy()
plt.subplot(ll, 3, _kid * 3 + 3)
plt.imshow(np.transpose(imgz, (1, 2, 0)), interpolation='nearest')
plt.title('Transformed space')
# show image
buf = io.BytesIO()
plt.savefig(buf, format='png')
plt.close(figure)
buf.seek(0)
# Convert PNG buffer to TF image
image = tf.image.decode_png(buf.getvalue(), channels=4)
# Add the batch dimension
image = tf.expand_dims(image, 0)
return image
def plt2img(pllt, figure):
buf = io.BytesIO()
pllt.savefig(buf, format='png')
pllt.close(figure)
buf.seek(0)
# Convert PNG buffer to TF image
image = tf.image.decode_png(buf.getvalue(), channels=4)
# Add the batch dimension
image = tf.expand_dims(image, 0)
return image
def savefig_clusters(n_epoch, nK_total, ks, dataset, xs0, mu_K, path='./res', sz=8, phase=1,
datasetname='mnist', cluster_idx_lst=None, plot=False, model=None, tensor_board=True):
if datasetname == 'mnist':
W, H, aeW, aeH = 28, 28, 2, 5
elif datasetname == 'cifar10':
W, H, aeW, aeH = 32, 32, 6, 8
N_CLASS = 15
N_CLASS_SAMPLE = 9
N_IMAGE_ROW = 3
# time consuming operation
if cluster_idx_lst is None:
cluster_idx_lst = list(range(min(nK_total, N_CLASS)))
figure = plt.figure()
plt.subplots(nrows=len(cluster_idx_lst), ncols=3, figsize=(sz, sz * len(cluster_idx_lst) / 3))
plt.tight_layout()
ll = len(cluster_idx_lst)
for _kid in range(ll):
# 1. show cluster mean vector
k = cluster_idx_lst[_kid]
plt.subplot(ll, 3, _kid * 3 + 1)
plt.imshow(np.asarray(mu_K[k]).reshape(aeW, aeH))
plt.colorbar()
plt.title('k:%d/%d |C_k|=%d' % (k, nK_total, len(np.where(ks == k)[0])))
# sample images idx from each cluster
cnt = min(N_CLASS_SAMPLE, len(np.where(ks == k)[0]))
idx = np.random.choice(np.where(ks == k)[0], cnt, replace=False)
# 2. plot raw image
imgx = make_grid(dataset.data[idx].view(-1, 1, W, H).float(), nrow=N_IMAGE_ROW, normalize=True).detach().numpy()
plt.subplot(ll, 3, _kid * 3 + 2)
plt.imshow(np.transpose(imgx, (1, 2, 0)), interpolation='nearest')
plt.title('Origin space')
# 3. plot flow repr of image
if xs0 is None:
image_z = [model.cpu().f(dataset[_idx][0].view(-1)).view(1, aeW, aeH).float() for _idx in idx]
else:
image_z = torch.from_numpy(xs0[idx]).view(-1, 1, aeW, aeH).float()
imgz = make_grid(image_z, nrow=N_IMAGE_ROW, normalize=True).detach().numpy()
plt.subplot(ll, 3, _kid * 3 + 3)
plt.imshow(np.transpose(imgz, (1, 2, 0)), interpolation='nearest')
plt.title('Transformed space')
# show image
if tensor_board:
buf = io.BytesIO()
plt.savefig(buf, format='png')
plt.close(figure)
buf.seek(0)
# Convert PNG buffer to TF image
image = tf.image.decode_png(buf.getvalue(), channels=4)
# Add the batch dimension
image = tf.expand_dims(image, 0)
return image
elif plot:
plt.show()
plt.close(figure)
elif path is not None:
plt.savefig('%s/%s_phase%d_epoch%d.png' % (path, datasetname, phase, n_epoch))
plt.close(figure)
def transform_z(model, dataset, N, idx=None, xs0=None, _n=5000, normalize=False):
s = time.time()
model.cuda().eval()
if idx is None:
# xs0 = model.f(dataset.ae_x)[0].detach().cpu().numpy()
lst = []
xs0 = dataset
for b in range(N):
_, xi, _, _ = xs0[b]
xi = xi.view(-1)
xi = xi.cuda()
xi, _ = model.f(xi)
xi = xi.detach().cpu().numpy().reshape(1, -1)
lst.append(xi)
xs0 = np.concatenate(lst, axis=0)
print(f'transfrom consumed {time.time()-s:.1f}s')
return xs0
# def transform_ae(model, data, N):
# xs0 = data
# lst = []
# for b in range(N):
# xi, _, _ = xs0[b]
# xi = xi.view(-1)
# if cfg['USE_CUDA']: xi = xi.cuda()
# xi, _ = model(xi)
# xi = xi.detach().cpu().numpy().reshape(1,-1)
# mu = np.mean(xi)
# sig = np.std(xi)
# xi = (xi-mu)/sig
# lst.append(xi)
# xs0 = np.concatenate(lst, axis=0)
# return xs0
def mvnrvs(mu, sig2):
"""multi-variant normal random variables"""
sig = np.sqrt(sig2)
d = len(mu)
res = np.random.randn(d) * sig + mu
return res
def mvnlogpdf(x, mu, sig2):
"""multi-variant normal log probability density fucntion"""
sig = np.sqrt(sig2)
d = len(mu)
res = -0.5 * d * log(2 * pi * sig2) - 0.5 * np.sum((x - mu) ** 2) / sig2
return res