-
Notifications
You must be signed in to change notification settings - Fork 19
/
create_dataset.py
206 lines (168 loc) · 8.72 KB
/
create_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
import argparse
import numpy as np
import tensorflow as tf
from PIL import Image
from tqdm import tqdm
from sklearn.cluster import KMeans
from sklearn.neighbors import NearestNeighbors
import pickle
import h5py
import os
from joblib import dump, load
default_batch_size = 1000 #number of consecutively processed latents
target_directory = 'data'
if __name__ == '__main__':
parser = argparse.ArgumentParser(description = 'create descriptors and cluster tool')
parser.add_argument('path', metavar='F', type=str, help='file path to network pickle')
parser.add_argument('num_latents', metavar='N', type=int, help='number of latents in database', default=200000)
parser.add_argument('t_size', metavar='T', type=int, help='size of padded output descriptors', default=20)
parser.add_argument('num_clusters', metavar='C', type=int, help='number of clusters', default=12)
parser.add_argument('network_name', metavar='NN', type=str, help='network name')
args = parser.parse_args()
########################################################################################################################
#PARSE ARGUMENTS
########################################################################################################################
filename = args.path
num_latents = args.num_latents
t_size = args.t_size
num_clusters = args.num_clusters
network_name = args.network_name
directory = os.path.join(target_directory, network_name)
########################################################################################################################
#CREATE DIRECTORY
########################################################################################################################
#create directory, if it doesn't exist
if not os.path.exists(directory):
os.makedirs(directory)
# LAYER 2 3 4 5 6 7 8 9
depths = [512, 512, 512, 512, 512, 256, 128, 64, 32] # depth of intermediate network output
minibatch_size = 8
channels = 3
print('Initializing TensorFlow...')
tf.InteractiveSession()
random_state = np.random.RandomState()
########################################################################################################################
# LOAD NETWORK
########################################################################################################################
try:
with open(filename, 'rb') as file:
G, D, Gs = pickle.load(file)
except TypeError:
print('You must specify the path to a valid trained ProGAN network.')
#saving only generator network for later use
with open(os.path.join(directory, '{}_network.pkl'.format(network_name)), 'wb') as file:
pickle.dump(Gs, file, protocol=pickle.HIGHEST_PROTOCOL)
########################################################################################################################
# WRITE INFO FILE
########################################################################################################################
f = open(os.path.join(directory, 'data.txt'), "w+")
f.write("num_latents=%d\r\n" % num_latents)
f.write("descriptor size=%d\r\n" % t_size)
f.write("num clusters=%d\r\n" % num_clusters)
f.close()
########################################################################################################################
# GENERATE LATENTS AND REPRESENTATIONS
########################################################################################################################
strdesc = ''
strdesc += str(t_size)
hdf5_path = os.path.join(directory, '{}_descriptors.hdf5'.format(network_name))
create_descriptors = not os.path.exists(hdf5_path)
hdf5_file = h5py.File(hdf5_path)
total_desc_size = t_size * t_size * channels
print('total descriptor length: ', total_desc_size)
if create_descriptors:
descriptors = np.zeros((default_batch_size, total_desc_size), dtype=np.uint8)
chunk_descriptor_shape = (default_batch_size, total_desc_size)
chunk_latents_shape = (default_batch_size, 512)
chunk_descriptors = np.empty(chunk_descriptor_shape)
chunk_latents = np.empty(chunk_latents_shape)
descriptors_dataset = hdf5_file.create_dataset("descriptors", (0, total_desc_size), maxshape=(None, total_desc_size), dtype='uint8', chunks=chunk_descriptor_shape)
latents_dataset = hdf5_file.create_dataset("latents", (0, 512), maxshape=(None, 512), dtype='float32', chunks=chunk_latents_shape)
chunk_count = 0
for run in tqdm(np.arange(num_latents // default_batch_size)):
latents = np.random.randn(default_batch_size, *Gs.input_shape[1:]).astype(np.float32)
labels = np.zeros([latents.shape[0], 0], np.float32)
padded_images = Gs.run(latents, labels, minibatch_size=minibatch_size, num_gpus=1, out_mul=127.5, out_add=127.5, out_dtype=np.uint8)
batch_index = 0
# generate descriptors
for image in padded_images:
image = np.squeeze(image)
if image.shape[0] == 3:
image = np.rollaxis(image, 0, 3)
im = Image.fromarray(image)
im.thumbnail((t_size, t_size))
descriptors[batch_index, :] = np.ravel(im)
batch_index += 1
descriptors_dataset.resize((chunk_count + default_batch_size, total_desc_size))
descriptors_dataset[chunk_count:] = descriptors
latents_dataset.resize((chunk_count + default_batch_size, 512))
latents_dataset[chunk_count:] = latents
chunk_count += default_batch_size
########################################################################################################################
# GENERATE CLUSTERS
########################################################################################################################
hdf5_file = h5py.File(hdf5_path)
latents = hdf5_file["latents"].value
padded_descriptors = hdf5_file["descriptors"].value
print('descriptors shape: {}'.format(padded_descriptors.shape))
kmeans_path = os.path.join(directory, '{}_kmeans.joblib'.format(network_name))
if os.path.isfile(kmeans_path):
kmeans = load(kmeans_path)
print('loaded KMeans!')
else:
print('calculating {}-means for {} descriptors'.format(num_clusters, len(padded_descriptors)))
kmeans = KMeans(n_clusters=num_clusters, random_state=0).fit(padded_descriptors)
dump(kmeans, kmeans_path)
if create_descriptors:
hdf5_file.create_dataset("clusters", data=kmeans.labels_)
print('dumped KMeans!')
hdf5_file.close()
def get_averaged_cluster_centers(clustering, num_clusters):
"""Generates plot to view the cluster centroids."""
average_images = np.zeros((num_clusters, t_size, t_size, 3), dtype=np.uint8)
for c in range(num_clusters):
pixData = clustering.cluster_centers_[c, :]
pixImg = np.reshape(pixData, (t_size, t_size, 3))
average_images[c] = pixImg
return average_images
def get_upsampled_cluster_centers(clustering, num_clusters, latents, descriptors):
"""Generates plot to view the cluster centroids."""
m = 1
cluster_images = np.zeros((num_clusters, Gs.output_shape[3], Gs.output_shape[3], 3), dtype=np.uint8)
for c in range(num_clusters):
distances = clustering.transform(descriptors)[:, c]
ind = np.argsort(distances)[::][:m] # n closest latents
closest_latents = latents[ind]
labels = np.zeros([closest_latents.shape[0], 0], np.float32)
outputs = Gs.run(closest_latents, labels, minibatch_size=minibatch_size, num_gpus=1, out_mul=127.5, out_add=127.5, out_dtype=np.uint8)
cluster_images[c] = np.rollaxis(outputs[0], 0, 3)
return cluster_images
average_images = get_averaged_cluster_centers(kmeans, num_clusters)
cluster_images = get_upsampled_cluster_centers(kmeans, num_clusters, latents, padded_descriptors)
clusters = kmeans.labels_
count = np.bincount(clusters)
hdf5_cluster_path = os.path.join(directory, '{}_{}_clusters.hdf5'.format(network_name, num_clusters))
hdf5_file_clusters = h5py.File(hdf5_cluster_path, "w")
for cluster in range(num_clusters):
this_cluster = np.argwhere(clusters == cluster)
this_descriptors = padded_descriptors[clusters == cluster]
print('cluster {}'.format(cluster))
assert (len(this_descriptors) == count[cluster])
print(this_descriptors.shape)
hdf5_file_clusters.create_dataset("{}".format(cluster), data=this_cluster)
distances = kmeans.transform(this_descriptors)[:, cluster]
indices = np.argsort(distances)[::][:]
print('max index=', np.argmax(indices))
max_dist = np.amax(distances)
print(max_dist)
probabilities = 1 - np.clip(distances / 3000, 0, 1)
probabilities /= probabilities.sum()
print('probabilities between {} and {}'.format(np.amin(probabilities), np.amax(probabilities)))
cdf = [probabilities[0]]
for i in np.arange(1, len(probabilities)):
cdf.append(cdf[-1] + probabilities[i])
assert (len(cdf) == count[cluster])
hdf5_file_clusters.create_dataset("{}_cdf".format(cluster), data=cdf)
hdf5_file_clusters.create_dataset("images", data=cluster_images)
hdf5_file_clusters.create_dataset("averages", data=average_images)
hdf5_file_clusters.close()