-
Notifications
You must be signed in to change notification settings - Fork 0
/
Kernel3avgPool16x16x32.py
277 lines (229 loc) · 10.2 KB
/
Kernel3avgPool16x16x32.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
import tensorflow as tf
from keras import losses
from sklearn.manifold import TSNE
from sklearn.decomposition import PCA
from matplotlib.image import BboxImage
from matplotlib.transforms import Bbox, TransformedBbox
import numpy as np
import os
import tifffile as tiff
from matplotlib import pyplot as plt, gridspec
import time
K = tf.keras.backend
latent_dim = 1024
file_tag = os.path.splitext(os.path.basename(__file__))[0] + "_dim" + str(latent_dim)
batch_size = 128
epochs = 50
num_examples_to_generate = 16
test_train_ratio = 1 / 8 # is only used if test_size is null
test_size = 200 # if test_size is null the test_train_ratio will be used
# data_source_dir = "Track1-RGB/Track1-RGB128x128" #jupyter notebook
# data_source_dir = "Track1-RGB128x128" # xla machine unibw
data_source_dir = "test-RGB128x128split8" # local
generation_path = "generatedImages/" + file_tag + "/"
class Sampling(tf.keras.layers.Layer):
def call(self, inputs):
mean, log_var = inputs
return K.random_normal(tf.shape(log_var)) * K.exp(log_var / 2) + mean
inputs = tf.keras.layers.Input(shape=[128, 128, 3])
z = tf.keras.layers.Conv2D(8, kernel_size=3, padding="same", activation="selu")(inputs)
z = tf.keras.layers.AvgPool2D(pool_size=2)(z)
z = tf.keras.layers.Conv2D(16, kernel_size=3, padding="same", activation="selu")(z)
z = tf.keras.layers.AvgPool2D(pool_size=2)(z)
z = tf.keras.layers.Conv2D(32, kernel_size=3, padding="same", activation="selu")(z)
z = tf.keras.layers.AvgPool2D(pool_size=2)(z)
z = tf.keras.layers.Flatten()(z)
codings_mean = tf.keras.layers.Dense(latent_dim)(z) # μ
codings_log_var = tf.keras.layers.Dense(latent_dim)(z) # γ
codings = Sampling()([codings_mean, codings_log_var])
variational_encoder = tf.keras.Model(
inputs=[inputs], outputs=[codings_mean, codings_log_var, codings])
decoder_inputs = tf.keras.layers.Input(shape=[latent_dim])
x = tf.keras.layers.Dense(16 * 16 * 32)(decoder_inputs)
x = tf.keras.layers.Reshape([16, 16, 32])(x)
x = tf.keras.layers.Conv2DTranspose(16, kernel_size=3, strides=2, padding="same", activation="selu")(x)
x = tf.keras.layers.Conv2DTranspose(8, kernel_size=3, strides=2, padding="same", activation="selu")(x)
outputs = tf.keras.layers.Conv2DTranspose(3, kernel_size=3, strides=2, padding="same", activation="sigmoid")(x)
variational_decoder = tf.keras.Model(inputs=[decoder_inputs], outputs=[outputs])
_, _, codings = variational_encoder(inputs)
reconstructions = variational_decoder(codings)
variational_ae = tf.keras.Model(inputs=[inputs], outputs=[reconstructions])
latent_loss = -0.5 * K.sum(
1 + codings_log_var - K.exp(codings_log_var) - K.square(codings_mean),
axis=-1)
variational_ae.add_loss(K.mean(latent_loss) / (128. * 128. * 3.))
variational_ae.compile(loss=losses.mean_absolute_error, optimizer="rmsprop", metrics=['accuracy', 'mae'])
def get_image_paths(dir):
image_paths = os.listdir(dir)
data = list()
for i in image_paths:
data.append(dir + "/" + i)
return data
def get_image_batch(files, index, batchsize):
batch = list()
for i in range(index, index + batchsize):
batch.append(tiff.imread(files[i]))
batch = np.array(batch, dtype=np.float32)
batch /= 255.
return batch
file_paths = get_image_paths(data_source_dir)
np.random.shuffle(file_paths)
n_files = len(file_paths)
if (test_size != None and n_files <= test_size):
print("error: test_size is larger than the amount of files in the training directory")
split_index = int(n_files // test_train_ratio) if test_size == None else n_files - test_size
training_paths, test_paths = file_paths[:split_index], file_paths[split_index:]
print("number of training images: " + str(len(training_paths)))
print("number of test images: " + str(len(test_paths)))
train_iterations = len(training_paths) // batch_size
test_iterations = len(test_paths) // batch_size
log_file = "log/" + file_tag + "{}.log".format(time.strftime("%d_%m_%Y_%H_%M_%S"))
if not os.path.exists("log"):
os.mkdir("log")
def train():
epoch_means = []
for epoch in range(1, epochs + 1):
epoch_mean = []
for iteration in range(train_iterations):
first_index = iteration * batch_size
batch = get_image_batch(training_paths, first_index, batch_size)
history = variational_ae.train_on_batch(batch, batch)
result = str(history) + "\n" + "Epoch: {}/{}...".format(epoch, epochs) + "Iteration: {}/{}...".format(
iteration + 1, train_iterations) + "Images: {}/{}...".format((iteration + 1) * batch_size,
train_iterations * batch_size) + "\n"
print(result)
f = open(log_file, "a")
f.write(result)
f.close()
epoch_mean.append(history)
epoch_mean = np.mean(epoch_mean, 0)
epoch_means.append(epoch_mean)
print("epoch mean: " + str(epoch_mean))
f = open(log_file, "a")
f.write("epoch mean: " + str(epoch_mean))
f.close()
epoch_means = np.array(epoch_means)
plt.clf()
plt.plot(epoch_means[:,0])
plt.ylim([0.06,0.2])
plt.ylabel("Loss")
plt.xlabel("Epochs")
plt.savefig(generation_path + "loss_graph" + file_tag, bbox_inches='tight', pad_inches=0)
plt.clf()
plt.plot(epoch_means[:, 1])
plt.ylim([0.3, 1])
plt.ylabel("Accuracy")
plt.xlabel("Epochs")
plt.savefig(generation_path + "accuracy_graph" + file_tag, bbox_inches='tight', pad_inches=0)
plt.clf()
plt.plot(epoch_means[:, 2])
plt.ylim([0.06,0.2])
plt.ylabel("Reconstruction Loss MAE")
plt.xlabel("Epochs")
plt.savefig(generation_path + "mae_graph" + file_tag, bbox_inches='tight', pad_inches=0)
if not os.path.exists(generation_path):
os.mkdir(generation_path)
def predictions_and_generations():
pred_on = list()
file_names = os.listdir("RGB-From-Track1128x128split8")
for filename in file_names[0:10]:
pred_on.append(tiff.imread("RGB-From-Track1128x128split8/" + filename))
pred_on = np.array(pred_on, dtype=np.float32)
pred_on /= 255.
latent_log_var, latent_mean, latent_vars = variational_encoder.predict(pred_on)
predictions = variational_decoder.predict(latent_vars)
number_predictions = len(predictions)
fig1, axs1 = plt.subplots(int(number_predictions/2), 1, gridspec_kw={'wspace': 0, 'hspace': 0})
for i in range(0, int(number_predictions/2)):
axs1[i].imshow(pred_on[i])
axs1[i].axis('off')
plt.tight_layout(pad=0)
plt.savefig(generation_path + "inputsCol1" + file_tag, bbox_inches='tight', pad_inches=0)
fig1, axs1 = plt.subplots(int(number_predictions/2), 1, gridspec_kw={'wspace': 0, 'hspace': 0})
for i in range(0, int(number_predictions/2)):
axs1[i].imshow(pred_on[int(number_predictions/2) + i])
axs1[i].axis('off')
plt.tight_layout(pad=0)
plt.savefig(generation_path + "inputsCol2" + file_tag, bbox_inches='tight', pad_inches=0)
fig2, axs2 = plt.subplots(int(number_predictions/2), 1, gridspec_kw={'wspace': 0, 'hspace': 0})
for i in range(0, int(number_predictions/2)):
axs2[i].imshow(predictions[i])
axs2[i].axis('off')
plt.tight_layout(pad=0)
plt.savefig(generation_path + "reconstructionsCol1" + file_tag, bbox_inches='tight', pad_inches=0)
fig2, axs2 = plt.subplots(int(number_predictions/2), 1, gridspec_kw={'wspace': 0, 'hspace': 0})
for i in range(0, int(number_predictions / 2)):
axs2[i].imshow(predictions[int(number_predictions/2) + i])
axs2[i].axis('off')
plt.tight_layout(pad=0)
plt.savefig(generation_path + "reconstructionsCol2" + file_tag, bbox_inches='tight', pad_inches=0)
codings = tf.random.normal(shape=[12, latent_dim])
images_generated = variational_decoder.predict(codings, steps=1)
print(type(images_generated))
print(images_generated.shape)
for i in range(0, len(images_generated)):
plt.clf()
plt.imshow(images_generated[i])
plt.savefig(generation_path + "generated" + str(i))
def get_predominant_class_as_color(img):
flat = img.flatten()
most_common = np.bincount(flat).argmax()
if most_common == 2: # ground
return 'gray'
elif most_common == 5: # vegetation
return 'g'
elif most_common == 6: # building
return 'r'
elif most_common == 9: # water
return 'b'
elif most_common == 65: # clutter
return 'y'
def tsne_vis():
predominant_classes = []
locations = []
average_heights = []
images = []
file_names = os.listdir("RGB-From-Track1128x128split8")
print(len(file_names))
for filename in file_names[0:5000]:
image = tiff.imread("RGB-From-Track1128x128split8/" + filename)
cls = tiff.imread("Track1-Truth128x128split8/" + filename.replace("RGB", "CLS"))
predominant_classes.append(get_predominant_class_as_color(cls))
images.append(image)
if "OMA" in filename:
locations.append('.')
else:
locations.append('s')
images = np.array(images, dtype=np.float32)
images /= 255.
latent_log_var, latent_mean, latent_vars = variational_encoder.predict(images)
predictions = variational_decoder.predict(latent_vars)
print("t-SNE started with: " + str(latent_vars.shape))
embedded = TSNE(perplexity=15, n_iter=5000).fit_transform(latent_vars)
# embedded = PCA(n_components=2).fit_transform(latent_vars)
print("t-SNE finished")
plt.clf()
# plt.scatter(embedded[:, 0], embedded[:, 1], s=1, c=predominant_classes)
fig = plt.figure()
ax = fig.add_subplot(111)
for i in range(0, len(embedded)):
x = embedded[i, 0]
y = embedded[i, 1]
bb = Bbox.from_bounds(x, y, 1, 1)
bb2 = TransformedBbox(bb, ax.transData)
bbox_image = BboxImage(bb2,
norm=None,
origin=None,
clip_on=False)
bbox_image.set_data(images[i])
ax.add_artist(bbox_image)
plt.show()
#start_time = time.time()
#train()
#f = open(log_file, "a")
#f.write(str(time.time() - start_time))
#f.close()
#variational_ae.save_weights("./savedModels/" + file_tag + ".h5")
variational_ae.load_weights("./savedModels/" + file_tag + ".h5")
variational_ae.summary()
predictions_and_generations()