-
Notifications
You must be signed in to change notification settings - Fork 1
/
pipeline.py
292 lines (257 loc) · 11 KB
/
pipeline.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
import albumentations as A
import cv2
import json
import numpy as np
import os
import seaborn as sns
import sys
import tensorflow as tf
import keras
import shutil
from keras_unet_collection import losses, models
from matplotlib import pyplot as plt
from skimage.transform import resize
from sklearn.model_selection import KFold
from tensorflow import float64
from tensorflow.keras.callbacks import Callback
from tensorflow.keras.optimizers import Adam
shape = (128, 128)
batch_size = 8
K = 10 # denotes cross-validation folds
ms = 3 # denotes pyplot 'marksersize'
save_dir = '/content/saved_models' # specific to Colab instance
sns.set_theme()
model_dict = {
'unet_2d': models.unet_2d,
# 'att_unet_2d': models.att_unet_2d
}
def lr_scheduler(epoch, lr):
decay_rate = 0.5
decay_step = 50
if epoch % decay_step == 0 and epoch > 1:
return lr * decay_rate
return lr
class SegPipeline:
def __init__(self, backbone, weights, model_name, train_dir, test_dir, test_save_dir,
lr, epochs, filters):
self.backbone = backbone
self.weights = weights
self.model_name = model_name
self.model = model_dict[model_name]
self.lr = lr
self.epochs = epochs
self.filters = filters
self.filter_num = []
for i in range(filters):
self.filter_num.append(64*(2**i))
self.x_train, self.y_train = self.get_data(train_dir)
self.x_test, self.y_test = self.get_data(test_dir)
self.save_path = os.path.join(save_dir, model_name + '_{:.0e}_{:1}'.format(self.lr, self.filters))
if not os.path.exists(self.save_path):
os.makedirs(self.save_path)
self.test_save_dir = test_save_dir
if not os.path.exists(self.test_save_dir):
os.makedirs(self.test_save_dir)
def get_data(self, data_dir):
print('Getting data...')
print(data_dir, 'path exists:', os.path.exists(data_dir))
ids = next(os.walk(data_dir))[1]
print('IDs in {0}: {1}'.format(data_dir, ids))
x = np.zeros((len(ids), shape[0], shape[1], 3), dtype=np.uint8)
y = np.zeros((len(ids), shape[0], shape[1], 1), dtype=bool)
for n, name in enumerate(ids):
img_dir = os.path.join(data_dir, name, 'Images')
img_file = os.path.join(img_dir, os.listdir(img_dir)[0])
img = cv2.imread(img_file)
img = resize(img, (shape[0], shape[1]), mode='constant', preserve_range=True)
x[n] = img
mask = np.zeros((shape[0], shape[1], 1), dtype=bool)
mask_dir = os.path.join(data_dir, name, 'Masks')
mask_file = os.path.join(mask_dir, os.listdir(mask_dir)[0])
mask_read = cv2.imread(mask_file, cv2.IMREAD_GRAYSCALE)
mask_read = np.expand_dims(resize(mask_read, (shape[0], shape[1]), mode='constant', preserve_range=True), axis=-1)
mask = np.maximum(mask, mask_read)
y[n] = mask
return x, y
def plot(self, history, path):
fig, axs = plt.subplots(3)
fig.set_figheight(20)
fig.set_figwidth(10)
x = range(1, len(history.history['loss'])+1)
axs[0].plot(x, history.history['loss'], marker='o', markersize=ms, color='mediumvioletred', linewidth=0.5)
#axs[0].plot(x, history.history['val_loss'], marker='o', markersize=ms, color='mediumpurple', linewidth=0.5)
axs[0].set_title('Losses')
axs[0].set_ylim((0, 1))
#axs[0].legend(['train', 'val'], loc='upper right')
axs[1].plot(x, history.history['accuracy'], marker='o', markersize=ms, color='mediumvioletred', linewidth=0.5)
#axs[1].plot(x, history.history['val_accuracy'], marker='o', markersize=ms, color='mediumpurple', linewidth=0.5)
axs[1].set_title('Accuracies')
axs[2].plot(x, history.history['dice_coef'], marker='o', markersize=ms, color='mediumvioletred', linewidth=0.5)
#axs[2].plot(x, history.history['val_dice_coef'], marker='o', markersize=ms, color='mediumpurple', linewidth=0.5)
axs[2].set_title('Dice coef plots')
if not os.path.exists(path):
os.makedirs(path)
plt.savefig(os.path.join(path, 'loss_acc_dice.png'))
plt.show()
# uses Albumentations library
def alb_aug(self, image, mask):
mask = np.float32(mask) # img already converted
transform = A.Compose([
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(p=0.5), #default (-0.2, 0.2) for contrast and brightness
A.Affine(scale=[0.75, 1], shear=[-30, 30], p=0.5)
])
transformed = transform(image=image, mask=mask)
return transformed['image'], transformed['mask']
# no EarlyStop
# no val
def train_single(self):
aug_x_train = np.zeros((self.x_train.shape[0], 128, 128, 3))
aug_y_train = np.zeros((self.y_train.shape[0], 128, 128, 1))
for ind in range(self.x_train.shape[0]):
aug_x_train[ind], aug_y_train[ind] = self.alb_aug(self.x_train[ind], self.y_train[ind])
seg_model = self.model(
input_size=(128, 128, 3),
filter_num=self.filter_num,
n_labels=1,
stack_num_down=2,
stack_num_up=2,
activation='ReLU',
output_activation='Sigmoid',
name='seg_model',
backbone=self.backbone,
weights=self.weights
)
seg_model.compile(
loss='binary_crossentropy',
optimizer=Adam(learning_rate = self.lr),
metrics=['accuracy', losses.dice_coef]
)
history = seg_model.fit(
aug_x_train, aug_y_train,
verbose=1,
batch_size=batch_size,
shuffle=False,
epochs=self.epochs,
callbacks=[
tf.keras.callbacks.LearningRateScheduler(
lr_scheduler, verbose=1
)
]
)
self.plot(history, os.path.join(self.test_save_dir))
seg_model.save(os.path.join(self.test_save_dir, 'model'), overwrite=True)
loss, acc, dice = seg_model.evaluate(
self.x_test, self.y_test, verbose=1
)
with open(os.path.join(self.test_save_dir, 'test_scores.txt'), 'w') as text_file:
text_file.write(
'Test loss >>> ' + str(loss) + '\n\n'
+ 'Test acc >>> ' + str(acc) + '\n\n'
+ 'Test dice >>> ' + str(dice) + '\n\n'
)
def train(self):
histories = []
val_losses = []
val_accs = []
val_dices = []
kfold = KFold(n_splits=K, shuffle=True)
for i, (train, val) in enumerate(kfold.split(self.x_train, self.y_train)):
model_path = os.path.join(self.save_path, 'model'+str(i))
print("Beginning training on fold {0}".format(i))
print('Train: {0}'.format(train.shape))
print('Val: {0}'.format(val.shape))
aug_x_train = np.zeros((self.x_train.shape[0], 128, 128, 3))
aug_y_train = np.zeros((self.y_train.shape[0], 128, 128, 1))
for ind in range(self.x_train.shape[0]):
aug_x_train[ind], aug_y_train[ind] = self.alb_aug(self.x_train[ind], self.y_train[ind])
seg_model = self.model(
input_size=(128, 128, 3),
filter_num=self.filter_num,
n_labels=1,
stack_num_down=2,
stack_num_up=2,
activation='ReLU',
output_activation='Sigmoid',
name='seg_model',
backbone=self.backbone,
weights=self.weights
)
seg_model.compile(
loss='binary_crossentropy',
optimizer=Adam(learning_rate = self.lr),
metrics=['accuracy', losses.dice_coef]
)
history = seg_model.fit(
aug_x_train[train], aug_y_train[train],
verbose=1,
batch_size=batch_size,
validation_data=(
self.x_train[val],
self.y_train[val]),
shuffle=False,
epochs=self.epochs,
callbacks=[
tf.keras.callbacks.ModelCheckpoint(
filepath=model_path,
monitor='val_dice_coef',
verbose=1,
save_best_only=True,
save_freq='epoch', # required for custom metrics, periods=20 supplies true frequency
mode='max',
period=20
),
tf.keras.callbacks.EarlyStopping(
monitor='val_dice_coef',
verbose=1,
patience=40,
mode='max'
),
tf.keras.callbacks.LearningRateScheduler(
lr_scheduler, verbose=1
)
]
)
print('Metrics names:', seg_model.metrics_names)
self.plot(history, os.path.join(self.save_path, 'figs', 'fold{0}'.format(i)))
print('Retraining model from', model_path)
saved_model = tf.keras.models.load_model(
os.path.join(model_path),
custom_objects = {'dice_coef': losses.dice_coef}
)
val_loss, val_acc, val_dice = saved_model.evaluate(self.x_train[val], self.y_train[val])
val_losses.append(val_loss)
val_accs.append(val_acc)
val_dices.append(val_dice)
print('Val loss, val acc, val dice resp:', val_loss, val_acc, val_dice)
shutil.rmtree(model_path)
with open(os.path.join(self.save_path, 'cv_scores.txt'), 'a') as text_file:
text_file.write(
'Val losses >>> ' + str(val_losses) + '\n\n'
+ 'Val accs >>> ' + str(val_accs) + '\n\n'
+ 'Val dices >>> ' + str(val_dices) + '\n\n'
+ 'Val dice avg: ' + str(sum(val_dices)/len(val_dices))
)
if __name__ == '__main__':
print('Running seg_pipline.py...')
if not os.path.exists(save_dir):
os.makedirs(save_dir)
with open(os.path.join(save_dir, "specs.txt"), "w") as text_file:
text_file.write(
'Backbone: '+sys.argv[1]+'\n'
+'Weights: '+sys.argv[2]+'\n'
+'Model: '+sys.argv[3]+'\n'
+'LR: '+str(sys.argv[6])+'\n'
+'Epochs: '+str(sys.argv[7])+'\n'
+'Filters: '+str(sys.argv[8])
)
seg_pipeline = SegPipeline(
*sys.argv[1:6],
sys.argv[6],
float(sys.argv[7]),
int(sys.argv[8]),
int(sys.argv[9])
)
seg_pipeline.train()
# for testing
# seg_pipeline.train_single()