-
Notifications
You must be signed in to change notification settings - Fork 0
/
supcon.py
395 lines (372 loc) · 13.5 KB
/
supcon.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
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
"""Supervised contrastive learning using two CNN encoder types and either tabular and image data or just image data. No balanced batching."""
import sys
import numpy as np
import pandas as pd
import tensorflow as tf
from sklearn.model_selection import train_test_split
from functions import (
SupervisedContrastiveLoss,
add_metrics,
add_projection_head,
create_classifier,
create_classifier_imgs_only,
create_data_augmentation_module,
create_encoder,
)
# Weights should be loaded from pretext task.
# Full training data
train_df = pd.read_csv("Training_data_ready_with_target_and_ID.csv")
train_df.drop(columns="Unnamed: 0", inplace=True)
# Get list of images in training dataset
trainDataDir = "D:/Downloads/siim-isic-melanoma-classification/jpeg_adj_sample/train/"
# trainDataDir = "D:/Downloads/siim-isic-melanoma-classification/jpeg_adj_final/train/"
image_generator = tf.keras.preprocessing.image.ImageDataGenerator().flow_from_directory(
trainDataDir,
shuffle=False,
class_mode="binary",
target_size=(224, 224),
batch_size=400, # 33126 #400,
) # total amount of 0 and 1 values
images, labels = next(image_generator)
# Getting the ordered list of filenames for the images
image_files = pd.Series(image_generator.filenames)
# Form = 0\ISIC_0015719.jpg
image_files = image_files.str.split("\\", expand=True)[1].str[:-4]
# Form = ISIC_0015719
image_files = list(image_files)
# Now reduce the training data df to reflect chosen data
train_df = train_df[train_df["image_name"].isin(image_files)]
# Check the order of records in the training data vs image_files
if image_files == list(train_df.image_name):
print("The lists are identical")
# Drop image name
train_df.drop(columns=["image_name"], inplace=True)
else:
print("The lists are not identical")
sys.exit()
# Set up the train/ validation split
(train_tabular, valid_tabular, train_images, valid_images) = train_test_split(
train_df, images, test_size=0.2, random_state=42
) # 33,126 images initially
print(f"Training tabular data shape = {train_tabular.shape}")
print(f"Validation tabular data shape = {valid_tabular.shape}")
print(f"Training image data shape = {train_images.shape}")
print(f"Validation image data shape = {valid_images.shape}")
# Set aside the target
train_target = train_tabular["target"]
train_target = np.array(train_target)
print(f"Train target shape = {train_target.shape}")
valid_target = valid_tabular["target"]
valid_target = np.array(valid_target)
print(f"Validation target shape = {valid_target.shape}")
# Now drop target from the original dataframes
train_tabular.drop(columns=["target"], inplace=True)
valid_tabular.drop(columns=["target"], inplace=True)
# ------------------------------------------------------
# Modelling
# ------------------------------------------------------
# Shared parameters
numClasses = 2
input_shape = (224, 224, 3)
learningRate = 0.001
batchSize = 8 # 256
hiddenUnits = 512
projectionUnits = 128
numEpochs = 100
dropoutRate = 0.1
temperature = 0.05
optimiser = tf.keras.optimizers.Adam(learning_rate=learningRate)
# Shared functions
# Early stopping on validation loss
callback_EarlyStopping = tf.keras.callbacks.EarlyStopping(
monitor="val_loss", patience=5, restore_best_weights=True
)
# Image augmentation
data_aug = create_data_augmentation_module()
# Setting the state of the normalization layer.
data_aug.layers[0].adapt(train_images)
encoder_types = ["ResNet50V2", "InceptionV3"]
# ---------------------------------------------------------------------------------
# Baseline classification models using images only
# ---------------------------------------------------------------------------------
for enc in encoder_types:
print(f"\nBaseline classification model using {enc}, only images\n")
# Setup encoder
encoder = create_encoder(
encoder_type=enc,
input_shape=input_shape,
data_augmentation=data_aug,
encoder_name=f"{enc}_encoder",
encoder_weights_location=f"saved_models/encoder_weights_{enc}_MSE.h5",
)
encoder.summary()
# Setup classifier
classifier = create_classifier_imgs_only(
encoder_module=encoder,
model_name=f"{enc}_baseline_classifier",
input_shape=input_shape,
hidden_units=hiddenUnits,
dropout_rate=dropoutRate,
optimizer=optimiser,
trainable=True,
)
classifier.summary()
# Define all the callbacks
# Logging
callback_CSVLogger = tf.keras.callbacks.CSVLogger(
f"CSVLogger/train_baseline_{enc}.csv"
)
# Training
history = classifier.fit(
x=train_images,
y=train_target.reshape(
-1,
),
batch_size=batchSize,
epochs=numEpochs,
validation_data=[
valid_images,
valid_target.reshape(
-1,
),
],
callbacks=[callback_EarlyStopping, callback_CSVLogger],
verbose=2,
)
# Save the entire model as a SavedModel.
classifier.save(f"saved_models/train_baseline_{enc}")
# -----------------------------------------------------------------------------------
# Supervised contrastive learning model with images only
# -----------------------------------------------------------------------------------
for enc in encoder_types:
print(
f"\nSupervised contrastive learning classification model using {enc}, only images\n"
)
# Pre-train the encoder
encoder = create_encoder(
encoder_type=enc,
input_shape=input_shape,
data_augmentation=data_aug,
encoder_name=f"{enc}_encoder_supcon",
encoder_weights_location=f"saved_models/encoder_weights_{enc}_MSE.h5",
)
encoder.summary()
encoder_with_projection_head = add_projection_head(
encoder_module=encoder,
model_name=f"{enc}_encoder_with_projection_head",
input_shape=input_shape,
projection_units=projectionUnits,
)
encoder_with_projection_head.compile(
optimizer=optimiser,
loss=SupervisedContrastiveLoss(temperature),
)
encoder_with_projection_head.summary()
# Logging
callback_CSVLogger = tf.keras.callbacks.CSVLogger(
f"CSVLogger/supcon_pretrained_encoder_{enc}.csv"
)
# Pre-training encoder
history = encoder_with_projection_head.fit(
x=train_images,
y=train_target.reshape(
-1,
),
batch_size=batchSize,
epochs=numEpochs,
validation_data=[
valid_images,
valid_target.reshape(
-1,
),
],
callbacks=[callback_EarlyStopping, callback_CSVLogger],
verbose=2,
)
# Train the classifier with the frozen encoder
classifier = create_classifier_imgs_only(
encoder_module=encoder,
model_name=f"supcon_classifier_{enc}_frozen_encoder",
input_shape=input_shape,
hidden_units=hiddenUnits,
dropout_rate=dropoutRate,
optimizer=optimiser,
trainable=False,
)
classifier.summary()
# Logging
callback_CSVLogger = tf.keras.callbacks.CSVLogger(
f"CSVLogger/supcon_encoder_{enc}.csv"
)
# Train the classifier with the frozen encoder
history = classifier.fit(
x=train_images,
y=train_target.reshape(
-1,
),
batch_size=batchSize,
epochs=numEpochs,
validation_data=[
valid_images,
valid_target.reshape(
-1,
),
],
callbacks=[callback_EarlyStopping, callback_CSVLogger],
verbose=2,
)
# Save the entire model as a SavedModel.
classifier.save(f"saved_models/supcon_encoder_{enc}")
# ---------------------------------------------------------------------------------
# Baseline classification models using image and tabular data
# ---------------------------------------------------------------------------------
for enc in encoder_types:
print(f"\nBaseline classification model using {enc}, images and data\n")
# Setup encoder
encoder = create_encoder(
encoder_type=enc,
input_shape=input_shape,
data_augmentation=data_aug,
encoder_name=f"{enc}_encoder",
encoder_weights_location=f"saved_models/encoder_weights_{enc}_MSE.h5",
)
encoder.summary()
# Setup classifier
classifier = create_classifier(
encoder_module=encoder,
model_name=f"{enc}_baseline_classifier",
input_shape=input_shape,
hidden_units=hiddenUnits,
dropout_rate=dropoutRate,
optimizer=optimiser,
trainable=True,
)
classifier.summary()
# Define all the callbacks
# Logging
callback_CSVLogger = tf.keras.callbacks.CSVLogger(
f"CSVLogger/train_baseline_{enc}_incl_tabular.csv"
)
# Training
history = classifier.fit(
x=[train_images, train_tabular],
y=np.array(train_target).reshape(-1, 1),
batch_size=batchSize,
epochs=numEpochs,
validation_data=[
[valid_images, valid_tabular],
np.array(valid_target).reshape(-1, 1),
],
callbacks=[callback_EarlyStopping, callback_CSVLogger],
verbose=2,
)
# Save the entire model as a SavedModel.
classifier.save(f"saved_models/train_baseline_{enc}_incl_tabular")
# -----------------------------------------------------------------------------------
# Supervised contrastive learning model with image and tabular data
# -----------------------------------------------------------------------------------
for enc in encoder_types:
print(f"\nSupervised contrastive learning classification model using {enc}\n")
# Pre-train the encoder
encoder = create_encoder(
encoder_type=enc,
input_shape=input_shape,
data_augmentation=data_aug,
encoder_name=f"{enc}_encoder_supcon",
encoder_weights_location=f"saved_models/encoder_weights_{enc}_MSE.h5",
)
encoder.summary()
encoder_with_projection_head = add_projection_head(
encoder_module=encoder,
model_name=f"{enc}_encoder_with_projection_head",
input_shape=input_shape,
projection_units=projectionUnits,
)
encoder_with_projection_head.compile(
optimizer=optimiser,
loss=SupervisedContrastiveLoss(temperature),
)
encoder_with_projection_head.summary()
# Logging
callback_CSVLogger = tf.keras.callbacks.CSVLogger(
f"CSVLogger/supcon_pretrained_encoder_{enc}_incl_tabular.csv"
)
# Pre-training encoder
history = encoder_with_projection_head.fit(
x=train_images,
y=np.array(train_target).reshape(-1, 1),
batch_size=batchSize,
epochs=numEpochs,
validation_data=[valid_images, np.array(valid_target).reshape(-1, 1)],
callbacks=[callback_EarlyStopping, callback_CSVLogger],
verbose=2,
)
# Train the classifier with the frozen encoder
classifier = create_classifier(
encoder_module=encoder,
model_name=f"supcon_classifier_{enc}_frozen_encoder",
input_shape=input_shape,
hidden_units=hiddenUnits,
dropout_rate=dropoutRate,
optimizer=optimiser,
trainable=False,
)
classifier.summary()
# Logging
callback_CSVLogger = tf.keras.callbacks.CSVLogger(
f"CSVLogger/supcon_encoder_{enc}_incl_tabular.csv"
)
# Train the classifier with the frozen encoder
history = classifier.fit(
x=[train_images, train_tabular],
y=np.array(train_target).reshape(-1, 1),
batch_size=batchSize,
epochs=numEpochs,
validation_data=[
[valid_images, valid_tabular],
np.array(valid_target).reshape(-1, 1),
],
callbacks=[callback_EarlyStopping, callback_CSVLogger],
verbose=2,
)
# Save the entire model as a SavedModel.
classifier.save(f"saved_models/supcon_encoder_{enc}_incl_tabular")
# -----------------------------------------------------------------------------------
# Adding metrics
# -----------------------------------------------------------------------------------
# Add metrics to csv file
add_metrics(
hist_filelocation="CSVLogger/train_baseline_InceptionV3_incl_tabular.csv",
saved_name="CSVLogger/train_baseline_InceptionV3_incl_tabular_added_metrics.csv",
)
add_metrics(
hist_filelocation="CSVLogger/train_baseline_ResNet50V2_incl_tabular.csv",
saved_name="CSVLogger/train_baseline_ResNet50V2_incl_tabular_added_metrics.csv",
)
add_metrics(
hist_filelocation="CSVLogger/supcon_encoder_InceptionV3_incl_tabular.csv",
saved_name="CSVLogger/supcon_encoder_InceptionV3_incl_tabular_added_metrics.csv",
)
add_metrics(
hist_filelocation="CSVLogger/supcon_encoder_ResNet50V2_incl_tabular.csv",
saved_name="CSVLogger/supcon_encoder_ResNet50V2_incl_tabular_added_metrics.csv",
)
add_metrics(
hist_filelocation="CSVLogger/supcon_encoder_InceptionV3.csv",
saved_name="CSVLogger/supcon_encoder_InceptionV3_added_metrics.csv",
)
add_metrics(
hist_filelocation="CSVLogger/supcon_encoder_ResNet50V2.csv",
saved_name="CSVLogger/supcon_encoder_ResNet50V2_added_metrics.csv",
)
add_metrics(
hist_filelocation="CSVLogger/supcon_encoder_InceptionV3.csv",
saved_name="CSVLogger/supcon_encoder_InceptionV3_added_metrics.csv",
)
add_metrics(
hist_filelocation="CSVLogger/train_baseline_InceptionV3.csv",
saved_name="CSVLogger/train_baseline_InceptionV3_added_metrics.csv",
)
# -----------------------------------------------------------------------------------
# Plotting is in another python file
# -----------------------------------------------------------------------------------