-
Notifications
You must be signed in to change notification settings - Fork 6
/
tcga_dann_4reps.py
208 lines (175 loc) · 8.69 KB
/
tcga_dann_4reps.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
import tensorflow as tf
from keras.backend.tensorflow_backend import set_session
import keras
configtf = tf.ConfigProto()
configtf.gpu_options.allow_growth = True
configtf.gpu_options.visible_device_list = "0"
set_session(tf.Session(config=configtf))
import glob
import os
import sys
from PIL import *
sys.path.insert(0, '/home/sebastian/stain_adversarial_learning/utils/')
from utils_patches import color_augment_patches
from config import *
import numpy as np
import pandas as pd
sys.path.insert(0, '/home/sebastian/stain_adversarial_learning/models_code/')
from models_code.models_keras import GradientReversal
from keras import Model
from keras.layers.core import Dense
from sklearn.metrics import *
from keras.applications.mobilenet import MobileNet
import keras.backend as K
from keras.models import Model
from keras.layers import Dense
from keras import optimizers
from utils.evaluation_utils import evaluate_model_test_tma,evaluate_model_validation_TCGA
from keras_utils import LR_SGD
from imp import reload
import utils.utils_patches
reload(utils.utils_patches)
from utils.utils_patches import patchgen_tcga_only_domains, simplePatchGeneratorTCGA
#Defining batch generators
training_gen_tcga =simplePatchGeneratorTCGA(
input_dir='/mnt/nas3/bigdatasets/Desuto/tcga/graded_diagnostic_patch_dataset/subset_stain_norm/training/',
batch_size=64,
img_shape = (224,224,3),
augmentation_fn=None)
validation_gen_tcga =simplePatchGeneratorTCGA(
input_dir='/mnt/nas3/bigdatasets/Desuto/tcga/graded_diagnostic_patch_dataset/subset_stain_norm/training/',
batch_size=64,
img_shape = (224,224,3),
augmentation_fn=None)
domains_gen_tcga = patchgen_tcga_only_domains(
input_dir='/mnt/nas3/bigdatasets/Desuto/tcga/graded_diagnostic_patch_dataset/subset_stain_norm/all_patches/',
batch_size=64,
img_shape = (224,224,3),
augmentation_fn=None)
mpath = '/home/sebastian/stain_adversarial_learning/models/tcga_dann_gen/'
if not os.path.exists(mpath):
os.makedirs(mpath)
LL =training_gen_tcga.next()
print(LL[0].shape)
lmit = K.variable(1.) # To be able to modify them later.
ldom = K.variable(1.)
history_losses = []
###################Model Training
mpath = '/home/sebastian/stain_adversarial_learning/models/tcga_dann_4reps/'
if not os.path.exists(mpath):
os.makedirs(mpath)
for rep in range(4):
history_losses = []
history_filepath = mpath +'history_'+'rep_'+str(rep)+'_'
#Base model extraction from pre-trained mobilnet with alpha=0.5
base_mnet = MobileNet(input_shape=(224,224,3), alpha=0.5, depth_multiplier=1, dropout=.2, include_top=False,
weights="imagenet", input_tensor=None, pooling='avg', classes=2)
n_domains = len(training_gen_tcga.domains_dict_train) #The domain branch of the network depends on the total #domains
hp_lambda_domain = K.variable(1)
x_top = base_mnet.output
gr = GradientReversal(hp_lambda=hp_lambda_domain)
xd_reversal = gr(x_top)
feats1_gp = Dense(256, name='feats1_gp', activation='relu')(x_top)
feats2_gp = Dense(128, name='feats2_gp', activation='relu')(feats1_gp)
y_gleason = Dense(2, name='output', activation='softmax')(feats2_gp)
feats1_dom = Dense(256, name='feats1_dom', activation='relu')(xd_reversal)
feats2_dom = Dense(128, name='feats2_dom', activation='relu')(feats1_dom)
y_dom = Dense(n_domains, name='output_dom', activation='softmax')(feats2_dom)
model = Model(base_mnet.input, [y_gleason,y_dom])
#model.summary()
losses = {
"output": "categorical_crossentropy",
"output_dom": "categorical_crossentropy" } #categorical_crossentropy vs categorical_crossentropy
# first: train only the top layers (which were randomly initialized)
# i.e. freeze all convolutional layers
for layer in base_mnet.layers:
layer.trainable = False
LR_mult_dict = {'output':0.25, "output_dom":1} #This allows to have a lr of 0.0025 for the domain classifier: 0.0025 = 0.01 * x
sgd_opt = LR_SGD(lr=0.001, momentum=0.9, decay=0.0005, nesterov=False, multipliers = LR_mult_dict)
#With the SGD LR multipliers the modifiers don't work deterministically
model.compile(optimizer=optimizers.Adam(lr=0.001),
loss=losses, loss_weights=[lmit,ldom], metrics = ['accuracy'])
lgp = K.variable(1.) # To be able to modify them later.
ldom = K.variable(1.)
print(model.metrics_names)
K.set_value(model.loss_weights[0],1.)
K.set_value(model.loss_weights[1],0.)
gg = model.layers[-7]
K.set_value(gg.hp_lambda,0.)
history = model.fit_generator(
generator=training_gen_tcga,
steps_per_epoch=20,
epochs=1,
validation_data=training_gen_tcga,
validation_steps=2,
verbose=1
)
change_iter = 100
#for iter_num in range(4000):
print(model.metrics_names)
for epoch in range(5):
for iter_num in range(1000):
#GP features gets updated
K.set_value(lmit,1.)
K.set_value(ldom,0.)
K.set_value(model.loss_weights[0],1.)
K.set_value(model.loss_weights[1],0.)
K.set_value(gg.hp_lambda,0.)
batch_images,labels = training_gen_tcga.next()
history_losses.append(model.train_on_batch(x=batch_images,y=labels))
print(history_losses[-1],str(iter_num))
#the domain features gets updated
K.set_value(lmit,1.)
K.set_value(ldom,1.)
K.set_value(model.loss_weights[0],0.)
K.set_value(model.loss_weights[1],1.)
K.set_value(gg.hp_lambda,-1.)
batch_images,domain_labels = domains_gen_tcga.next()
history_losses.append(model.train_on_batch(x=batch_images,y=domain_labels))
print(history_losses[-1],str(iter_num))
#Adversarial update
K.set_value(lmit,1.)
#alpha_val = (2 /(1+np.exp(-20.*iter_num/change_iter))) - 1.
K.set_value(ldom,1)
K.set_value(model.loss_weights[0],1.)
K.set_value(model.loss_weights[1],1.)
K.set_value(gg.hp_lambda,1.)
batch_images,labels = training_gen_tcga.next()
history_losses.append(model.train_on_batch(x=batch_images,y=labels))
print(history_losses[-1],str(iter_num))
###
df = pd.DataFrame(history_losses, columns=model.metrics_names)
full_path_model_weights = mpath + 'weights_'+ str(lambda_m)+'_rep_'+str(rep) + '_epoch_' + str(epoch) +'_iter_' + str(iter_num)+'.hdf5'
model.save_weights(full_path_model_weights)
df.to_csv(history_filepath + str(lambda_m) + '_iter_'+str(iter_num)+'.log', index=False)
all_external_gp3 = glob.glob('/mnt/nas3/bigdatasets/Desuto/tcga/graded_diagnostic_patch_dataset/subset_stain_norm/external_test/GP3/*jpg')
all_external_gp4 = glob.glob('/mnt/nas3/bigdatasets/Desuto/tcga/graded_diagnostic_patch_dataset/subset_stain_norm/external_test/GP4/*jpg')
all_internal_gp3 = glob.glob('/mnt/nas3/bigdatasets/Desuto/tcga/graded_diagnostic_patch_dataset/subset_stain_norm/internal_test/GP3/*png')
all_internal_gp4 = glob.glob('/mnt/nas3/bigdatasets/Desuto/tcga/graded_diagnostic_patch_dataset/subset_stain_norm/internal_test/GP4/*png')
validation_images_path = '/mnt/nas3/bigdatasets/Desuto/tcga/graded_diagnostic_patch_dataset/subset_stain_norm/val/'
val_models_paths = []
models_measures = []
for rep in range(4):
model_weights_paths = glob.glob(mpath + 'weights_'+ str(lambda_m)+'_rep_'+str(rep)+'*hdf5')
best_val_path, best_val_thres,best_val_max_f1,best_val_roc_auc = '',0,0,0
for wpath in model_weights_paths:
model.load_weights(wpath)
best_thres,max_f1,roc_auc = evaluate_model_validation_TCGA(model,validation_images_path,perf_meassure='f1', func_model=True)
if roc_auc > best_val_roc_auc:
best_val_roc_auc = roc_auc
best_val_max_f1 = max_f1
best_val_thres = best_thres
best_val_path = wpath
print("Better validation model found with the weights path: " + wpath)
print(best_val_roc_auc,best_val_max_f1,best_val_thres)
val_models_paths.append(best_val_path)
model.load_weights(best_val_path)
int_auc, int_f1 = evaluate_model_test_tma(model,all_internal_gp3,all_internal_gp4,best_thres,
return_probs=False, func_model=True,internal_tcga_test = True)
ext_auc, ext_f1 = evaluate_model_test_tma(model,all_external_gp3,all_external_gp4,best_thres,
return_probs=False, func_model=True)
models_measures.append([int_auc, int_f1,ext_auc,ext_f1])
print(val_models_paths)
print('int_auc, int_f1, ext_auc, ext_f1')
print(np.mean(models_measures,axis=0))
print(np.std(models_measures,axis=0))