/
train_VQ-VAE.py
165 lines (112 loc) · 6.04 KB
/
train_VQ-VAE.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
import tensorflow as tf
from tensorflow.keras.losses import cosine_proximity, categorical_crossentropy
from tensorflow.keras.optimizers import SGD, Adam
from tensorflow.keras.callbacks import LearningRateScheduler, EarlyStopping, ModelCheckpoint
from tensorflow.keras.callbacks import TensorBoard
import math
import os
import argparse
from scipy.spatial.distance import cosine
from dataloader import *
from models import *
from utility import *
from tensorflow.keras.backend import set_session
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
set_session(tf.Session(config=gpu_config))
#################################################### parameters ##########################################################
if not files_exist():
look_ups()
if not os.listdir(str(os.path.dirname(os.path.abspath(__file__))) +'/data/glove_data'):
print('Download the GloVe embeddings from the webpage first!')
if not len(os.listdir(str(os.path.dirname(os.path.abspath(__file__))) +'/data/subjects'))==15:
print('Download the fMRI data from all the subjects from the webpage first!')
parser = argparse.ArgumentParser()
parser.add_argument('-subject', type = str, default = 'M15')
parser.add_argument('-class_model', type= int, default = 1)
args = parser.parse_args()
subject = args.subject
save_file =str(os.path.dirname(os.path.abspath(__file__))) + '/model_weights/Subject' +subject+'_VQ_VAE_model.h5'
#Glove prediction model or classifiction model:
class_model = args.class_model
if class_model:
class_weight = 1.5
glove_weight = 1.0
else:
class_weight = 0.0
glove_weight = 1.8
#################################################### data load ##########################################################
data_train, data_test, glove_train, glove_test, data_fine, data_fine_test, glove_fine, glove_fine_test = dataloader_sentence_word_split_new_matching_all_subjects(subject)
class_fine_test =np.eye(180)
class_fine = np.reshape(np.tile(class_fine_test,(42,1)),(7560,180))
class_fine_test = np.reshape(np.tile(class_fine_test,(3,1)),(540,180))
class_train = np.zeros((4530,180))
#################################################### losses & schedule ##########################################################
initial_lrate = 0.001
epochs_drop = 10.0
epochs=100
batch_size= 180
commitment_cost = 0.25
def step_decay(epoch):
drop = 0.3
lrate = initial_lrate * math.pow(drop,
math.floor((1+epoch)/epochs_drop))
return lrate
def vq_vae_loss_wrapper(commitment_cost, quantized, x_inputs):
def vq_vae_loss(x, x_hat):
recon_loss = cosine_proximity(x,x_hat)
e_latent_loss = K.mean((K.stop_gradient(quantized) - x_inputs) ** 2)
q_latent_loss = K.mean((quantized - K.stop_gradient(x_inputs)) ** 2)
loss = q_latent_loss + commitment_cost * e_latent_loss
return recon_loss + loss
return vq_vae_loss
def mean_distance_loss(y_true, y_pred):
total = 0
total_two = 0
val = 179
for i in range((val+1)):
if i == 0:
total += (val*cosine_proximity(y_true,y_pred))
else:
rolled = tf.manip.roll(y_pred, i, axis=0)
total_two -= cosine_proximity(y_true,rolled)
return total_two/val + total/val
#################################################### model ##########################################################
optimizer = Adam(lr=1e-3,amsgrad=True)
model, enc, enc_inputs = encdec_VQVAE()
loss_fMRI = vq_vae_loss_wrapper(commitment_cost, enc, enc_inputs)
model.compile(loss= {'pred_glove':loss_fMRI,'fMRI_rec':cosine_proximity, 'pred_class':categorical_crossentropy},loss_weights=[1.0,glove_weight,class_weight],optimizer=optimizer, metrics=['accuracy'])
##################################################### callbacks ########################################################
callback_list = []
if class_model:
reduce_lr = LearningRateScheduler(step_decay)
callback_list.append(reduce_lr)
early = EarlyStopping(monitor= 'val_pred_class_loss', patience=10)
callback_list.append(early)
model_checkpoint_callback = ModelCheckpoint(filepath=save_file,save_weights_only=True,monitor='val_pred_class_acc',mode='max',save_best_only=True)
callback_list.append(model_checkpoint_callback)
else:
reduce_lr = LearningRateScheduler(step_decay)
callback_list.append(reduce_lr)
early = EarlyStopping(monitor= 'val_pred_glove_loss', patience=10)
callback_list.append(early)
model_checkpoint_callback = ModelCheckpoint(filepath=save_file,save_weights_only=True,monitor='val_pred_glove_loss',mode='min',save_best_only=True)
callback_list.append(model_checkpoint_callback)
##################################################### fit & save #######################################################
model.fit([data_fine], [data_fine,glove_fine,class_fine], batch_size=batch_size, epochs=epochs, verbose=2, callbacks=callback_list, validation_data=([data_fine_test], [data_fine_test,glove_fine_test,class_fine_test])) #,workers=5,use_multiprocessing=True , validation_data=([data_test], [data_test,glove_test])
model.load_weights(save_file)
##################################################### Evaluation #######################################################
rec, glove_pred_test, pred_class = model.predict(data_fine_test)
text_all = open(str(os.path.dirname(os.path.abspath(__file__))) + "/results/Subject" +subject+"_VQ_VAE_model.txt", "a+")
if class_model:
accuracy, accuracy_five, accuracy_ten = top_5(pred_class,class_fine_test)
text_all.write('Accuracy_top1: ' + str(accuracy) + ' Accuracy_top5: ' + str(accuracy_five) + 'Accuracy_top10: ' + str(accuracy_ten))
text_all.write('\n')
else:
accuracy_test = np.zeros((3))
accuracy_test[0] = evaluation(glove_pred_test[:180,:],glove_fine_test[:180,:])
accuracy_test[1] = evaluation(glove_pred_test[180:360,:],glove_fine_test[180:360,:])
accuracy_test[2] = evaluation(glove_pred_test[360:,:],glove_fine_test[360:,:])
text_all.write('Pairwise Accuracy: ' + str(np.mean(accuracy_test)))
text_all.write('\n')
text_all.close()