-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain_bayesian_meshnet.py
executable file
·152 lines (132 loc) · 6.53 KB
/
train_bayesian_meshnet.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
#!/usr/bin/env python3
# Imports
#import nobrainer
import tensorflow as tf
import os
from kwyk_data import get_dataset
from baysian_meshnet import variational_meshnet
#from kwyk_losses import MaskedCategoricalCrossEntropy#,Dice_Cce,DiceLoss
from kwyk_utils import save_parameters, save_output, calcualte_dice
# constants
#root_path = '/om/user/satra/kwyk/tfrecords/'
#root_path = '/om2/user/hodaraja/kwyk/nobrainer_scripts/'
root_path = "data/"
# to run the code on Satori
#root_path = "/nobackup/users/abizeul/kwyk/tfrecords/"
train_pattern = root_path+"single_volume-000.tfrec"
eval_pattern = root_path + "single_volume-000.tfrec"
#train_pattern = root_path+"data-train_shard-*.tfrec"
#eval_pattern = root_path + "data-evaluate_shard-*.tfrec"
n_classes =50
volume_shape = (256, 256, 256)
block_shape = (64,64,64)
EPOCHS = 150
lr = 5e-04
BATCH_SIZE = 4
num_training_brains = 1
#num_training_brains = 10600
num_examples = int(((volume_shape[0]/block_shape[0])**3)* num_training_brains/BATCH_SIZE)
#num_examples=1
one_hot_label=True
initial_epoch = 0 ; scaling_start_epoch=10; scaling_increase_per_epoch = 0.3
warmup_factor=0
model_name = "kwyk_cce_kl_nbg_b{}_cl{}".format(block_shape[0], n_classes)
checkpoint_dir = os.path.join("training_files",model_name,"training_checkpoints")
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt_{epoch}")
print("--------- Loading data -------------")
dataset_train = get_dataset(train_pattern,volume_shape, BATCH_SIZE, block_shape, n_classes,
one_hot_label=one_hot_label,
filter_background=True)
dataset_eval = get_dataset(eval_pattern,volume_shape, BATCH_SIZE, block_shape, n_classes, training= False,one_hot_label=one_hot_label)
print("_________ data loaded with block_size: {}, batch_size: {}___________".format(block_shape[0], BATCH_SIZE))
if initial_epoch >= scaling_start_epoch:
warmup_factor = tf.convert_to_tensor(min(1., warmup_factor + (initial_epoch - scaling_start_epoch) * scaling_increase_per_epoch))
kl_beta=tf.Variable(warmup_factor, dtype=tf.float32)
# create model
model = variational_meshnet(
n_classes = n_classes,
input_shape = block_shape+(1,),
receptive_field=129,
filters=96,
scale_factor = num_examples,
dropout=None,
batch_size= BATCH_SIZE,
warmup_factor=kl_beta,
)
optimizer = tf.keras.optimizers.Adam(lr=lr)
#loss_fn = DiceLoss(axis=(1,2,3))
#loss_fn = MaskedCategoricalCrossEntropy()
#loss_fn = Dice_Cce(axis =(1,2,3), ignore_background = True)
loss_fn= tf.keras.losses.CategoricalCrossentropy()
checkpoint = tf.train.Checkpoint(optimizer=optimizer, model=model)
model.compile(optimizer, loss=loss_fn,
metrics=['categorical_accuracy'],
#loss_weights=class_weights,
experimental_run_tf_function=False)
#training loop
train_accuracy, train_loss = [], []
valid_accuracy, valid_loss = [], []
for epoch in range(EPOCHS):
print('Epoch number ',epoch)
epoch_accuracy, epoch_loss, epoch_dice = [], [], []
for steps, (batch_x,batch_y) in enumerate(dataset_train):
batch_loss, batch_accuracy = model.train_on_batch(batch_x, batch_y)
epoch_accuracy.append(batch_accuracy)
epoch_loss.append(batch_loss)
# calculate dice
result = model.predict_on_batch(batch_x)
dice_score = calcualte_dice(batch_y,result,n_classes,axis=(1,2,3),one_hot_label=one_hot_label)
epoch_dice.append(dice_score)
# save checkpoint and output every 10 epoch
if epoch % 10 == 0:
checkpoint.save(checkpoint_prefix.format(epoch=epoch))
output_path = "./training_files/" + model_name + "/out_epoch-{}".format(epoch)
save_output(output_path, model, dataset_eval, volume_shape, block_shape, one_hot_label=one_hot_label)
save_parameters(output_path+"_prm.out",model_name,loss=tf.reduce_mean(epoch_loss).numpy().tolist(),
accuracy=tf.reduce_mean(epoch_accuracy).numpy().tolist(),
dice=tf.reduce_mean(epoch_dice).numpy().tolist())
print("loss:{}, accuracy:{}, dice:{}".format(tf.reduce_mean(epoch_loss),
tf.reduce_mean(epoch_accuracy),
tf.reduce_mean(epoch_dice)))
#adjusting the warmup factor
if epoch >= scaling_start_epoch:
new_warmup_factor = tf.convert_to_tensor(min(1., warmup_factor + (epoch - scaling_start_epoch) * scaling_increase_per_epoch), dtype=tf.float32 )
kl_beta.assign(new_warmup_factor)
print("epoch {}, new kl_factor {}".format(epoch, kl_beta.numpy()))
#evaluation
epoch_val_accuracy = []
epoch_val_loss = []
eval_dice = []
for eval_x, eval_y in dataset_eval.take(num_examples):
batch_val_loss, batch_val_accuracy = model.test_on_batch(eval_x, eval_y)
epoch_val_loss.append(batch_val_loss)
epoch_val_accuracy.append(batch_val_accuracy)
# calculate dice
result = model.predict_on_batch(eval_x)
dice_score = calcualte_dice(eval_y, result,n_classes,axis=(1,2,3), one_hot_label=one_hot_label)
eval_dice.append(dice_score)
print("Eval_loss: {}, Eval_accuracy: {}, Eval_dice: {}".format(tf.reduce_mean(epoch_val_loss),
tf.reduce_mean(epoch_val_accuracy),
tf.reduce_mean(eval_dice)))
# save model
#saved_model_path=os.path.join("./training_files",model_name,"saved_model/")
#model.save(saved_model_path, save_format='tf')
saved_weight_path=os.path.join("./training_files",model_name,"model_weights.hd5/")
model.save_weights(saved_weight_path)
saved_param_path = os.path.join("./training_files",model_name,"model_parameters.json")
save_parameters(saved_param_path,model_name,
block_shape = block_shape,
batch_size = BATCH_SIZE,
n_classes = n_classes,
lr = lr,
n_epochs = EPOCHS,
num_training_brains = num_training_brains,
loss_fn = loss_fn.name,
kl_warmup = scaling_start_epoch,
one_hot_label = one_hot_label
)
# test and save output
print("------------ test--------------")
test_dataset = get_dataset(train_pattern, volume_shape, BATCH_SIZE, block_shape, n_classes, training= False)
output_file = os.path.join("training_files",model_name,"output_test_b{}_cl{}".format(block_shape[0],n_classes))
save_output(output_file,model,test_dataset,volume_shape,block_shape)