-
Notifications
You must be signed in to change notification settings - Fork 3
/
fpiSubmit.py
268 lines (204 loc) · 9.21 KB
/
fpiSubmit.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
import numpy as np
import itertools
import copy
from datetime import datetime
import os
import pickle
from sklearn.metrics import average_precision_score
import tensorflow as tf
import readData
import self_sup_task
from models.wide_residual_network import create_wide_residual_network_selfsup
from scipy.signal import savgol_filter
from utils import save_roc_pr_curve_data
import gc
def train_folder(input_dir,output_dir,mode,data):
gpu = tf.config.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(gpu[0], True)
data_frame = get_data_frame(data,input_dir,shuffle_order=True)
mdl = get_mdl(data,data_frame,restore=False)
submit_train(mdl,data_frame,output_dir,data)
return
def predict_folder(input_dir,output_dir,mode,data):
#K.manual_variable_initialization(True)
gpu = tf.config.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(gpu[0], True)
data_frame = get_data_frame(data,input_dir,shuffle_order=False)
mdl = get_mdl(data,data_frame,restore=True)
submit_test(mdl,data_frame,output_dir,mode)
return
def get_data_frame(data,input_dir,shuffle_order=False,load_labels=False):
if 'brain' in data:
batch_dim = [256,256,256,1]
primary_axis = 2
elif 'abdom' in data:
batch_dim = [512,512,512,1]
primary_axis = 1
else:
raise ValueError("data type not correctly defined. Either choose 'brain','abdom', or add a new definition")
data_frame = readData.data_frame(batch_dim,primary_axis)
input_list = os.listdir(input_dir)
data_frame.load_data(input_list,input_dir,shuffle_order=shuffle_order,load_labels=load_labels)
return data_frame
def get_mdl(data,data_frame,restore=False):
if 'brain' in data:
n, k = (16,4)#network size
net_f='create_wide_residual_network_dec'
n_classes = 1
model_dir = '/workspace/restore_dir/brain/'
elif 'abdom' in data:
n, k = (19,4)#network size
net_f='create_wide_residual_network_decdeeper'
n_classes = 5
model_dir = '/workspace/restore_dir/abdom/'
else:
raise ValueError("data type not correctly defined. Either choose 'brain','abdom', or add a new definition")
if restore:
#grab weights and build model
model_fnames = os.listdir(model_dir)
model_fnames = [fn for fn in model_fnames if 'weights' in fn][0]
model_path = os.path.join(model_dir,model_fnames)
print(model_path)
mdl = tf.keras.models.load_model(model_path)
else:
#build new model
mdl = create_wide_residual_network_selfsup(data_frame.batch_dim[1:],
n_classes, n, k, net_f=net_f)
return mdl
@tf.function
def train_step(mdl,x, y):
loss_fn = mdl.compiled_loss
with tf.GradientTape() as tape:
logits = mdl(x, training=True)
loss_value = loss_fn(y, logits)
grads = tape.gradient(loss_value, mdl.trainable_weights)
mdl.optimizer.apply_gradients(zip(grads, mdl.trainable_weights))
mdl.compiled_metrics.update_state(y, logits)
return loss_value
@tf.function
def test_step(mdl,x, y):
loss_fn = mdl.compiled_loss
logits = mdl(x, training=False)
loss_value = loss_fn(y, logits)
return loss_value
@tf.function
def pred_step(mdl,x):
pred = mdl(x, training=False)
return pred
def grouped(iterable, n):
#get n elements at a time
return zip(*[iter(iterable)]*n)
def submit_train(mdl,data_frame,output_dir,data,epochs=50,cyclic_epochs=0,save_name='selfsup_mdl',training_batch_size=32):
print('training start: {}'.format(datetime.now().strftime('%Y-%m-%d-%H%M')))
num_classes = mdl.output_shape[-1]
num_classes = None if num_classes <= 1 else num_classes
fpi_args = {'num_classes':num_classes,
'core_percent':0.5 if 'brain' in data else 0.8,
'tolerance': None if 'brain' in data else 1E-3
}
elem_in_epoch = len(data_frame.file_list)
if cyclic_epochs>0:
half_cycle_len = elem_in_epoch//4
lr_min = 1E-4
lr_max = 1E-1
half1 = np.linspace(lr_min,lr_max,half_cycle_len)
half2 = np.linspace(lr_max,lr_min,half_cycle_len)
lr_cycle = np.concatenate((half1,half2),0)
for epoch_i in range(epochs+cyclic_epochs):
if epoch_i>epochs and elem_i < len(lr_cycle):
#cyclic training portion, adjust learning rate
tf.keras.backend.set_value(mdl.optimizer.lr, lr_cycle[elem_i])
#get subjects in pairs for mixing
for batch_in,batch_in2 in grouped(data_frame.tf_dataset,2):
#apply fpi on batch
pex1,pex2 = self_sup_task.patch_ex(batch_in,batch_in2,**fpi_args)
ind_sampler = index_sampling(len(pex1[0]))#randomize slices in batch
for _ in range(len(pex1[0])//training_batch_size):
cur_inds = ind_sampler.get_inds(training_batch_size)
train_step(mdl,tf.gather(pex1[0],cur_inds),tf.gather(pex1[1],cur_inds))
train_step(mdl,tf.gather(pex2[0],cur_inds),tf.gather(pex2[1],cur_inds))
print('epoch {}: {}'.format(str(epoch_i),datetime.now().strftime('%Y-%m-%d-%H%M')))
#measure loss
for batch_in,batch_in2 in grouped(data_frame.tf_dataset,2):
break
pex1,pex2 = self_sup_task.patch_ex(batch_in,batch_in2,**fpi_args)
avg_loss = []
ind_sampler = index_sampling(len(pex1[0]))#randomize slices in batch
for _ in range(len(pex1[0])//training_batch_size):
cur_inds = ind_sampler.get_inds(training_batch_size)
avg_loss.append(test_step(mdl,tf.gather(pex1[0],cur_inds),tf.gather(pex1[1],cur_inds)))
avg_loss.append(test_step(mdl,tf.gather(pex2[0],cur_inds),tf.gather(pex2[1],cur_inds)))
avg_loss = np.mean(avg_loss)
print('Avg loss: {}'.format(avg_loss))
if epoch_i == 0:
best_loss = avg_loss
elif avg_loss < best_loss:
best_loss = avg_loss
print('new best loss')
save_model(mdl,output_dir,save_name+'_bestLoss',time_stamp=False)
if epoch_i % 10 == 0 or epoch_i>epochs:
#save every 10 epochs or every epoch in cyclic mode
save_model(mdl,output_dir,save_name)
#save final model
save_model(mdl,output_dir,save_name+'_final')
return
def submit_test(mdl,data_frame,output_dir,mode,batch_size=1,save_name='selfsup_mdl'):
print('testing start: {}'.format(datetime.now().strftime('%Y-%m-%d-%H%M')))
nii_file = 0
for batch_in in data_frame.tf_dataset:
#predict for subject
pred = np.zeros(np.shape(batch_in))
for ind in range(len(batch_in)//batch_size):
pred[ind:(ind+1)*batch_size] = pred_step(mdl,batch_in[ind:(ind+1)*batch_size])
output_chan = np.shape(pred)[-1]
if output_chan > 1:
pred *= np.arange(output_chan)/(output_chan-1)
pred = np.sum(pred,-1,keepdims=True)
#save output as nifti and label with label suffix
#print(data_frame.file_list[0])#only data, not label names
fname_i = data_frame.file_list[nii_file].split('/')[-1]
if 'sample' in mode:
#derive subject-level score
im_level_score = np.mean(pred,axis=(1,2,3))
window_size = int((len(im_level_score)*0.1)//2)*2+1#take 10% sliding filter window
im_level_score_f = savgol_filter(im_level_score,window_size,3)#order 3 polynomial
im_level_score_s = sorted(im_level_score_f)
im_level_score_s = im_level_score_s[int(len(im_level_score_s)*0.75):]
sample_score = np.mean(im_level_score_s)#mean of top quartile values
with open(os.path.join(output_dir,fname_i + ".txt"), "w") as write_file:
write_file.write(str(sample_score))
if 'pixel' in mode:
data_frame.save_nii(pred,output_dir,fname_i)
nii_file += 1
return
def save_model(mdl,results_dir,fname,time_stamp=True):
#save model
if time_stamp:
#mdl_weights_name = fname+'_{}_weights.h5'.format(datetime.now().strftime('%Y-%m-%d-%H%M'))
mdl_weights_name = fname+'_{}_weights'.format(datetime.now().strftime('%Y-%m-%d-%H%M'))
else:
#mdl_weights_name = fname+'_weights.h5'
mdl_weights_name = fname+'_weights'
mdl_weights_path = os.path.join(results_dir, mdl_weights_name)
mdl.save(mdl_weights_path)
return
class index_sampling(object):
def __init__(self,total_len):
self.total_len = total_len
self.ind_generator = rand_ind_fisheryates(self.total_len)
def get_inds(self,batch_size):
cur_inds = list(itertools.islice(self.ind_generator,batch_size))
if len(cur_inds) < batch_size:
#end of iterator - reset/shuffle
self.ind_generator = rand_ind_fisheryates(self.total_len)
cur_inds = list(itertools.islice(self.ind_generator,batch_size))
return cur_inds
def reset():
self.ind_generator = rand_ind_fisheryates(self.total_len)
return
def rand_ind_fisheryates(num_inds):
numbers=np.arange(num_inds,dtype=np.uint32)
for ind_i in range(num_inds):
j=np.random.randint(ind_i,num_inds)
numbers[ind_i],numbers[j]=numbers[j],numbers[ind_i]
yield numbers[ind_i]