-
Notifications
You must be signed in to change notification settings - Fork 40
/
dsb_a_eliasx37_relias18_s5_p8a1.py
316 lines (233 loc) · 11.3 KB
/
dsb_a_eliasx37_relias18_s5_p8a1.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
import numpy as np
import data_transforms
import data_iterators
import pathfinder
import lasagne as nn
import nn_lung
from collections import namedtuple
from functools import partial
import lasagne.layers.dnn as dnn
import theano.tensor as T
import utils
import utils_lung
import os
from configs_luna_props_patch import r_elias_18 as cfg_prop
# TODO: import correct config here
candidates_config = 'dsb_relias18_s5_p8a1'
restart_from_save = None
rng = np.random.RandomState(42)
predictions_dir = utils.get_dir_path('model-predictions', pathfinder.METADATA_PATH)
candidates_path = predictions_dir + '/%s' % candidates_config
id2candidates_path = utils_lung.get_candidates_paths(candidates_path)
# transformations
p_transform = {'patch_size': (48, 48, 48),
'mm_patch_size': (48, 48, 48),
'pixel_spacing': (1., 1., 1.)
}
p_transform_augment = {
'translation_range_z': [-5, 5],
'translation_range_y': [-5, 5],
'translation_range_x': [-5, 5],
'rotation_range_z': [-10, 10],
'rotation_range_y': [-10, 10],
'rotation_range_x': [-10, 10]
}
n_candidates_per_patient = 8
def data_prep_function(data, patch_centers, pixel_spacing, p_transform,
p_transform_augment, **kwargs):
x = data_transforms.transform_dsb_candidates(data=data,
patch_centers=patch_centers,
p_transform=p_transform,
p_transform_augment=p_transform_augment,
pixel_spacing=pixel_spacing)
x = data_transforms.hu2normHU(x)
return x
def get_feature_dict(feature_vector):
d_feat = {}
unit_ptr = 0
for obj_idx, obj_name in enumerate(cfg_prop.order_objectives):
ptype = cfg_prop.property_type[obj_name]
if ptype == 'classification':
num_units = len(cfg_prop.property_bin_borders[obj_name])
d_feat[obj_name] = feature_vector[unit_ptr:unit_ptr+num_units]
unit_ptr += num_units
elif ptype == 'continuous':
d_feat[obj_name] = feature_vector[unit_ptr]
unit_ptr += 1
return d_feat
def logodds2p(lo):
if lo < -500:
# prevent underflow
return 0.
elif lo > 500:
# prevent overflow
return 1.
else:
odds = np.exp(lo)
p = odds / (1+odds)
return p
def candidates_prep_function(all_candidates, n_selection):
candidates_w_svalue = []
for candidate in all_candidates:
# print 'candidate', candidate
feature_vector = candidate[4:]
d_feat = get_feature_dict(feature_vector)
# print d_feat
sorting_value = d_feat['nodule'][1]
can_w_svalue = np.concatenate((candidate[:3], [sorting_value]))
# print can_w_svalue
candidates_w_svalue.append(can_w_svalue)
a = np.asarray(sorted(candidates_w_svalue, key=lambda x: x[-1], reverse=True))
a = a[:n_selection]
return a
data_prep_function_train = partial(data_prep_function, p_transform_augment=p_transform_augment,
p_transform=p_transform)
data_prep_function_valid = partial(data_prep_function, p_transform_augment=None,
p_transform=p_transform)
# data iterators
batch_size = 1
train_valid_ids = utils.load_pkl(pathfinder.VALIDATION_SPLIT_PATH)
train_pids, valid_pids, test_pids = train_valid_ids['training'], train_valid_ids['validation'], train_valid_ids['test']
print 'n train', len(train_pids)
print 'n valid', len(valid_pids)
print 'n test', len(test_pids)
train_data_iterator = data_iterators.DSBPatientsDataGenerator(data_path=pathfinder.DATA_PATH,
batch_size=batch_size,
transform_params=p_transform,
n_candidates_per_patient=n_candidates_per_patient,
data_prep_fun=data_prep_function_train,
candidates_prep_fun = candidates_prep_function,
id2candidates_path=id2candidates_path,
rng=rng,
patient_ids=train_pids,
random=True, infinite=True)
valid_data_iterator = data_iterators.DSBPatientsDataGenerator(data_path=pathfinder.DATA_PATH,
batch_size=1,
transform_params=p_transform,
n_candidates_per_patient=n_candidates_per_patient,
data_prep_fun=data_prep_function_valid,
candidates_prep_fun = candidates_prep_function,
id2candidates_path=id2candidates_path,
rng=rng,
patient_ids=valid_pids,
random=False, infinite=False)
test_data_iterator = data_iterators.DSBPatientsDataGenerator(data_path=pathfinder.DATA_PATH,
batch_size=1,
transform_params=p_transform,
n_candidates_per_patient=n_candidates_per_patient,
data_prep_fun=data_prep_function_valid,
candidates_prep_fun = candidates_prep_function,
id2candidates_path=id2candidates_path,
rng=rng,
patient_ids=test_pids,
random=False, infinite=False)
nchunks_per_epoch = train_data_iterator.nsamples / batch_size
max_nchunks = nchunks_per_epoch * 8
validate_every = int(0.5 * nchunks_per_epoch)
save_every = int(0.5 * nchunks_per_epoch)
learning_rate_schedule = {
0: 1e-4,
int(5 * nchunks_per_epoch): 5e-5,
int(6 * nchunks_per_epoch): 1e-5,
int(7 * nchunks_per_epoch): 5e-6,
int(9 * nchunks_per_epoch): 1e-6
}
# model
conv3d = partial(dnn.Conv3DDNNLayer,
filter_size=3,
pad='same',
W=nn.init.Orthogonal(),
nonlinearity=nn.nonlinearities.very_leaky_rectify)
max_pool3d = partial(dnn.MaxPool3DDNNLayer,
pool_size=2)
drop = nn.layers.DropoutLayer
dense = partial(nn.layers.DenseLayer,
W=nn.init.Orthogonal(),
nonlinearity=nn.nonlinearities.very_leaky_rectify)
def inrn_v2(lin):
n_base_filter = 32
l1 = conv3d(lin, n_base_filter, filter_size=1)
l2 = conv3d(lin, n_base_filter, filter_size=1)
l2 = conv3d(l2, n_base_filter, filter_size=3)
l3 = conv3d(lin, n_base_filter, filter_size=1)
l3 = conv3d(l3, n_base_filter, filter_size=3)
l3 = conv3d(l3, n_base_filter, filter_size=3)
l = nn.layers.ConcatLayer([l1, l2, l3])
l = conv3d(l, lin.output_shape[1], filter_size=1)
l = nn.layers.ElemwiseSumLayer([l, lin])
l = nn.layers.NonlinearityLayer(l, nonlinearity=nn.nonlinearities.rectify)
return l
def inrn_v2_red(lin):
# We want to reduce our total volume /4
den = 16
nom2 = 4
nom3 = 5
nom4 = 7
ins = lin.output_shape[1]
l1 = max_pool3d(lin)
l2 = conv3d(lin, ins // den * nom2, filter_size=3, stride=2)
l3 = conv3d(lin, ins // den * nom2, filter_size=1)
l3 = conv3d(l3, ins // den * nom3, filter_size=3, stride=2)
l4 = conv3d(lin, ins // den * nom2, filter_size=1)
l4 = conv3d(l4, ins // den * nom3, filter_size=3)
l4 = conv3d(l4, ins // den * nom4, filter_size=3, stride=2)
l = nn.layers.ConcatLayer([l1, l2, l3, l4])
return l
def feat_red(lin):
# We want to reduce the feature maps by a factor of 2
ins = lin.output_shape[1]
l = conv3d(lin, ins // 2, filter_size=1)
return l
def load_pretrained_model(l_in):
l = conv3d(l_in, 64)
l = inrn_v2_red(l)
l = inrn_v2(l)
l = inrn_v2_red(l)
l = inrn_v2(l)
l = inrn_v2_red(l)
l = inrn_v2_red(l)
l = drop(l, name='can_dropout')
l = dense(l, 512, name='can_dense')
final_layers = []
for obj_idx, obj_name in enumerate(cfg_prop.order_objectives):
ptype = cfg_prop.property_type[obj_name]
if ptype == 'classification':
num_units = len(cfg_prop.property_bin_borders[obj_name])
l_fin = nn.layers.DenseLayer(l, num_units=num_units,
W=nn.init.Orthogonal(),
b=nn.init.Constant(cfg_prop.init_values_final_units[obj_name]),
nonlinearity=nn.nonlinearities.softmax, name='dense_'+ptype+'_'+obj_name)
elif ptype == 'continuous':
l_fin = nn.layers.DenseLayer(l, num_units=1,
W=nn.init.Orthogonal(),
b=nn.init.Constant(cfg_prop.init_values_final_units[obj_name]),
nonlinearity=nn.nonlinearities.softplus, name='dense_'+ptype+'_'+obj_name)
else:
raise
final_layers.append(l_fin)
l_out = nn.layers.ConcatLayer(final_layers, name = 'final_concat_layer')
metadata = utils.load_pkl(os.path.join("/home/eavsteen/dsb3/storage/metadata/dsb3/models/eavsteen/","r_elias_18-20170329-182238.pkl"))
nn.layers.set_all_param_values(l_out, metadata['param_values'])
features = nn.layers.get_all_layers(l_out)[(-2-len(final_layers))]
print 'features layer', features.name
return features
def build_model():
l_in = nn.layers.InputLayer((None, n_candidates_per_patient, ) + p_transform['patch_size'])
l_in_rshp = nn.layers.ReshapeLayer(l_in, (-1, 1,) + p_transform['patch_size'])
l_target = nn.layers.InputLayer((batch_size,))
penultimate_layer = load_pretrained_model(l_in_rshp)
l = nn.layers.DenseLayer(penultimate_layer, num_units=1, W=nn.init.Orthogonal(),
nonlinearity=nn.nonlinearities.sigmoid, name='dense_p_benign')
l = nn.layers.ReshapeLayer(l, (-1, n_candidates_per_patient, 1), name='reshape2patients')
l_out = nn_lung.LogMeanExp(l, r=8, axis=(1, 2), name='LME')
return namedtuple('Model', ['l_in', 'l_out', 'l_target'])(l_in, l_out, l_target)
def build_objective(model, deterministic=False, epsilon=1e-12):
p = nn.layers.get_output(model.l_out, deterministic=deterministic)
targets = T.flatten(nn.layers.get_output(model.l_target))
p = T.clip(p, epsilon, 1.-epsilon)
bce = T.nnet.binary_crossentropy(p, targets)
return T.mean(bce)
def build_updates(train_loss, model, learning_rate):
params = nn.layers.get_all_params(model.l_out, trainable=True)
updates = nn.updates.adam(train_loss, params, learning_rate)
return updates