|
a |
|
b/main.py |
|
|
1 |
#main.py |
|
|
2 |
|
|
|
3 |
import timeit |
|
|
4 |
|
|
|
5 |
from run_experiment import DukeCTModel |
|
|
6 |
from models import custom_models_ctnet, custom_models_alternative, custom_models_ablation |
|
|
7 |
from load_dataset import custom_datasets |
|
|
8 |
|
|
|
9 |
#Note that here NUM_EPOCHS is set to 2 for the purposes of quickly demonstrating |
|
|
10 |
#the code on the fake data. In all of the experiments reported in the paper, |
|
|
11 |
#NUM_EPOCHS was set to 100. No model actually trained all the way to 100 epochs |
|
|
12 |
#due to use of early stopping. |
|
|
13 |
NUM_EPOCHS = 2 |
|
|
14 |
|
|
|
15 |
if __name__=='__main__': |
|
|
16 |
#################################### |
|
|
17 |
# CTNet-83 Model on Whole Data Set #---------------------------------------- |
|
|
18 |
#################################### |
|
|
19 |
tot0 = timeit.default_timer() |
|
|
20 |
DukeCTModel(descriptor = 'CTNet83', |
|
|
21 |
custom_net = custom_models_ctnet.CTNetModel, |
|
|
22 |
custom_net_args = {'n_outputs':83}, |
|
|
23 |
loss = 'bce', loss_args = {}, |
|
|
24 |
num_epochs=NUM_EPOCHS, patience = 15, |
|
|
25 |
batch_size = 2, device = 'all', data_parallel = True, |
|
|
26 |
use_test_set = False, task = 'train_eval', |
|
|
27 |
old_params_dir = '', |
|
|
28 |
dataset_class = custom_datasets.CTDataset_2019_10, |
|
|
29 |
dataset_args = {'label_type_ld':'disease_new', |
|
|
30 |
'label_meanings':'all', |
|
|
31 |
'num_channels':3, |
|
|
32 |
'pixel_bounds':[-1000,200], |
|
|
33 |
'data_augment':True, |
|
|
34 |
'crop_type':'single', |
|
|
35 |
'selected_note_acc_files':{'train':'','valid':''}}) |
|
|
36 |
tot1 = timeit.default_timer() |
|
|
37 |
print('Total Time', round((tot1 - tot0)/60.0,2),'minutes') |
|
|
38 |
|
|
|
39 |
################################### |
|
|
40 |
# CTNet-9 Model on Whole Data Set #----------------------------------------- |
|
|
41 |
################################### |
|
|
42 |
tot0 = timeit.default_timer() |
|
|
43 |
DukeCTModel(descriptor = 'CTNet9', |
|
|
44 |
custom_net = custom_models_ctnet.CTNetModel, |
|
|
45 |
custom_net_args = {'n_outputs':9}, |
|
|
46 |
loss = 'bce', loss_args = {}, |
|
|
47 |
num_epochs=NUM_EPOCHS, patience = 15, |
|
|
48 |
batch_size = 2, device = 'all', data_parallel = True, |
|
|
49 |
use_test_set = False, task = 'train_eval', |
|
|
50 |
old_params_dir = '', |
|
|
51 |
dataset_class = custom_datasets.CTDataset_2019_10, |
|
|
52 |
dataset_args = {'label_type_ld':'disease_new', |
|
|
53 |
'label_meanings':['nodule','opacity','atelectasis','pleural_effusion','consolidation','mass','pericardial_effusion','cardiomegaly','pneumothorax'], |
|
|
54 |
'num_channels':3, |
|
|
55 |
'pixel_bounds':[-1000,200], |
|
|
56 |
'data_augment':True, |
|
|
57 |
'crop_type':'single', |
|
|
58 |
'selected_note_acc_files':{'train':'','valid':''}}) |
|
|
59 |
tot1 = timeit.default_timer() |
|
|
60 |
print('Total Time', round((tot1 - tot0)/60.0,2),'minutes') |
|
|
61 |
|
|
|
62 |
#################################################### |
|
|
63 |
# CTNet-83 Model on 2000 Train and 1000 Val Subset #------------------------ |
|
|
64 |
#################################################### |
|
|
65 |
tot0 = timeit.default_timer() |
|
|
66 |
DukeCTModel(descriptor = 'CTNet83_SmallData', |
|
|
67 |
custom_net = custom_models_ctnet.CTNetModel, |
|
|
68 |
custom_net_args = {'n_outputs':83}, |
|
|
69 |
loss = 'bce', loss_args = {}, |
|
|
70 |
num_epochs=NUM_EPOCHS, patience = 15, |
|
|
71 |
batch_size = 2, device = 'all', data_parallel = True, |
|
|
72 |
use_test_set = False, task = 'train_eval', |
|
|
73 |
old_params_dir = '', |
|
|
74 |
dataset_class = custom_datasets.CTDataset_2019_10, |
|
|
75 |
dataset_args = {'label_type_ld':'disease_new', |
|
|
76 |
'label_meanings':'all', |
|
|
77 |
'num_channels':3, |
|
|
78 |
'pixel_bounds':[-1000,200], |
|
|
79 |
'data_augment':True, |
|
|
80 |
'crop_type':'single', |
|
|
81 |
'selected_note_acc_files':{'train':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgtrain_random2000.csv', |
|
|
82 |
'valid':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgvalid_a_random1000.csv'}}) |
|
|
83 |
tot1 = timeit.default_timer() |
|
|
84 |
print('Total Time', round((tot1 - tot0)/60.0,2),'minutes') |
|
|
85 |
|
|
|
86 |
###################################################################### |
|
|
87 |
# Alternative Arch: BodyConv Model on 2000 Train and 1000 Val Subset #------ |
|
|
88 |
###################################################################### |
|
|
89 |
tot0 = timeit.default_timer() |
|
|
90 |
DukeCTModel(descriptor = 'BodyConv_SmallData', |
|
|
91 |
custom_net = custom_models_alternative.BodyConv, |
|
|
92 |
custom_net_args = {'n_outputs':83}, |
|
|
93 |
loss = 'bce', loss_args = {}, |
|
|
94 |
num_epochs=NUM_EPOCHS, patience = 15, |
|
|
95 |
batch_size = 2, device = 'all', data_parallel = True, |
|
|
96 |
use_test_set = False, task = 'train_eval', |
|
|
97 |
old_params_dir = '', |
|
|
98 |
dataset_class = custom_datasets.CTDataset_2019_10, |
|
|
99 |
dataset_args = {'label_type_ld':'disease_new', |
|
|
100 |
'label_meanings':'all', |
|
|
101 |
'num_channels':3, |
|
|
102 |
'pixel_bounds':[-1000,200], |
|
|
103 |
'data_augment':True, |
|
|
104 |
'crop_type':'single', |
|
|
105 |
'selected_note_acc_files':{'train':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgtrain_random2000.csv', |
|
|
106 |
'valid':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgvalid_a_random1000.csv'}}) |
|
|
107 |
tot1 = timeit.default_timer() |
|
|
108 |
print('Total Time', round((tot1 - tot0)/60.0,2),'minutes') |
|
|
109 |
|
|
|
110 |
#################################################################### |
|
|
111 |
# Alternative Arch: 3DConv Model on 2000 Train and 1000 Val Subset #-------- |
|
|
112 |
#################################################################### |
|
|
113 |
tot0 = timeit.default_timer() |
|
|
114 |
DukeCTModel(descriptor = 'ThreeDConv_SmallData', |
|
|
115 |
custom_net = custom_models_alternative.ThreeDConv, |
|
|
116 |
custom_net_args = {'n_outputs':83}, |
|
|
117 |
loss = 'bce', loss_args = {}, |
|
|
118 |
num_epochs=NUM_EPOCHS, patience = 15, |
|
|
119 |
batch_size = 4, device = 'all', data_parallel = True, |
|
|
120 |
use_test_set = False, task = 'train_eval', |
|
|
121 |
old_params_dir = '', |
|
|
122 |
dataset_class = custom_datasets.CTDataset_2019_10, |
|
|
123 |
dataset_args = {'label_type_ld':'disease_new', |
|
|
124 |
'label_meanings':'all', |
|
|
125 |
'num_channels':1, |
|
|
126 |
'pixel_bounds':[-1000,200], |
|
|
127 |
'data_augment':True, |
|
|
128 |
'crop_type':'single', |
|
|
129 |
'selected_note_acc_files':{'train':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgtrain_random2000.csv', |
|
|
130 |
'valid':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgvalid_a_random1000.csv'}}) |
|
|
131 |
tot1 = timeit.default_timer() |
|
|
132 |
print('Total Time', round((tot1 - tot0)/60.0,2),'minutes') |
|
|
133 |
|
|
|
134 |
#################################################################### |
|
|
135 |
# Ablation Study: CTNet-83 (Pool) on 2000Train and 1000 Val Subset #-------- |
|
|
136 |
#################################################################### |
|
|
137 |
tot0 = timeit.default_timer() |
|
|
138 |
DukeCTModel(descriptor = 'CTNet83AblatePool_SmallData', |
|
|
139 |
custom_net = custom_models_ablation.CTNetModel_Ablate_PoolInsteadOf3D, |
|
|
140 |
custom_net_args = {'n_outputs':83}, |
|
|
141 |
loss = 'bce', loss_args = {}, |
|
|
142 |
num_epochs=NUM_EPOCHS, patience = 15, |
|
|
143 |
batch_size = 2, device = 'all', data_parallel = True, |
|
|
144 |
use_test_set = False, task = 'train_eval', |
|
|
145 |
old_params_dir = '', |
|
|
146 |
dataset_class = custom_datasets.CTDataset_2019_10, |
|
|
147 |
dataset_args = {'label_type_ld':'disease_new', |
|
|
148 |
'label_meanings':'all', |
|
|
149 |
'num_channels':3, |
|
|
150 |
'pixel_bounds':[-1000,200], |
|
|
151 |
'data_augment':True, |
|
|
152 |
'crop_type':'single', |
|
|
153 |
'selected_note_acc_files':{'train':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgtrain_random2000.csv', |
|
|
154 |
'valid':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgvalid_a_random1000.csv'}}) |
|
|
155 |
tot1 = timeit.default_timer() |
|
|
156 |
print('Total Time', round((tot1 - tot0)/60.0,2),'minutes') |
|
|
157 |
|
|
|
158 |
##################################################################### |
|
|
159 |
# Ablation Study: CTNet-83 (Rand) on 2000 Train and 1000 Val Subset #------- |
|
|
160 |
##################################################################### |
|
|
161 |
tot0 = timeit.default_timer() |
|
|
162 |
DukeCTModel(descriptor = 'CTNet83AblateRand_SmallData', |
|
|
163 |
custom_net = custom_models_ablation.CTNetModel_Ablate_RandomInitResNet, |
|
|
164 |
custom_net_args = {'n_outputs':83}, |
|
|
165 |
loss = 'bce', loss_args = {}, |
|
|
166 |
num_epochs=NUM_EPOCHS, patience = 15, |
|
|
167 |
batch_size = 2, device = 'all', data_parallel = True, |
|
|
168 |
use_test_set = False, task = 'train_eval', |
|
|
169 |
old_params_dir = '', |
|
|
170 |
dataset_class = custom_datasets.CTDataset_2019_10, |
|
|
171 |
dataset_args = {'label_type_ld':'disease_new', |
|
|
172 |
'label_meanings':'all', |
|
|
173 |
'num_channels':3, |
|
|
174 |
'pixel_bounds':[-1000,200], |
|
|
175 |
'data_augment':True, |
|
|
176 |
'crop_type':'single', |
|
|
177 |
'selected_note_acc_files':{'train':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgtrain_random2000.csv', |
|
|
178 |
'valid':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgvalid_a_random1000.csv'}}) |
|
|
179 |
tot1 = timeit.default_timer() |
|
|
180 |
print('Total Time', round((tot1 - tot0)/60.0,2),'minutes') |
|
|
181 |
|
|
|
182 |
################################################### |
|
|
183 |
# CTNet-1 Model on 2000 Train and 1000 Val Subset #------------------------- |
|
|
184 |
################################################### |
|
|
185 |
for abnormality in ['nodule', 'opacity', 'atelectasis', 'pleural_effusion', |
|
|
186 |
'consolidation', 'mass', 'pericardial_effusion', |
|
|
187 |
'cardiomegaly', 'pneumothorax']: |
|
|
188 |
print('\n\n\n\n********** Working on',abnormality,'**********') |
|
|
189 |
tot0 = timeit.default_timer() |
|
|
190 |
DukeCTModel(descriptor = 'CTNet-'+abnormality, |
|
|
191 |
custom_net = custom_models_ctnet.CTNetModel, |
|
|
192 |
custom_net_args = {'n_outputs':1}, |
|
|
193 |
loss = 'bce', loss_args = {}, |
|
|
194 |
num_epochs=NUM_EPOCHS, patience = 15, |
|
|
195 |
batch_size = 2, device = 'all', data_parallel = True, |
|
|
196 |
use_test_set = False, task = 'train_eval', |
|
|
197 |
old_params_dir = '', |
|
|
198 |
dataset_class = custom_datasets.CTDataset_2019_10, |
|
|
199 |
dataset_args = {'label_type_ld':'disease_new', |
|
|
200 |
'label_meanings':[abnormality], #can be 'all' or a list of strings |
|
|
201 |
'num_channels':3, |
|
|
202 |
'pixel_bounds':[-1000,200], |
|
|
203 |
'data_augment':True, |
|
|
204 |
'crop_type':'single', |
|
|
205 |
'selected_note_acc_files':{'train':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgtrain_random2000.csv', |
|
|
206 |
'valid':'/load_dataset/fakedata/predefined_subsets/2020-01-10-imgvalid_a_random1000.csv'}} |
|
|
207 |
) |
|
|
208 |
tot1 = timeit.default_timer() |
|
|
209 |
print('Total Time', round((tot1 - tot0)/60.0,2),'minutes') |
|
|
210 |
|