test_data_path = Path(TEST_DATA_PATH)
IMG_SIZE = 512
BS = int(os.environ.get('BS', 20))
NUM_EPOCHS = int(os.environ.get('NUM_EPOCHS', 16))
DCM_CONF = json.loads(os.environ['DCM_CONF'])
open_dcm_image_func = simple_open_dcm_image_factory(DCM_CONF)
fastai.vision.image.open_image = open_dcm_image_func
fastai.vision.data.open_image = open_dcm_image_func
open_image = open_dcm_image_func
# TRAIN
scans = get_scans(data_path)
validation_patients = []
print('Validation patients: ', validation_patients)
data = get_data(scans,
HOME_PATH,
validation_patients,
normalize_stats=DCM_CONF[NORM_STATS],
bs=BS)
learn = get_learner(data, metrics=[], model_save_path=MODEL_SAVE_PATH)
learn.unfreeze()
learn.fit_one_cycle(NUM_EPOCHS, 1e-4)
learn.save(MODEL_NAME)
NUM_EPOCHS = int(os.environ.get('NUM_EPOCHS', 16))
DCM_LOAD_FUNC = {
SIMPLE_WINDOW_SMALL: simple_open_dcm_image_factory,
SIMPLE_MULTIPLE_WINDOWS: simple_open_dcm_image_factory,
FREQS_NO_LIMIT_WINDOWS: freqs_open_dcm_image_factory,
FREQS_LIMIT_WINDOWS: freqs_open_dcm_image_factory,
SIMPLE_WINDOW_1CHANNEL_CLASSIC_UNET: simple_open_dcm_image_factory,
SIMPLE_WINDOW_3CHANNEL_CLASSIC_UNET: simple_open_dcm_image_factory,
}
# CV
scans = get_scans(data_path)
folds_df = pd.read_csv(FOLDS_PATH, encoding='utf-8')
results = []
by_patient_results = []
for idx, fold in folds_df.iterrows():
conf = json.loads(fold[EXPERIMENT_NAME])
open_dcm_image_func = DCM_LOAD_FUNC[EXPERIMENT_NAME](conf)
fastai.vision.image.open_image = open_dcm_image_func
fastai.vision.data.open_image = open_dcm_image_func
open_image = open_dcm_image_func
validation_patients = json.loads(fold[VALIDATION_PATIENTS])
print('Validation patients: ', validation_patients)
training_patients = [p for p in patients if p not in VALIDATION_PATIENTS]
print(
f'patients: {len(patients)}, training: {len(training_patients)}, validation: {len(VALIDATION_PATIENTS)}'
)
print(f'validation patients: {VALIDATION_PATIENTS}')
results = []
for sample_size in SAMPLE_SIZES:
print(f'Sample size: {sample_size}')
for sampling_round in range(SAMPLING_ROUNDS):
sampled_patients = sample(training_patients, sample_size)
print(f'{sampling_round} sampled patients: {sampled_patients}')
scans = get_scans(data_path,
patients=sampled_patients + VALIDATION_PATIENTS)
data = get_data(
scans,
HOME_PATH,
VALIDATION_PATIENTS,
normalize_stats=DCM_CONF[NORM_STATS],
bs=BS,
)
learn = get_learner(data,
metrics=[dice],
model_save_path=MODEL_SAVE_PATH)
learn.unfreeze()
learn.fit_one_cycle(16, 1e-4)
return {WINDOWS: windows, NORM_STATS: (means.tolist(), stds.tolist())}
folds = []
kfold = KFold(N_FOLDS, shuffle=True, random_state=SEED)
for train_index, val_index in kfold.split(patients):
result = {}
train_patients = patients[train_index]
val_patients = patients[val_index]
print(val_patients)
result[TRAIN_PATIENTS] = json.dumps(train_patients.tolist())
result[VALIDATION_PATIENTS] = json.dumps(val_patients.tolist())
scans = get_scans(data_path, patients=train_patients)
# freqs
result[FREQS_NO_LIMIT_WINDOWS] = json.dumps(get_freqs_method_dict(scans, None))
result[FREQS_LIMIT_WINDOWS] = json.dumps(
get_freqs_method_dict(scans, STANDARD_WINDOWS[0])
)
# simple
result[SIMPLE_WINDOW_SMALL] = json.dumps(
get_standard_method_dict(scans, STANDARD_WINDOWS)
)
result[SIMPLE_MULTIPLE_WINDOWS] = json.dumps(
get_standard_method_dict(scans, EXTENDED_WINDOWS)
)