def get_exam_aggregator(name):
if not name in AGGREGATOR_REGISTRY:
error_msg = NO_AGGREGATOR_ERR.format(name, AGGREGATOR_REGISTRY.keys())
logger.error(error_msg)
raise Exception(error_msg)
logger.info(AGGREGATOR_SUCCESS_MSG.format(name))
return AGGREGATOR_REGISTRY[name]
def get_risk_factors(args, ssn, exam, risk_metadata_json, json_dir, logger):
'''
args:
- args:
- ssn:
- exam:
- json_dir:
- logger:
returns:
- risk_factor_vector:
'''
try:
os.makedirs(json_dir)
except Exception as e:
pass
args.metadata_path = "{}.json".format(
os.path.join(json_dir, str(uuid.uuid4())))
args.risk_factor_metadata_path = "{}.json".format(
os.path.join(json_dir, str(uuid.uuid4())))
# Write current request to a file to use as a metadata path
prior_hist = risk_metadata_json[ssn]['any_breast_cancer'] == 1
metadata_json = [{
'ssn':
ssn,
'accessions': [{
'accession': exam,
'prior_hist': prior_hist
}]
}]
try:
json.dump(metadata_json, open(args.metadata_path, 'w'))
json.dump(risk_metadata_json, open(args.risk_factor_metadata_path,
'w'))
except Exception as e:
delete_jsons(args)
err_msg = FAIL_TO_SAVE_METADATA_MESSAGE.format(ssn, exam, e, args)
logger.error(err_msg)
raise Exception(err_msg)
# Load risk factor vector from metadata file and del metadata json
try:
risk_factor_vectorizer = RiskFactorVectorizer(args)
sample = {'ssn': ssn, 'exam': exam}
risk_factor_vector = risk_factor_vectorizer.get_risk_factors_for_sample(
sample)
logger.info(SUCCESS_RISK_VEC_MESSAGE.format(ssn, exam, args))
delete_jsons(args)
return risk_factor_vector
except Exception as e:
delete_jsons(args)
err_msg = FAIL_TO_GET_RISK_VECTOR_MESSAGE.format(ssn, exam, e, args)
logger.error(err_msg)
raise Exception(err_msg)
def serve():
'''
API to serve a model from OncoNet
Takes a list of dicom files, and a list of optional keys
i.e { dicoms: [bytes, bytes, bytes], data: { optional metadata} }
and returns:
{ prediction: Y, metadata: {}, model_name : NAME,
oncoserve_version: X.X.X, onconet_version: X.X.X,
oncodata_version: X.X.X}
Prediction is the exam level prediction over dicoms.
Model_name defines what predictor is running: i.e density, risk etc.
Metadata is meant to contain things like MRN, ACCESSION and any
additional metadata for object tracking.
OncoServe_version is the version of the model deployment framework
OncoNet_version is the version of the research model framework
OncoData_version is the version of the dicom conversion framework
The configuration of the model used to produce Y is set in the app configuration. See config.py for config objects.
'''
logger.info("Serving request...")
response = {
'model_name': app.config['NAME'],
'oncoserve_version': app.config['ONCOSERVE_VERSION'],
'onconet_version': app.config['ONCONET_VERSION'],
'oncodata_version': app.config['ONCODATA_VERSION'],
'log_file': LOG_FILE
}
try:
dicoms = request.files.getlist('dicom')
metadata = request.form
response['metadata'] = metadata
images = oncodata_wrapper.get_pngs(dicoms, oncodata_args, logger)
logger.info(ONCODATA_SUCCESS_MSG)
if onconet_args.use_risk_factors:
assert 'mrn' in metadata
assert 'accession' in metadata
risk_factors = json.loads(
request.files.getlist('risk_factors')[0].read())
risk_factor_vector = oncoqueries_wrapper.get_risk_factors(
onconet_args, metadata['mrn'], metadata['accession'],
risk_factors, oncodata_args.temp_img_dir, logger)
logger.info(ONCOQUERIES_SUCCESS_MSG)
else:
risk_factor_vector = None
y = onconet.process_exam(images, risk_factor_vector)
logger.info(ONCONET_SUCCESS_MSG)
msg = 'OK'
response['prediction'] = y
response['msg'] = msg
return jsonify(response), HTTP_200_OK
except Exception as e:
msg = ONCOSERVE_FAIL_MSG.format(str(e))
response['prediction'] = None
response['msg'] = msg
return jsonify(response), HTTP_500_INTERNAL_SERVER_ERROR
def __init__(self, args, aggregator_name, logger):
logger.info(INIT_MESSAGE)
self.args = args
args.cuda = args.cuda and torch.cuda.is_available()
args.test_image_transformers = parsing.parse_transformers(
args.test_image_transformers)
args.test_tensor_transformers = parsing.parse_transformers(
args.test_tensor_transformers)
test_transformers = transformer_factory.get_transformers(
args.test_image_transformers, args.test_tensor_transformers, args)
self.transformer = ComposeTrans(test_transformers)
logger.info(TRANSF_MESSAGE)
self.model = torch.load(args.snapshot, map_location='cpu')
# Unpack models taht were trained as data parallel
if isinstance(self.model, nn.DataParallel):
self.model = self.model.module
# Add use precomputed hiddens for models trained before it was introduced.
# Assumes a resnet base backbone
try:
self.model._model.args.use_precomputed_hiddens = args.use_precomputed_hiddens
self.model._model.args.cuda = args.cuda
except Exception as e:
pass
# Load callibrator if desired
if args.callibrator_path is not None:
self.callibrator = pickle.load(open(args.callibrator_path, 'rb'))
else:
self.callibrator = None
logger.info(MODEL_MESSAGE.format(args.snapshot))
self.aggregator = aggregator_factory.get_exam_aggregator(
aggregator_name)
self.logger = logger
logger.info(ONCODATA_SUCCESS_MSG)
if onconet_args.use_risk_factors:
assert 'mrn' in metadata
assert 'accession' in metadata
risk_factors = json.loads(
request.files.getlist('risk_factors')[0].read())
risk_factor_vector = oncoqueries_wrapper.get_risk_factors(
onconet_args, metadata['mrn'], metadata['accession'],
risk_factors, oncodata_args.temp_img_dir, logger)
logger.info(ONCOQUERIES_SUCCESS_MSG)
else:
risk_factor_vector = None
y = onconet.process_exam(images, risk_factor_vector)
logger.info(ONCONET_SUCCESS_MSG)
msg = 'OK'
response['prediction'] = y
response['msg'] = msg
return jsonify(response), HTTP_200_OK
except Exception as e:
msg = ONCOSERVE_FAIL_MSG.format(str(e))
response['prediction'] = None
response['msg'] = msg
return jsonify(response), HTTP_500_INTERNAL_SERVER_ERROR
if __name__ == '__main__':
port = app.config['PORT']
logger.info("Launching app at port {}".format(port))
app.run(host='0.0.0.0', port=port)