def predict(self, image_paths):
path = Path(image_paths[0]).parent
logger.info(f'Running model on {len(image_paths)} images in {path}')
probs, labels = self.model.predict(image_paths)
pred_list = [{
'prob': float(prob),
'label': label
} for prob, label in zip(probs, labels)]
return {'predictions': pred_list}
def save_images(self, doc, path):
images_doc = doc['images'][:self.images_limit]
logger.info(f"Saving {len(images_doc)} images to {path}")
image_paths = []
for i, image_doc in enumerate(images_doc):
img_bytes = base64.b64decode(image_doc['content'])
image_path = path / f'image-{i}.png'
with open(image_path, 'wb') as f:
f.write(img_bytes)
image_paths.append(image_path)
return image_paths
def on_post(self, req, resp):
path = None
try:
req_doc = json.load(req.bounded_stream)
path = self.data_path / str(uuid4())
path.mkdir()
image_paths = self.save_images(req_doc, path)
resp_doc = self.predict(image_paths)
resp.body = json.dumps(resp_doc)
except IOError as e:
logger.error(e, exc_info=True)
raise falcon.HTTPError('500')
finally:
if path:
logger.info(f'Cleaning path {path}')
shutil.rmtree(path)
def test_service():
DATA_PATH.mkdir(exist_ok=True)
logger.info(f'Downloading input data from {TEST_DATA_URL} to {DATA_PATH}')
curl_sp = subprocess.Popen(f'curl -s {TEST_DATA_URL}'.split(' '),
stdout=subprocess.PIPE)
subprocess.Popen(f'tar xz -C {DATA_PATH}'.split(' '), stdin=curl_sp.stdout)
path_label_df = pd.read_csv(DATA_PATH / 'path_label.csv')
logger.info('Label counts: {}'.format(
path_label_df.label.value_counts().to_dict()))
image_paths = [DATA_PATH / p for p in path_label_df.path.values]
image_path_chunks = make_chunk_gen(image_paths, chunk_size=32)
predictions = []
with ThreadPoolExecutor(2) as pool:
for ch_preds in pool.map(predict_chunk, image_path_chunks):
predictions.extend(ch_preds)
labels = (path_label_df.label == 'off').astype(int).values
assert len(predictions) == len(labels)
preds = [p['prob'] > 0.5 for p in predictions]
test_accuracy = (preds == labels).mean()
logger.info(f'Test accuracy: {test_accuracy}')
assert test_accuracy > 0.95, f'Not enough model test set accuracy: {test_accuracy}'
import falcon
from off_sample.resources import PingResource, PredictResource
from off_sample.utils import logger
def create_app():
api = falcon.API()
api.add_route('/off-sample', PingResource())
api.add_route('/off-sample/predict', PredictResource())
return api
def get_app():
return create_app()
if __name__ == '__main__':
logger.info('Creating app...')
app = get_app()
from wsgiref import simple_server
httpd = simple_server.make_server('0.0.0.0', 9876, app)
logger.info('Running debug server...')
httpd.serve_forever()