def classify_content(content):
# create a generator for fetching data
urls = []
for asset in content.assets:
urls.append(asset.url)
# prepare images for download
val_generator = UrlGenerator(urls,
cfg.resolve(cfg.RETINANET_MODEL, cfg.classes_file),
cfg.resolve(cfg.RETINANET_MODEL, cfg.labels_file))
response = Response()
# load image
for i, asset in enumerate(content.assets):
logger.info('Running classification on: {}'.format(asset.url))
# initialize result object
result = Result()
result.url = asset.url
result.asset_id = asset.asset_id
logger.info('Reading image bgr...')
try:
# fetch images
image = val_generator.read_image_bgr(i)
# index original image for searching
if content.insert:
index_original_image(image, asset)
except (OSError, ConnectTimeout, HTTPError, ReadTimeout, Timeout, ConnectionError):
logger.warning('Skipped: Unable to reach resource')
continue
except:
err = traceback.format_exc()
logger.error('Could not read image: {}'.format(err))
continue
# copy to draw on
logger.info('Drawing cvt color...')
draw = np.asarray(image.copy())
draw = cv2.cvtColor(draw, cv2.COLOR_BGR2RGB)
# pre-process the image for the network
logger.info('Processing image...')
image = val_generator.preprocess_image(image)
image, scale = val_generator.resize_image(image)
# classify image
start = time.time()
boxes, scores, labels = core.model.predict_on_batch(np.expand_dims(image, axis=0))
elapsed = time.time() - start
logger.info('Processing time: {}'.format(elapsed))
result.time = str(elapsed)
boxes /= scale
# process and save detections
idx = 0
for box, score, label in zip(boxes[0], scores[0], labels[0]):
if score < cfg.resolve_float(cfg.CLASSIFICATION, cfg.min_confidence):
continue
# get position data
box = boxes[0, idx, :4].astype(int)
label_name = val_generator.label_to_name(label)
# save meta-info for REST API response
caption = Caption(str(label),
label_name,
str(score),
'{};{}'.format(box[0], box[1]), # x1;y1
'{};{}'.format(box[2], box[3])) # x2;y2
result.captions.append(caption)
# Crop image for extraction
h = box[3] - box[1]
w = box[2] - box[0]
cropped_img = draw[box[1]:(box[1] + h), box[0]:(box[0] + w)]
if content.insert:
# update sequence to remove previous index if available
remove_cropped_if_asset_exists(asset)
# process cropped image fragment for searching
cropped_file_name = index_cropped_image(asset, cropped_img, label_name, idx, insert=content.insert)
features = extract_features(cropped_file_name)
faiss_features = features.reshape((1, cfg.resolve_int(cfg.FAISS_SETTINGS, cfg.index_size)))
# add or clean image
if content.insert:
# add feature to faiss index
core.index.add(faiss_features)
else:
# clean temp image again
os.remove(cropped_file_name)
# index caption
if content.insert:
index_asset_meta(asset, idx, caption, features.tolist(), core.index.ntotal - 1)
# find similar suggestions and handle response
asset_metas = get_similar_asset_metas(faiss_features,
cfg.resolve_int(cfg.FAISS_SETTINGS, cfg.index_n_similar_results))
handle_suggestion_response(result, asset.asset_id, asset_metas)
idx += 1
# add result to response list
response.result_list.append(result)
return response
