def __init__(self, args):
self.args = args
with open(os.path.join(args.model_dir, args.log_name)) as the_log:
log_data = json.load(the_log)[0]
self.target_shape = Size._make(log_data['target_size'])
self.image_size = Size._make(log_data['image_size'])
# Step 1: build network
localization_net_class_name, localization_module_name = self.get_class_and_module(log_data['localization_net'])
module = self.load_module(os.path.abspath(os.path.join(args.model_dir, localization_module_name)))
localization_net_class = eval('module.{}'.format(localization_net_class_name))
localization_net = self.build_localization_net(localization_net_class)
recognition_net_class_name, recognition_module_name = self.get_class_and_module(log_data['recognition_net'])
module = self.load_module(os.path.abspath(os.path.join(args.model_dir, recognition_module_name)))
recognition_net_class = eval('module.{}'.format(recognition_net_class_name))
recognition_net = self.build_recognition_net(recognition_net_class)
fusion_net_class_name, fusion_module_name = self.get_class_and_module(log_data['fusion_net'])
module = self.load_module(os.path.abspath(os.path.join(args.model_dir, fusion_module_name)))
fusion_net_class = eval('module.{}'.format(fusion_net_class_name))
self.net = self.build_fusion_net(fusion_net_class, localization_net, recognition_net)
if args.gpu >= 0:
self.net.to_gpu(args.gpu)
# Step 2: load weights
with np.load(os.path.join(args.model_dir, args.snapshot_name)) as f:
chainer.serializers.NpzDeserializer(f).load(self.net)
# Step 3: open gt and do evaluation
with open(args.char_map) as the_map:
self.char_map = json.load(the_map)
self.xp = chainer.cuda.cupy if args.gpu >= 0 else np
with open(args.eval_gt) as eval_gt:
reader = csv.reader(eval_gt, delimiter='\t')
self.lines = [l for l in reader]
self.blank_symbol = args.blank_symbol
self.num_correct_lines = 0
self.num_correct_words = 0
self.num_lines = 0
self.num_words = 0
self.num_word_x_correct = [0 for _ in range(args.timesteps)]
self.num_word_x = [0 for _ in range(args.timesteps)]
self.model_dir = args.model_dir
self.metrics = self.create_metrics()
self.save_rois = args.save_rois
self.bbox_plotter = None
if self.save_rois:
self.create_bbox_plotter()
args = parser.parse_args()
# set standard args that should always hold true if using the supplied model
args.is_original_fsns = True
args.log_name = 'log'
args.dropout_ratio = 0.5
args.blank_symbol = 0
# max number of text regions in the image
args.timesteps = 4
# max number of characters per word
args.num_labels = 21
# open log and extract meta information
with open(os.path.join(args.model_dir, args.log_name)) as the_log:
log_data = json.load(the_log)[0]
target_shape = Size._make(log_data['target_size'])
image_size = Size._make(log_data['image_size'])
xp = chainer.cuda.cupy if args.gpu >= 0 else np
network = create_network(args, log_data)
# load weights
with np.load(os.path.join(args.model_dir, args.snapshot_name)) as f:
chainer.serializers.NpzDeserializer(f).load(network)
# load char map
with open(args.char_map) as the_map:
char_map = json.load(the_map)
# load image
image = load_image(args.image_path, xp)
