# attributes that need to be adjusted, once the Curriculum decides to use
# a more difficult dataset
# this is a 'map' of attribute name to a path in the trainer object
attributes_to_adjust = [
('num_timesteps', ['predictor', 'localization_net']),
('num_timesteps', ['predictor', 'recognition_net']),
('num_timesteps', ['lossfun', '__self__']),
('num_labels', ['predictor', 'recognition_net']),
]
# create train curriculum
curriculum = BabyStepCurriculum(
args.dataset_specification,
FileBasedDataset,
args.blank_label,
attributes_to_adjust=attributes_to_adjust,
trigger=(args.test_interval, 'iteration'),
min_delta=0.1,
)
train_dataset, validation_dataset = curriculum.load_dataset(0)
# the metrics object calculates the loss
metrics = SoftmaxMetrics(
args.blank_label,
args.char_map,
train_dataset.num_timesteps,
image_size,
area_loss_factor=args.area_factor,
aspect_ratio_loss_factor=args.aspect_factor,
area_scaling_factor=args.area_scale_factor,
uses_original_data=args.is_original_fsns,
# attributes that need to be adjusted, once the Curriculum decides to use
# a more difficult dataset
# this is a 'map' of attribute name to path in trainer object
attributes_to_adjust = [
('num_timesteps', ['predictor', 'localization_net']),
('num_timesteps', ['predictor', 'recognition_net']),
('num_timesteps', ['lossfun', '__self__']),
('num_labels', ['predictor', 'recognition_net']),
]
curriculum = BabyStepCurriculum(args.dataset_specification,
TextRecFileDataset,
args.blank_label,
attributes_to_adjust=attributes_to_adjust,
trigger=(args.test_interval, 'iteration'),
min_delta=1.0,
dataset_args={
'char_map': args.char_map,
'resize_size': target_shape,
'blank_label': args.blank_label,
})
train_dataset, validation_dataset = curriculum.load_dataset(0)
train_dataset.resize_size = image_size
validation_dataset.resize_size = image_size
metrics = TextRecSoftmaxMetrics(
args.blank_label,
args.char_map,
train_dataset.num_timesteps,
image_size,