def __init__(
self,
mode: DataSetMode,
args: dict,
):
""" Create a dataset from memory
Since this dataset already contains all data in the memory, this dataset may not be loaded
Parameters
----------
"""
super().__init__(mode)
self.loaded = False
self.lines_per_epoch = 10000
self._samples = [{
'id': '{}'.format(i)
} for i in range(self.lines_per_epoch)]
self.text_generator_params = args.get('text_generator_params',
TextGeneratorParameters())
self.line_generator_params = args.get('line_generator_params',
LineGeneratorParameters())
self.manager = Manager()
self.data_queue = self.manager.Queue(50)
self.data_generators = [
LineGeneratorProcess(
self.data_queue,
self.text_generator_params,
self.line_generator_params,
"{}".format(i),
) for i in range(8)
]
for d in self.data_generators:
d.start()
def run(cfg: CfgNode):
# check if loading a json file
if len(cfg.DATASET.TRAIN.PATH) == 1 and cfg.DATASET.TRAIN.PATH[0].endswith(
"json"):
import json
with open(cfg.DATASET.TRAIN.PATH[0], 'r') as f:
json_args = json.load(f)
for key, value in json_args.items():
if key == 'dataset' or key == 'validation_dataset':
setattr(cfg, key, DataSetType.from_string(value))
else:
setattr(cfg, key, value)
# parse whitelist
whitelist = cfg.MODEL.CODEX.WHITELIST
if len(whitelist) == 1:
whitelist = list(whitelist[0])
whitelist_files = glob_all(cfg.MODEL.CODEX.WHITELIST_FILES)
for f in whitelist_files:
with open(f) as txt:
whitelist += list(txt.read())
if cfg.DATASET.TRAIN.GT_EXTENSION is False:
cfg.DATASET.TRAIN.GT_EXTENSION = DataSetType.gt_extension(
cfg.DATASET.TRAIN.TYPE)
if cfg.DATASET.VALID.GT_EXTENSION is False:
cfg.DATASET.VALID.GT_EXTENSION = DataSetType.gt_extension(
cfg.DATASET.VALID.TYPE)
text_generator_params = TextGeneratorParameters()
line_generator_params = LineGeneratorParameters()
dataset_args = {
'line_generator_params': line_generator_params,
'text_generator_params': text_generator_params,
'pad': None,
'text_index': 0,
}
# Training dataset
dataset = create_train_dataset(cfg, dataset_args)
# Validation dataset
validation_dataset_list = create_test_dataset(cfg, dataset_args)
params = CheckpointParams()
params.max_iters = cfg.SOLVER.MAX_ITER
params.stats_size = cfg.STATS_SIZE
params.batch_size = cfg.SOLVER.BATCH_SIZE
params.checkpoint_frequency = cfg.SOLVER.CHECKPOINT_FREQ if cfg.SOLVER.CHECKPOINT_FREQ >= 0 else cfg.SOLVER.EARLY_STOPPING_FREQ
params.output_dir = cfg.OUTPUT_DIR
params.output_model_prefix = cfg.OUTPUT_MODEL_PREFIX
params.display = cfg.DISPLAY
params.skip_invalid_gt = not cfg.DATALOADER.NO_SKIP_INVALID_GT
params.processes = cfg.NUM_THREADS
params.data_aug_retrain_on_original = not cfg.DATALOADER.ONLY_TRAIN_ON_AUGMENTED
params.early_stopping_at_acc = cfg.SOLVER.EARLY_STOPPING_AT_ACC
params.early_stopping_frequency = cfg.SOLVER.EARLY_STOPPING_FREQ
params.early_stopping_nbest = cfg.SOLVER.EARLY_STOPPING_NBEST
params.early_stopping_best_model_prefix = cfg.EARLY_STOPPING_BEST_MODEL_PREFIX
params.early_stopping_best_model_output_dir = \
cfg.EARLY_STOPPING_BEST_MODEL_OUTPUT_DIR if cfg.EARLY_STOPPING_BEST_MODEL_OUTPUT_DIR else cfg.OUTPUT_DIR
if cfg.INPUT.DATA_PREPROCESSING is False or len(
cfg.INPUT.DATA_PREPROCESSING) == 0:
cfg.INPUT.DATA_PREPROCESSING = [
DataPreprocessorParams.DEFAULT_NORMALIZER
]
params.model.data_preprocessor.type = DataPreprocessorParams.MULTI_NORMALIZER
for preproc in cfg.INPUT.DATA_PREPROCESSING:
pp = params.model.data_preprocessor.children.add()
pp.type = DataPreprocessorParams.Type.Value(preproc) if isinstance(
preproc, str) else preproc
pp.line_height = cfg.INPUT.LINE_HEIGHT
pp.pad = cfg.INPUT.PAD
# Text pre processing (reading)
params.model.text_preprocessor.type = TextProcessorParams.MULTI_NORMALIZER
default_text_normalizer_params(
params.model.text_preprocessor.children.add(),
default=cfg.INPUT.TEXT_NORMALIZATION)
default_text_regularizer_params(
params.model.text_preprocessor.children.add(),
groups=cfg.INPUT.TEXT_REGULARIZATION)
strip_processor_params = params.model.text_preprocessor.children.add()
strip_processor_params.type = TextProcessorParams.STRIP_NORMALIZER
# Text post processing (prediction)
params.model.text_postprocessor.type = TextProcessorParams.MULTI_NORMALIZER
default_text_normalizer_params(
params.model.text_postprocessor.children.add(),
default=cfg.INPUT.TEXT_NORMALIZATION)
default_text_regularizer_params(
params.model.text_postprocessor.children.add(),
groups=cfg.INPUT.TEXT_REGULARIZATION)
strip_processor_params = params.model.text_postprocessor.children.add()
strip_processor_params.type = TextProcessorParams.STRIP_NORMALIZER
if cfg.SEED > 0:
params.model.network.backend.random_seed = cfg.SEED
if cfg.INPUT.BIDI_DIR:
# change bidirectional text direction if desired
bidi_dir_to_enum = {
"rtl": TextProcessorParams.BIDI_RTL,
"ltr": TextProcessorParams.BIDI_LTR,
"auto": TextProcessorParams.BIDI_AUTO
}
bidi_processor_params = params.model.text_preprocessor.children.add()
bidi_processor_params.type = TextProcessorParams.BIDI_NORMALIZER
bidi_processor_params.bidi_direction = bidi_dir_to_enum[
cfg.INPUT.BIDI_DIR]
bidi_processor_params = params.model.text_postprocessor.children.add()
bidi_processor_params.type = TextProcessorParams.BIDI_NORMALIZER
bidi_processor_params.bidi_direction = TextProcessorParams.BIDI_AUTO
params.model.line_height = cfg.INPUT.LINE_HEIGHT
params.model.network.learning_rate = cfg.SOLVER.LR
params.model.network.lr_decay = cfg.SOLVER.LR_DECAY
params.model.network.lr_decay_freq = cfg.SOLVER.LR_DECAY_FREQ
params.model.network.train_last_n_layer = cfg.SOLVER.TRAIN_LAST_N_LAYER
network_params_from_definition_string(cfg.MODEL.NETWORK,
params.model.network)
params.model.network.clipping_norm = cfg.SOLVER.GRADIENT_CLIPPING_NORM
params.model.network.backend.num_inter_threads = 0
params.model.network.backend.num_intra_threads = 0
params.model.network.backend.shuffle_buffer_size = cfg.DATALOADER.SHUFFLE_BUFFER_SIZE
if cfg.MODEL.WEIGHTS == "":
weights = None
else:
weights = cfg.MODEL.WEIGHTS
# create the actual trainer
trainer = Trainer(
params,
dataset,
validation_dataset=validation_dataset_list,
data_augmenter=SimpleDataAugmenter(),
n_augmentations=cfg.INPUT.N_AUGMENT,
weights=weights,
codec_whitelist=whitelist,
keep_loaded_codec=cfg.MODEL.CODEX.KEEP_LOADED_CODEC,
preload_training=not cfg.DATALOADER.TRAIN_ON_THE_FLY,
preload_validation=not cfg.DATALOADER.VALID_ON_THE_FLY,
)
trainer.train(auto_compute_codec=not cfg.MODEL.CODEX.SEE_WHITELIST,
progress_bar=not cfg.NO_PROGRESS_BAR)
def run(args):
# check if loading a json file
if len(args.files) == 1 and args.files[0].endswith("json"):
import json
with open(args.files[0], 'r') as f:
json_args = json.load(f)
for key, value in json_args.items():
if key == 'dataset' or key == 'validation_dataset':
setattr(args, key, DataSetType.from_string(value))
else:
setattr(args, key, value)
# parse whitelist
whitelist = args.whitelist
if len(whitelist) == 1:
whitelist = list(whitelist[0])
whitelist_files = glob_all(args.whitelist_files)
for f in whitelist_files:
with open(f) as txt:
whitelist += list(txt.read())
if args.gt_extension is None:
args.gt_extension = DataSetType.gt_extension(args.dataset)
if args.validation_extension is None:
args.validation_extension = DataSetType.gt_extension(args.validation_dataset)
if args.text_generator_params is not None:
with open(args.text_generator_params, 'r') as f:
args.text_generator_params = json_format.Parse(f.read(), TextGeneratorParameters())
else:
args.text_generator_params = TextGeneratorParameters()
if args.line_generator_params is not None:
with open(args.line_generator_params, 'r') as f:
args.line_generator_params = json_format.Parse(f.read(), LineGeneratorParameters())
else:
args.line_generator_params = LineGeneratorParameters()
dataset_args = {
'line_generator_params': args.line_generator_params,
'text_generator_params': args.text_generator_params,
'pad': args.dataset_pad,
'text_index': args.pagexml_text_index,
}
# Training dataset
dataset = create_train_dataset(args, dataset_args)
# Validation dataset
if args.validation:
print("Resolving validation files")
validation_image_files = glob_all(args.validation)
if not args.validation_text_files:
val_txt_files = [split_all_ext(f)[0] + args.validation_extension for f in validation_image_files]
else:
val_txt_files = sorted(glob_all(args.validation_text_files))
validation_image_files, val_txt_files = keep_files_with_same_file_name(validation_image_files, val_txt_files)
for img, gt in zip(validation_image_files, val_txt_files):
if split_all_ext(os.path.basename(img))[0] != split_all_ext(os.path.basename(gt))[0]:
raise Exception("Expected identical basenames of validation file: {} and {}".format(img, gt))
if len(set(val_txt_files)) != len(val_txt_files):
raise Exception("Some validation images are occurring more than once in the data set.")
validation_dataset = create_dataset(
args.validation_dataset,
DataSetMode.TRAIN,
images=validation_image_files,
texts=val_txt_files,
skip_invalid=not args.no_skip_invalid_gt,
args=dataset_args,
)
print("Found {} files in the validation dataset".format(len(validation_dataset)))
else:
validation_dataset = None
params = CheckpointParams()
params.max_iters = args.max_iters
params.stats_size = args.stats_size
params.batch_size = args.batch_size
params.checkpoint_frequency = args.checkpoint_frequency if args.checkpoint_frequency >= 0 else args.early_stopping_frequency
params.output_dir = args.output_dir
params.output_model_prefix = args.output_model_prefix
params.display = args.display
params.skip_invalid_gt = not args.no_skip_invalid_gt
params.processes = args.num_threads
params.data_aug_retrain_on_original = not args.only_train_on_augmented
params.early_stopping_frequency = args.early_stopping_frequency
params.early_stopping_nbest = args.early_stopping_nbest
params.early_stopping_best_model_prefix = args.early_stopping_best_model_prefix
params.early_stopping_best_model_output_dir = \
args.early_stopping_best_model_output_dir if args.early_stopping_best_model_output_dir else args.output_dir
if args.data_preprocessing is None or len(args.data_preprocessing) == 0:
args.data_preprocessing = [DataPreprocessorParams.DEFAULT_NORMALIZER]
params.model.data_preprocessor.type = DataPreprocessorParams.MULTI_NORMALIZER
for preproc in args.data_preprocessing:
pp = params.model.data_preprocessor.children.add()
pp.type = DataPreprocessorParams.Type.Value(preproc) if isinstance(preproc, str) else preproc
pp.line_height = args.line_height
pp.pad = args.pad
# Text pre processing (reading)
params.model.text_preprocessor.type = TextProcessorParams.MULTI_NORMALIZER
default_text_normalizer_params(params.model.text_preprocessor.children.add(), default=args.text_normalization)
default_text_regularizer_params(params.model.text_preprocessor.children.add(), groups=args.text_regularization)
strip_processor_params = params.model.text_preprocessor.children.add()
strip_processor_params.type = TextProcessorParams.STRIP_NORMALIZER
# Text post processing (prediction)
params.model.text_postprocessor.type = TextProcessorParams.MULTI_NORMALIZER
default_text_normalizer_params(params.model.text_postprocessor.children.add(), default=args.text_normalization)
default_text_regularizer_params(params.model.text_postprocessor.children.add(), groups=args.text_regularization)
strip_processor_params = params.model.text_postprocessor.children.add()
strip_processor_params.type = TextProcessorParams.STRIP_NORMALIZER
if args.seed > 0:
params.model.network.backend.random_seed = args.seed
if args.bidi_dir:
# change bidirectional text direction if desired
bidi_dir_to_enum = {"rtl": TextProcessorParams.BIDI_RTL, "ltr": TextProcessorParams.BIDI_LTR,
"auto": TextProcessorParams.BIDI_AUTO}
bidi_processor_params = params.model.text_preprocessor.children.add()
bidi_processor_params.type = TextProcessorParams.BIDI_NORMALIZER
bidi_processor_params.bidi_direction = bidi_dir_to_enum[args.bidi_dir]
bidi_processor_params = params.model.text_postprocessor.children.add()
bidi_processor_params.type = TextProcessorParams.BIDI_NORMALIZER
bidi_processor_params.bidi_direction = TextProcessorParams.BIDI_AUTO
params.model.line_height = args.line_height
network_params_from_definition_string(args.network, params.model.network)
params.model.network.clipping_mode = NetworkParams.ClippingMode.Value("CLIP_" + args.gradient_clipping_mode.upper())
params.model.network.clipping_constant = args.gradient_clipping_const
params.model.network.backend.fuzzy_ctc_library_path = args.fuzzy_ctc_library_path
params.model.network.backend.num_inter_threads = args.num_inter_threads
params.model.network.backend.num_intra_threads = args.num_intra_threads
params.model.network.backend.shuffle_buffer_size = args.shuffle_buffer_size
# create the actual trainer
trainer = Trainer(params,
dataset,
validation_dataset=validation_dataset,
data_augmenter=SimpleDataAugmenter(),
n_augmentations=args.n_augmentations,
weights=args.weights,
codec_whitelist=whitelist,
keep_loaded_codec=args.keep_loaded_codec,
preload_training=not args.train_data_on_the_fly,
preload_validation=not args.validation_data_on_the_fly,
)
trainer.train(
auto_compute_codec=not args.no_auto_compute_codec,
progress_bar=not args.no_progress_bars
)
for d in self.data_generators:
d.start()
def _load_sample(self, sample, text_only):
return self.data_queue.get()
def create_generator(self, mp_context, output_queue) -> DatasetGenerator:
return GeneratedLineDatasetGenerator(mp_context,
output_queue, self.mode,
self.samples(), self.data_queue)
if __name__ == "__main__":
args = dict()
params = TextGeneratorParameters()
params.word_length_mean = 11
params.word_length_sigma = 3
params.number_of_words_mean = 7
params.number_of_words_mean = 4
params.word_separator = " "
params.sub_script_p = 0.2
params.super_script_p = 0.2
params.letter_spacing_p = 0.5
params.letter_spacing_mean = 0.5
params.letter_spacing_sigma = 0.05
params.bold_p = 0.5
params.italic_p = 0.5
params.codec.charset.extend(
list(
"ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789{}[]()_-.;:'\""
out.append(
Word(self.params.word_separator, Script.NORMAL, 0,
FontVariantType.NORMAL))
charset = [self.charset, self.super_charset,
self.sub_charset][script]
s = "".join(np.random.choice(charset, word_length))
s = s.strip()
out.append(Word(s, script, letter_spacing, variant))
return out
if __name__ == "__main__":
params = TextGeneratorParameters()
params.word_length_mean = 11
params.word_length_sigma = 3
params.number_of_words_mean = 7
params.number_of_words_mean = 4
params.word_separator = " "
params.sub_script_p = 0.0
params.super_script_p = 0.2
params.letter_spacing_p = 0.5
params.letter_spacing_mean = 1
params.letter_spacing_sigma = 0.1
params.charset.extend(
list(
"ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789{}[]()_-.;:'\" "
))
params.super_charset.extend(list("ABCDEFGHIJKLMNOPQRSTUVWXYZ"))