def apply(self, sample: Sample) -> Sample:
inputs = sample.inputs
outputs = sample.outputs
assert inputs["img_len"].shape == (1, )
inputs = inputs.copy()
outputs = outputs.copy()
inputs["img_len"] = inputs["img_len"][0]
def reshape_outputs(suffix):
out_len = "out_len" + suffix
if out_len in outputs and outputs[out_len].shape == (1, ):
outputs[out_len] = outputs[out_len][0]
for name in {
"logits",
"softmax",
"blank_last_logits",
"blank_last_softmax",
}:
name += suffix
if name in outputs:
outputs[name] = outputs[name][:outputs[out_len]]
reshape_outputs("")
for i in range(self.data_params.ensemble):
reshape_outputs(f"_{i}")
return sample.new_inputs(inputs).new_outputs(outputs)
def apply(self, sample: Sample) -> Sample:
assert self.data_params.downscale_factor > 0 # Not instantiated yet
codec = self.data_params.codec
# final preparation
if self.mode in {PipelineMode.TRAINING, PipelineMode.EVALUATION}:
text = np.array(codec.encode(sample.targets) if sample.targets else np.zeros((0,), dtype="int32"))
else:
text = None
line = sample.inputs
# gray or binary input, add missing axis
if len(line.shape) == 2:
line = np.expand_dims(line, axis=-1)
# Validate if the line is valid for training
if not self.is_valid_line(
text, len(line) // self.data_params.downscale_factor, len(line), sample.meta.get("id", "Unknown Sample ID")
):
return sample.new_invalid()
if text is not None:
sample = sample.new_targets(
{
"gt": np.asarray(text),
"gt_len": np.asarray([len(text)]),
"fold_id": np.asarray([sample.meta.get("fold_id", -1)]),
}
)
return sample.new_inputs({"img": line.astype(np.uint8), "img_len": np.asarray([len(line)])})
def apply(self, sample: Sample) -> Sample:
# data augmentation
if not self.data_aug_params.no_augs() \
and sample.inputs is not None \
and self.data_augmenter \
and np.random.rand() <= self.data_aug_params.to_rel():
line, text = self.augment(sample.inputs, sample.targets, sample.meta)
return sample.new_inputs(line).new_targets(text)
return sample
def apply(self, sample: Sample) -> Sample:
try:
return sample.new_inputs(
self._apply_single(sample.inputs, sample.meta))
except Exception as e:
logger.exception(e)
logger.warning(
"There was an unknown error when processing a line image. The line is skipped.\n"
f"The error was caused by the line with meta data: {sample.meta}.\n"
f"Please report this as an issue including the meta data, stack trace, the respective "
f"image file and call.\n"
f"You can ignore this error if it occurs only very rarely, only this particular line will "
f"be skipped.")
return sample.new_invalid()
def apply(self, sample: Sample) -> Sample:
inputs = sample.inputs
outputs = sample.outputs
assert (inputs['img_len'].shape == (1,))
inputs = inputs.copy()
outputs = outputs.copy()
inputs['img_len'] = inputs['img_len'][0]
def reshape_outputs(suffix):
out_len = 'out_len' + suffix
if out_len in outputs and outputs[out_len].shape == (1,):
outputs[out_len] = outputs[out_len][0]
for name in {'logits', 'softmax', 'blank_last_logits', 'blank_last_softmax'}:
name += suffix
if name in outputs:
outputs[name] = outputs[name][:outputs[out_len]]
reshape_outputs('')
for i in range(self.data_params.ensemble):
reshape_outputs(f"_{i}")
return sample.new_inputs(inputs).new_outputs(outputs)
def apply(self, sample: Sample) -> Sample:
assert (self.data_params.downscale_factor > 0) # Not instantiated yet
codec = self.data_params.codec
# final preparation
text = np.array(
codec.encode(sample.targets) if sample.targets else np.zeros(
(0, ), dtype='int32'))
line = sample.inputs
# gray or binary input, add missing axis
if len(line.shape) == 2:
line = np.expand_dims(line, axis=-1)
if self.mode in {PipelineMode.TRAINING, PipelineMode.EVALUATION
} and not self.is_valid_line(
text,
len(line) // self.data_params.downscale_factor):
# skip longer outputs than inputs (also in evaluation due to loss computation)
logger.warning(
f"Skipping line with longer outputs than inputs (id={sample.meta['id']})"
)
return sample.new_invalid()
if self.mode in {PipelineMode.TRAINING, PipelineMode.EVALUATION
} and len(text) == 0:
logger.warning(
f"Skipping empty line with empty GT (id={sample.meta['id']})")
return sample.new_invalid()
return sample.new_inputs({
'img': line.astype(np.uint8),
'img_len': [len(line)]
}).new_targets({
'gt': text,
'gt_len': [len(text)],
'fold_id': [sample.meta.get('fold_id', -1)]
})
def apply(self, sample: Sample) -> Sample:
return sample.new_inputs(self._apply_single(sample.inputs,
sample.meta))