def test_postprocessor_str():
postprocessor = _Postprocessor(
default_uncollate,
torch.relu,
torch.softmax,
None,
)
assert str(postprocessor) == (
"_Postprocessor:\n"
"\t(per_batch_transform): FuncModule(relu)\n"
"\t(uncollate_fn): FuncModule(default_uncollate)\n"
"\t(per_sample_transform): FuncModule(softmax)\n"
"\t(serializer): None")
def _create_uncollate_postprocessors(
self,
stage: RunningStage,
is_serving: bool = False,
) -> _Postprocessor:
save_per_sample = None
save_fn = None
postprocess: Postprocess = self._postprocess_pipeline
func_names: Dict[str, str] = {
k: self._resolve_function_hierarchy(k,
postprocess,
stage,
object_type=Postprocess)
for k in self.POSTPROCESS_FUNCS
}
# since postprocessing is exclusive for prediction, we don't have to check the resolution hierarchy here.
if postprocess._save_path:
save_per_sample: bool = self._is_overriden_recursive(
"save_sample",
postprocess,
Postprocess,
prefix=_STAGES_PREFIX[stage])
if save_per_sample:
save_per_sample: Callable = getattr(postprocess,
func_names["save_sample"])
else:
save_fn: Callable = getattr(postprocess,
func_names["save_data"])
return _Postprocessor(
getattr(postprocess, func_names["uncollate"]),
getattr(postprocess, func_names["per_batch_transform"]),
getattr(postprocess, func_names["per_sample_transform"]),
serializer=None if is_serving else self._serializer,
save_fn=save_fn,
save_per_sample=save_per_sample,
is_serving=is_serving,
)