Wrong number of parameter for flan_t5_small

[VictorY-Irreversible]

Describe the bug

Based on your expected output file for flan_t5_small you got 128,743,488 parameters.

The paper mention 80M and huggingface says 77M.
With my own script I arrive to 76,961,152, which matches with the expected order of magnitude.

To Reproduce

Run:

torchinfo/tests/torchinfo_xl_test.py

Lines 158 to 165 in 164597f

	def test_flan_t5_small() -> None:
	model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-small")
	inputs = {
	"input_ids": torch.zeros(3, 100).long(),
	"attention_mask": torch.zeros(3, 100).long(),
	"labels": torch.zeros(3, 100).long(),
	}
	summary(model, input_data=inputs)

Expected behavior

Get the good number of parameters.

[VictorY-Irreversible]

Similar issue for googlenet.
Based on the table 1 of the paper, the model should be around 7M parameters.
You got 13M in the output file for both inference and training.

I suspect that the leftover calculation is broken:

torchinfo/torchinfo/layer_info.py

Lines 307 to 325 in 164597f

	def leftover_params(self) -> int:
	"""
	Leftover params are the number of params this current layer has that are not
	included in the child num_param counts.
	"""
	return self.num_params - sum(
	child.num_params if child.is_leaf_layer else child.leftover_params()
	for child in self.children
	if not child.is_recursive
	)
	def leftover_trainable_params(self) -> int:
	return self.trainable_params - sum(
	child.trainable_params
	if child.is_leaf_layer
	else child.leftover_trainable_params()
	for child in self.children
	if not child.is_recursive
	)

I don't think that it is necessary to calculate the leftover parameters when you can simply use named_parameters(recurse=False).