Identity layer is considered as 'Unimplemented nodes'

Reproduction

Identity layer(no function in forward) causes the unimplemented nodes error.

import torch
import torch.nn as nn
import nobuco
from nobuco import ChannelOrder

class Identity(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self,x):
        return x

net = Identity()

net.eval()
input_tensor = torch.randn(10,20)

keras_mode = nobuco.pytorch_to_keras(
    net,
    args = [input_tensor], kwargs=None,
    inputs_channel_order=ChannelOrder.TENSORFLOW,
    outputs_channel_order=ChannelOrder.TENSORFLOW
)

---------------------------------------------------------------------------
---------------------------------------------------------------------------
Exception                                 Traceback (most recent call last)
Cell In[11], line 18
     15 net.eval()
     16 input_tensor = torch.randn(10,20)
---> 18 keras_mode = nobuco.pytorch_to_keras(
     19     net,
     20     args = [input_tensor], kwargs=None,
     21     inputs_channel_order=ChannelOrder.TENSORFLOW,
     22     outputs_channel_order=ChannelOrder.TENSORFLOW
     23 )

File ~nobuco/nobuco/convert.py:335, in pytorch_to_keras(model, args, kwargs, input_shapes, inputs_channel_order, outputs_channel_order, trace_shape, enable_torch_tracing, constants_to_variables, full_validation, validation_tolerance, return_outputs_pt, save_trace_html, debug_traces)
    333     print('Unimplemented nodes:')
    334     print(unimplemented_hierarchy.__str__(**vis_params))
--> 335     raise Exception('Unimplemented nodes')
    337 keras_op = keras_converted_node.keras_op
    339 args_tf, kwargs_tf = prepare_inputs_tf((args, kwargs), inputs_channel_order, input_shapes)

Exception: Unimplemented nodes

Reason

The Identity module is considered as a node. The node itself or its children should be registered in convert_dict. However, it is not registered, and it has no children.

def find_unimplemented(hierarchy: PytorchNodeHierarchy, converter_dict: Dict[object, Pytorch2KerasNodeConverter]) -> Optional[PytorchNodeHierarchy]:
    # Test if the node itself has a converter
    if has_converter(hierarchy.node, converter_dict):
        return None
    elif len(hierarchy.children) == 0:
        return PytorchNodeHierarchy(hierarchy.node, hierarchy.children)
    else:
        children_unimplemented = []
        for child in hierarchy.children:
            child_unimplemented = find_unimplemented(child, converter_dict)
            if child_unimplemented is not None:
                children_unimplemented.append(child_unimplemented)

        if len(children_unimplemented) > 0:
            # The node is unimplemented
            return PytorchNodeHierarchy(hierarchy.node, children_unimplemented)

AlexanderLutsenko / nobuco

Identity layer is considered as 'Unimplemented nodes' #38

Reproduction

Reason