keras-onnx icon indicating copy to clipboard operation
keras-onnx copied to clipboard

TF 2.3/2.4 support?

Open affranchi opened this issue 4 years ago • 0 comments

Hi,

I have a question regarding TF2.3 and TF2.4 support. I have a model which is not working with TF2.3 (It worked with TF2.2).

The error is

../lib/python3.6/site-packages/keras2onnx/main.py:75: in convert_keras
    parse_graph(topology, tf_graph, target_opset, output_names, output_dict)
../lib/python3.6/site-packages/keras2onnx/parser.py:841: in parse_graph
    ) if is_tf2 and is_tf_keras else _parse_graph_core(
../lib/python3.6/site-packages/keras2onnx/parser.py:699: in _parse_graph_core_v2
    _create_link_node(var_ts, top_scope, varset, adjust_batch_size=True)
../lib/python3.6/site-packages/keras2onnx/parser.py:380: in _create_link_node
    op.add_output(var0)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = <keras2onnx.common.intop.Operator object at 0x7fe0c05b4e10>
var = L0/while_2:0

    def add_output(self, var):
        if var.op_from is not None:
>           assert False, "Tensor {} already processed".format(var.full_name)
E           AssertionError: Tensor L0/while_2:0 already processed

Could you let me know why the problem happens? Also, is there any timeline to support TF2.3 or TF2.4 officially?

Thanks.

affranchi avatar Jan 06 '21 21:01 affranchi