Hello, I think there is a bug with the tensorflow max_pool_with_argmax op.
When running the op with onnxruntime, I find that the op returns the pooled values twice instead of the pooled values and the pooled indices. Here is a code reproducing the bug:
Note that I had to cast the resulting indices as float otherwise I get the following error during inference:
FAIL : Load model from test.onnx failed:Type Error: Type (tensor(int64)) of output arg (Identity:0) of node (bug/MaxPoolWithArgmax_transpose_2_1) does not match expected type (tensor(float)).
This was obtained on ubuntu 20.04, with tensorflow 2.4.1, onnx 1.8.1, keras2onnx 1.8.0, onnxruntime 1.7.0, onnxconverter-common 1.8.0.
Hello, I think there is a bug with the tensorflow max_pool_with_argmax op. When running the op with onnxruntime, I find that the op returns the pooled values twice instead of the pooled values and the pooled indices. Here is a code reproducing the bug:
In the end it produces:
Note that I had to cast the resulting indices as float otherwise I get the following error during inference:
This was obtained on ubuntu 20.04, with tensorflow 2.4.1, onnx 1.8.1, keras2onnx 1.8.0, onnxruntime 1.7.0, onnxconverter-common 1.8.0.