Onnx model converted from hdf5 model has lstm nodes with different dimensions for bias and kernel

[tijithomas]

The keras model:

After conversion, the onnx model has transpose after every node and the dimension for bias on lstm nodes have changed from [2000] to [1,4000]. Although after 2000, all the values are 0. Is the conversion correct in this case?

[jiafatom]

The dimension mismatch is because onnx and keras have different specs, please check onnx spec, it says the bias tensor has shape [num_directions, 8*hidden_size], others are 4*hidden_size, so it is 2x, which is correct.

[tijithomas]

@jiafatom Thanks for the quick reply!
The hdf5 model has dynamic batch size(?) in input/output placeholder, the output placeholder of the converted onnx (same model as above) shows 0 as batch size.

Should it not be '-1' or '?' instead of '0'?

[jiafatom]

Can you pull latest keras2onnx master to try? It should be 'N' for both input and output batch.

[tijithomas]

Thanks, it works!