Negative paddings

[jojaii]

The converter is outputting negative paddings in the graph whenever "same" padding is encountered. The model I used is the minixception model (accessible here).

This can be seen in the "conv2d_3" layer in the model. You may download the file here.

Here is a screenshot of the layer when opened with Netron:

Here is the sample code:

import onnx
from keras.models import load_model
from keras2onnx import convert_keras

model = load_model('fer2013_mini_XCEPTION.107-0.66.hdf5')
onnx_model = convert_keras(model)
onnx.save(onnx_model, 'minixception_sample.onnx')

[jiafatom]

@jojaii I have a fix for this, will check in soon.

[jojaii]

Thanks @jiafatom. Will wait for the fix.

[jiafatom]

Thanks @jiafatom. Will wait for the fix.

@jojaii code checked in, thanks for reporting the issue.