Kernel filtering not correctly replicated

pg020196 / Neural-Network-Translator

This repository contains a neural network translator which can translate neural network models to native code.

MIT License

8 stars 4 forks source link

Kernel filtering not correctly replicated #144

Open TerrestrialEclipse opened 3 years ago

TerrestrialEclipse commented 3 years ago

Keras and Pytorch apply kernels differently:

Keras cuts off the filter kernel at the edges
Pytorch implicitly invokes zero padding to apply the whole filter kernel on the padded image

This could potentially be solved by encoding the intended kernel filtering behavior into the intermediate format. This also involves adapting the backends.

pg020196 commented 3 years ago

Hi @TerrestrialEclipse,

thank you for pointing to this issue. In order to resolve this, I would agree to extend the intermediate format with an additional property for the kernel filtering behaviour, as you proposed in your initial post. This would result in simply setting the new property in the keras frontend and adjusting the backends accordingly.