Does CHaiDNN support the Depthwise convolution layer?

[mazhongzhong]

The document (https://github.com/Xilinx/CHaiDNN/blob/master/docs/SUPPORTED_LAYERS.md) says the CHaiDNN supports the Depthwise Convolution Layer. But when we actually run a Depthwise Convolution Layer on it, the CHaiDNN can run it and outputs a result, but the result is not correct. So, I wonder Does the CHaiDNN support the Depthwise convolution layer or not? If the CHaiDNN does support the Depthwise convolution layer, has anyone used this successfully? Is there a demo to show how to use this feature?

[ishida66]

It seems that the problem lies in the quantization rather than the layer implementation itself.

The paper below attempted a MobileNet model that could be used on fixed point quantization.

https://arxiv.org/pdf/1803.08607.pdf

[xlxxxxxl]

Hi @mazhongzhong @ishida66
Have you solved this problem please?

Thanks
xuelei