Feature Request: More padding mode.

[Soptq]

In many applications we want to make sure that the padding alters the local feature structure and global statistics the least (e.g. Image-to-image translation). Using zero padding in these applications will result in peaking edges[1][2].

However, only zero padding is supported by brevitas by now. I wonder is it possible to also support other padding mode like reflection or symmetric padding in brevitas? or is there a workaround?

[1] https://arxiv.org/abs/1703.10593
[2] https://arxiv.org/abs/1811.11718

[Giuseppe5]

I think it should be relatively straightforward to implement your own padding mode.
Inherit a new QuantConv from brevitas' one, then you'd need to override the init and this:

brevitas/src/brevitas/nn/quant_conv.py

Lines 200 to 205 in 2004568

	def inner_forward_impl(self, x: Tensor, quant_weight: Tensor, quant_bias: Optional[Tensor]):
	if self.is_same_padded_strided:
	return self.conv2d_same_zeros_pad_stride(x, quant_weight, quant_bias)
	else:
	return self._conv_forward(x, quant_weight, quant_bias)

in order to call your custom padding mode. As long as you call torch.nn.functional.conv2d (or 1d, etc.) with the input/weight/bias tensor that are passed to inner_forward_impl, everything else with respect to quantization is handled automatically.

[Giuseppe5]

If you think that this feature could be beneficial for others and are willing to contribute to Brevitas, we would be more than happy to accept a PR.

[Soptq]

Sure, I will try to make a PR later ;)

[Giuseppe5]

Closing this now. Feel free to re-open if you need more help!