Hello.
I need to implement a segmentation-aware conv operation, which means during conv, the weights of the filter is multiplied by a local segmentation-aware weights.
As the weights are different in different location, i need to unfold the image conv to act like matrix multiplcation.
Thanks.
you could either use torch.unfold from pytorch or im2col/col2im from https://github.com/szagoruyko/pyinn
1 Like
will unfold support dilated spacing in future? with which the dilated convolution can be implemented too.
will unfold support dilated spacing in future? with which the dilated convolution can be implemented too.
We now have nn.functional.unfold that supports batches, dilation and is pretty much the same as im2col for 4d inputs: https://pytorch.org/docs/stable/nn.html#torch.nn.functional.unfold
1 Like
Sorry for not watching the latest update… Is it available in version 0.3 ? I can’t visit the version selection website so can only see version 0.4…
Another question is about the functional caller, suppose I am customizing a operator which needs im2col operation, for a input variable v_input, after unfold, I get v_unfold, them v_unfold will do a product with another variable v_another, according to my understanding of autograd mechanism, to log the graph of and for the backprop for product operation, we need to save the operand v_unfold and v_another, however, after im2col, the variable v_unfold will consume to much memory. I think we can compose these two operator, the im2col, and product, them when forward, we only need to save the tensor of v_input, and for forward and backward, do the unfold manually, therefore, it can save many memory. However, when feed the functional a tensor, it will return a Variable, is there a way to do pure tensor operation with nn.functional ? thanks.