maureenzou / adaptive-anti-aliasing Goto Github PK

How about using sigmoid instead of softmax when generating the low-passing filter? Did you conduct the comparison experiments? I understand that all the kernels should be positive. But is it necessary to make the sum equal one?

Thanks for your impressive work. However, I noticed the operation in paper "Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR'21)" is highly similar with yours. Is there any difference? this paper just introduced the pasa into normal convolution rather than downsampling?

did you mead I should put ''Downsample_PASA_group_softmax()'' before my ''downsample block''? so I just did it and the miou drop a lot, could you help me? my downsample block is just a conv "self.conv1x1 = ConvBNPReLU(nIn, nOut, 3, 2)"

Hi author, thanks for the great work! Would you consider adding a license for this repo?

I think this implement will cause the downsample ops to choose at most one channel information from one group. Or it can be multiplied by channels per group?

I Don't find the metric of mASSC.Could you tell me the place you set?

Hi,
I wonder what 'pasa' stands for.
Thanks!

This is an interesting work. Could you provide the pretrained resnet50_k3_pasa_group8_softmax model? Thank you.

Is this the working code with DeepLabV3+ implementation from VainF? It seems this ResNet implementation is missing dilate for _make_layer() function and Bottleneck unit, and missing replace_stride_with_dilation argument to work with DeepLabV3+.

Thanks for your code.I wonder why using softmax in groups * k_size * k_size instead of k_size * k_size? My understanding is that the kernel shape is k_size * k_size and there are "groups" kernels. So the sum of each kernel should be 1 instead of "groups" kernels

Hi,

When do you plan to release the Instance segmentation branch? I would be extremely grateful if it were made available.

Additionally, if you could also provide ResNet-50/101 COCO pre-trained weights, that would be great!

Hi,

Thank you for you work. I am slightly confused by the presence of these two parameters: 'groups' and 'pasa_groups'. The ResNet class in resnet_pasa_group_softmax.py can take both of them, so I was wondering which one refers to the number of channel groups (which will have different predicted LBP filters), and what is the other one used for. Judging from your README examples and the build_model() function in network.py, I believe that 'pasa_group' is the parameter to set the number of channel groups. Is that right?

Thanks you I advance,
Xavi