conradry/max-deeplab

Unofficial implementation of MaX-DeepLab for Instance Segmentation

MaX-DeepLab

Unofficial implementation of MaX-DeepLab for Instance Segmentation: https://arxiv.org/abs/2012.00759v1.

Status

Only the MaX-DeepLab-S architecture is putatively implemented. Primarily, this code is intended as a reference; I can't make any guarantees that it will reproduce the results of the paper.

• Axial Attention block
• Dual Path Transformer block
• MaX-DeepLab-S architecture
• Hungarian Matcher
• PQ-style loss
• Auxiliary losses (Instance discrimination, Mask-ID cross-entropy, Semantic Segmentation)
• Coco Panoptic Dataset
• Simple inference

Usage

See example.ipynb.

max-deeplab's Issues

RuntimeError: size of dimension does not match previous size, operand 1, dim 1

In block file
qr = torch.einsum('bnci,cij->bnij', q, qemb) line 173, in forward

How can I solve this error?

Drop Path in the Paper

Hello, thanks for sharing this work! I have a question, do you implement the drop path in backbone as mentioned in the paper? Thank you.

Ignored classes and nestedTensor

Hi,

Thank you for the implementation. I have two questions regarding the code.

1. What is the significance of converting list of Tensors to nestedTensor in dataloader?
2. Are class labels for the background 0 and 201 ignored in the loss function?

what does the memory path mean?

what does the memory path mean? Is that read the original picture pixels from the memory?How does it work?

How about the performance with this code?

Could you provide it?

code

can you share your code and result ?

about the flops

Are the M-adds you reported in your paper equivalent to FLOPs？