Official Implementation for "Mask-Attention-Free Transformer for 3D Instance Segmentation"

Shell 0.09% Python 81.11% C++ 9.93% Cuda 7.13% C 1.74%

mask-attention-free-transformer's People

Contributors

Stargazers

Watchers

Forkers

cv-seg yangxin6 daoguizhang whuhxb xijunke

mask-attention-free-transformer's Issues

Question about mask in cross attention component

Hi @X-Lai , Thanks for sharing this great work!

What is the purpose of attn_masks in your transformer decoder? In your paper, you mentioned that mask-attention-free-transformer.

Mask-Attention-Free-Transformer/maft/model/transformer.py

Lines 108 to 115 in 4b5048c

 # get mask 

 pred_masks = torch.einsum('nbd,mbd->bnm', output_norm, mask_feats_batched) #[bsz, num_queries, max_length] 

 attn_masks = (pred_masks.sigmoid() < self.attn_mask_thresh).bool() #[bsz, tgt_len, src_len] 

 for b in range(lengths.shape[0]): 

 length = lengths[b] 

 attn_masks[b, (attn_masks[b, :, :length].sum(-1) == length)] = False 

 attn_masks[b, :, length:] = True 

 attn_masks = attn_masks.unsqueeze(1).expand(-1, self.nhead, -1, -1).contiguous().flatten(0,1)

Thank you.

What is the license for the code?

Hi,

I'm interested in potentially building on some of your code. Can you please clarify what usage license this repository is released under?

Question about the pretrained weights in training

I observed that you utilized the pretrained weights from SSTNet during the training process. However, the SSTNet model's original input does not encompass normal information. In the context of incorporating normal information into the model, the question arises: which pretrained weights should be employed post the integration of normal information?

What is CUDA and Torch version?

Can you provide more details regarding conda environment creation please?

Reproduction of the experimental results in Fig. 1

Hi, how can I reproduce the experimental results shown in Fig.1? I tried modifying the epoch parameter in the configs directly, but the results were significantly different. Can you assist me with this issue?

Questions about the training loss and backbone configuration

In loss.py, I think a score loss is also used to train the model but it is not mentioned in the paper.
Can you provide some insight into this? I might be overlooking something...

Additionally, is there a specific reason for using the Minkowski engine as a backbone in the S3DIS dataset?"

Great job! May I ask if you could provide the code for your work on the ScanNet200 and S3DIS datasets?

Testing MAFT for ScanNetv2 and ScanNet200 with Mask3D?

Hello!

Thanks for this great job! In this paper, you have tested MAFT for ScanNetv2 and ScanNet200 with SPFormer. Have you ever tested MAFT for ScanNetv2 and ScanNet200 with Mask3D? Thanks!

Best

How to generate superpoints for S3DIS

Dear authors,

Thanks for your job. My question is that are there any methods or repos can be referred to generate superpoints for S3DIS. And is there any plan to release the training code for S3DIS?

Best

Cannot import name 'colors_cityscapes' from 'maft.utils.visualize'

I think your code is not complete because some modules are missing. Could you tell me when you will complete it?

The normal information in training

Hi, XinLai, thank you for you great work.

I found that the code supports the sstnet pretrain weight, if I want to train with the normal information, could the pretrain backbone be released, or the backbone is trained the same as sstnet, just need to change the input channel into 9?
I'd like to know more about the S3DIS training too.
Thank you!

A question about environment configuration

Mask-Attention-Free-Transformer/lib/attention_rpe_ops/setup.py

Line 7 in 4b5048c

(opt,) = get_config_vars('OPT')

# install attention_rpe_ops
cd lib/attention_rpe_ops && python3 setup.py install && cd ../../

When I run the command, I was prompted that opt is null.

	# get mask
	pred_masks = torch.einsum('nbd,mbd->bnm', output_norm, mask_feats_batched) #[bsz, num_queries, max_length]
	attn_masks = (pred_masks.sigmoid() < self.attn_mask_thresh).bool() #[bsz, tgt_len, src_len]
	for b in range(lengths.shape[0]):
	length = lengths[b]
	attn_masks[b, (attn_masks[b, :, :length].sum(-1) == length)] = False
	attn_masks[b, :, length:] = True
	attn_masks = attn_masks.unsqueeze(1).expand(-1, self.nhead, -1, -1).contiguous().flatten(0,1)

dvlab-research / mask-attention-free-transformer Goto Github PK

mask-attention-free-transformer's People

Contributors

Stargazers

Watchers

Forkers

mask-attention-free-transformer's Issues

Recommend Projects

Recommend Topics

Recommend Org