postech-cvlab / fastpointtransformer Goto Github PK

Hello, I would like to ask you how much memory and FLOPs your model runs on S3DIS, I would like to quote your paper, but there is something wrong with my computer, so I thought I would ask you directly!

Hi

The paper mentions an inference time extremely reduced compared to the original PointTransformer, but I was also curious about the time it took to train the model. What GPUs did you use and how much time did it take to train compared to PointTransformer ?

Thx a lot

Thanks for your amazing work, and I have few questions about the implemention of this architecture. Thank you for any answers.

1. In the code of LightweightSelfAttentionLayer, why the inter postion embeding is initialized as a learnable random variable ?

self.inter_pos_enc = nn.Parameter(torch.FloatTensor(self.kernel_volume, self.num_heads, self.attn_channels))
nn.init.normal_(self.inter_pos_enc, 0, 1)

According to Fig 3 in the paper, shouldn't it be obtained from the coordinate difference between the current voxel and neighboring voxels ?

2. How many specific neighboring voxels are indexed in LightweightSelfAttentionLayer ？Is the number of neighboring voxels determined by kernel_size in the input parameter? Is the neighboring voxels the valid voxels contained in the kernel ?

Thanks for the great work!
I would like to learn if the code for 3d object detection is released?

Hi,Thanks for your great work first,but I have some questions about the attention block.
For example, I set batch size = 2, how can i find the query voxels around both two frames.

Could you share some environmental information? I can successfully install pytorch according to setup.py, but there is a conflict when I install openblas-devel.

Thanks to the author for such quality code. I have some questions for the author,
in the sparse_ops.py file.
import cuda_sparse_ops
keeps reporting errors.
How to solve it? ? ?
And how to run setup.py in src.cuda_ops？？？？？？

An error is reported when training the dataset, saying "Error while calling W&B API: permission denied (<Response [403]>)". I checked the documentation of wandb and it says I don't have the project permissions. I want to know how to get the permission, thanks for your help!

Before calling the script preprocess_s3dis.py,I need to download the data manually, right? Or automatic download in the script preprocess_s3dis.py? If I need to download it automatically,How to download the dataset?

Hello, I 'd like to know about whether in the implementation, the stride shoud be kept as 1. What other things I may need to do if I want
to expand it to an arbitrary number.

Thanks for the great work! I have some questions.

1. I wonder where the GPU memory budget cost in FastTrans. Because I only have GTX1080, 12G. Do you test the model for small-scale task like ShapeNet DataSet? And can I apply the lightweight self-attention to get a better feature embedding for ShapeNet classification in my 1080 machine? Maybe I can use a smaller FastTrans?
2. I notice that the voxel size is used for data augmention. If i use the model for ShapeNet classification which normalized coordinates lie between -1 and 1. Can I remove the data augmention and just use original coordinates? Or how to get a feasible voxel size ?

What version of pytorch_lightning are you using? Thank you

It occurs when I start to train on S3DIS datset via the command "python train.py config/s3dis/train_fpt.gin"

I tried installing the module using pip but it said "No matching distribution found for cuda_sparse_ops", and I didn't find any solution on the Internet. Is it because there was something wrong with my installation?

result, COO, vals = MEB.coo_spmm_average_int32(
RuntimeError: at /FastPointTransformer/thirdparty/MinkowskiEngine/src/gpu.cu:100
I encountered this issue during runtime.I don't know how to solve it. Can you help me take a look

Hi @chrockey, great work!

Can you guide me on how to set up multigpu training? I have only 20GB gpus available, and when using batch size of 2 I obtain poor performance (~6% lower mIoU and mAcc; probably due to the batch norm and batch size).

If I add multigpu support (DDP) according to the example from the ME repository the learning is blocked, i.e. it never starts.

Any help will be appreciated. You commented "multi-GPU training is currently not supported" in the code. Have you had similar issues as I mentioned?

Thanks!

Hi!
When I read your paper at reducing space complexity section, it says that the space complexity of δrel(vi-vj) is O(KD),I can't understand it.
I think there are K neighbors for each voxel,why not O(IKD)? I hope you can help me ,Thanks!

Hellow! Thank you for your awesome work!
I find that it takes 20 hours on A100 to train your model. Could I have a look at your training log on S3DIS?