Hi! I think there might be a detail missing in the code

[LAKawai]

In the React/thumos_dataset.py, and in the getitem () function, if it is training mode now, and when the self.provide_contrastive_data is True, we will get sample_gt as well as pos_sample_segment, neg_sample_segment and candidated_segments.

When we get pos_sample_segment, neg_sample_segment and candidated_segments, as the code shows:
pos_sample_segment = pos_sample_segment * (len(pos_feat) / self.clip_len)
neg_sample_segment = neg_sample_segment * (len(neg_feat) / self.clip_len)
candidated_segments = candidated_segments * (gt_len / self.clip_len)
Their start and end points are scaled to a value relative to 256.
However, the start and end of sample_gt are still relative to its own snippet_num.

Should a similar scaling operation be done for sample_gt here?

[dingfengshi]

Hi！Thank you for pointing this out! The sample_gt has already been scaled to its valid ratio in (see here). I will check the part of the code after the deadline of CVPR! Thank you very much!

[OpenSorceYCW]

hello, I meet some errors when installing the 1D Grid Sampling and RoI Align operators, the error message is 'roi_align/src/roi_align_kernel.cu:12:10: fatal error: ATen/ceil_div.h: No such file or directory
#include <ATen/ceil_div.h>
^~~~~~~~~~~~~~~~~
compilation terminated.
error: command '/usr/local/cuda-10.0/bin/nvcc' failed with exit status 1',
how to solve this problem? than you!

[dingfengshi]

hello, I meet some errors when installing the 1D Grid Sampling and RoI Align operators, the error message is 'roi_align/src/roi_align_kernel.cu:12:10: fatal error: ATen/ceil_div.h: No such file or directory #include <ATen/ceil_div.h> ^~~~~~~~~~~~~~~~~ compilation terminated. error: command '/usr/local/cuda-10.0/bin/nvcc' failed with exit status 1', how to solve this problem? than you!

hi! could you please try to uncomment the /roi_align/src/roi_align_kernel.cu in the line 7-9 and all other commented lines like: dim3 grid(std::min(THCCeilDiv((long)output_size, 512L), 4096L));
and comment the line 12-14 and all other lines which is below the commented line mention above, like: dim3 grid(std::min(at::ceil_div((long)output_size, 512L), 4096L));

That is because the pytorch1.9 replace the c++ operations from the THC library to the ATEN library.

After replace all operations, please setup the extension again.

I have not tried to build the code below the pytorch v1.9, so if you have any other questions, feel free to contact me. Thank you.

[guaguaxyf]

hello, I found some problems. My environment are python 3.7, torch 1.9.0+cu111, torchaudio 0.9.0 , torchvision 0.10.0+cu111, mmcv-full 1.7.1. When i was testing, it said

So I uninstall mmcv-full 1.7.1 ,install mmcv 1.4.0. But it said
which version of mmcv should i choose?

[dingfengshi]

hello, I found some problems. My environment are python 3.7, torch 1.9.0+cu111, torchaudio 0.9.0 , torchvision 0.10.0+cu111, mmcv-full 1.7.1. When i was testing, it said

So I uninstall mmcv-full 1.7.1 ,install mmcv 1.4.0. But it said which version of mmcv should i choose?

Hi, mmcv-full 1.4.0 is OK. We do not use ResNetTIN in our code. It seems that mmcv has not been installed correctly. Does the installed MMCV match the versions of CUDA and CUDNN? Or you can try to install mmcv-full from their source code.

[guaguaxyf]

thank you for your help. but i meet another problem, what should i do?