Test on kinetics 400

[manhcntt21]

Hi, thank for your implementation.

I have read your paper, and now just want to run test, I followed the steps here

I downloaded pre-trained on link

According to my limited knowledge, run pre-trained will not need dataset, is that true?

Some changes I made:

1. I created path_to_models folder inside video_classification folder
2. I changed _C.DIST_BACKEND = "gloo", because I'm windows user (nccl -> gloo)
3. I changed NUM_GPS 1 in test.sh and test.yaml
4. PRETRAIN_NAME: 'uniformer_small_k400_8x8' on test.yaml

I changed some config and run 'bash ./exp/uniformer_s8x8_k400/test.sh, error in your_path_to_data

So, my question is there a way to run test without downloading the data kinetics 400?, because my computer does not enough memory.

I would appreciate it if you have detailed instruction.

Thank you so much

[Andy1621]

Thanks for your question! What do you want to do for running the test?

• If you want to check the model, I suggest you use the model code in hugging_face
• If you want to check the codebase, you can create a tiny dataset with a few videos. You only need to prepare some CSV files as here. Then, you can change the data path in shell/config file.

[manhcntt21]

Thanks for your reply, I'm a graduate student, so I want to check the codebase and then conduct my own experiments. Thanks for your advice, I will try and provide my results for you.

[manhcntt21]

Hi, from the log on my terminal, i see the first file that I don't have "wXXiG-5cTX0.mp4".

I try to find it in test.csv and I saw it.

So, I just want to download this video to test, what should I do.

[manhcntt21]

To continue the learning process, I download a video on kinetics-400 test in https://s3.amazonaws.com/kinetics/400/test/part_0.tar.gz

And, then I insert its name into the file test.csv as below, number 1 is random number I added, because I don't know label what its label is

what should I do next? to view result of test. This is stdout.log

[Andy1621]

You have correctly run the codebase! Since your want to conduct your experiments, the first step is to communicate with your advisor about a research topic.

If you want to apply UniFormer for a specific domain, you may need to prepare your own dataset and fine-tune the pretrained models we provided!

[manhcntt21]

Thank you very much!