search

buxiangzhiren / ContextLoc

Code for the paper "Enriching Local and Global Contexts for Temporal Action Localization", ICCV 2021
Apache License 2.0
27 stars 2 forks source link

Only one result file #1

Open html4396 opened 3 years ago

html4396 commented 3 years ago

Hello, I want to get the two-stream result, you said in readme that i need to modify the result of rgb to rgb_result and the result of flow to flow_result, but I only get a result file.

buxiangzhiren commented 3 years ago

Hi, you should change the "train_ft_path" and "test_ft_path" in the "./data/dataset_cfg.yaml" to the other stream features. Then you can get the second result. In addition, you should retrain and test the model again.

html4396 commented 3 years ago

which checkpoint should I use during the two-stream test

buxiangzhiren commented 3 years ago

If you use the RGB features to train model and get checkpoints, you will obtain the RGB result in testing with these checkpoints. If you use the Flow features to train model and get checkpoints, you will obtain the Flow result in testing with these checkpoints. You can get two results. Then you should put them in the "results" folder.

html4396 commented 3 years ago

Thank you very much

buxiangzhiren commented 3 years ago

You're welcome, I hope these are useful to you.

html4396 commented 3 years ago

I use 4 GPUs to run this model, but its occupancy rate is very low(only 7%~17% per GPU). Can i modify its batchsize, if it is changed, will it affect its performance?

buxiangzhiren commented 3 years ago

I have not tried to change the batchsize because I need to keep the same setting as PGCN, so I do not know the performance.

html4396 commented 3 years ago

Thanks very much for answering my questions