gabeur
commented
3 years ago Indeed the ‘HowTo100M_full_train.pth’ I pre-trained after fixing Fix1 and Fix3 does not perform as well as before on ActivityNet as optimal hyperparameters seem to have changed.
I have experimented with a lower batch size (24) and obtained the following results:
command: python -m train --config configs_pub/eccv20/ActivityNet_val1_trainval_bs.json --load_checkpoint data/checkpoints/HowTo100M_full_train.pth
ActivityNet_val1_test:
t2v_metrics/R1/final_eval: 27.191376855806386
t2v_metrics/R5/final_eval: 59.42647956070775
t2v_metrics/R10/final_eval: 73.68314012609315
t2v_metrics/R50/final_eval: 94.30547081553793
t2v_metrics/MedR/final_eval: 4.0
t2v_metrics/MeanR/final_eval: 16.801098230628433
t2v_metrics/geometric_mean_R1-R5-R10/final_eval: 49.19562050180819
v2t_metrics/R1/final_eval: 27.293064876957494
v2t_metrics/R5/final_eval: 59.24344112263575
v2t_metrics/R10/final_eval: 73.3577384584096
v2t_metrics/R50/final_eval: 93.49196664632906
v2t_metrics/MedR/final_eval: 4.0
v2t_metrics/MeanR/final_eval: 18.19646125686394
v2t_metrics/geometric_mean_R1-R5-R10/final_eval: 49.1337040926001
Thank you for your generous sharing of your wonderful research results! But I have some problems in reproducing the results.
I found that using the ‘HowTo100M_full_train.pth’ you provided can not achieve the desired results on ActivityNet dataset. What should I do if I want to finetune on ActivityNet dataset? Is there any chance that you release the ‘HowTo100M_full_train.pth’ for ActivityNet dataset sometime?
Thank you!