Skip to content

Discrepancy reported and reproduced results of linear eval #54

@WannesVanwinsen

Description

@WannesVanwinsen

Good afternoon,

python main_classifier.py --net s3d --dataset ucf101 --seq_len 32 --ds 1 --batch_size 32 --train_what last --epochs 100 --schedule 60 80 --optim sgd --lr 1e-1 --wd 1e-3 --pretrain ../feature/CoCLR-k400-rgb-128-s3d.pth.tar

After running the linear evaluation on UCF101 using the suggested parameters I obtained a final top-1 accuracy of 66.2 rather than the expected 74.5 reported in Table 2 row 9 of your paper. I saw that you expect a margin of 1-2% difference in these results, which is much less than what I am seeing myself. Are the hyperparameters that are provided in the repository the exact same as those used to obtain the results given in your Table 2? I am running the experiments using a single T4 GPU. Thank you in advance!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions