8000 Experiment hyperparameters · Issue #6 · git-disl/recap · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content 8000
Experiment hyperparameters #6
Open
@danielvhvs

Description

@danielvhvs

I had a question regarding the hyperparameters from the paper. Looking at the code there are some hyper parameters that I missed. For the experimental results in the paper, what are the hyperparameters you used for: The amount of iterations in init_pruning (num_pi) for the amount of total iterations over the full pruning finetuning loop (num_pr) for the amount of evaluation steps in the config (training_steps) the amount of training epochs per iteration in the loop in the config (num_train_epochs) and the batch size.

And what method do you use for looking at the memory size on gpu?
Do you use nvidia-smi or torch.cuda.max_memory_allocated()? Because there is no metric logger in the script.

There was a mention in the paper about an appendix. I got the paper from the ieee/cvf proceeding, which does not have an appendix, is there some alternative version that does have an appendix with the extra information about the experiments?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions

      0