Skip to content

The validation set is way easier than the training set? #2

@Le-Anh-Duy

Description

@Le-Anh-Duy

Problems

I have experiemented with your dataset (Gym99) on my latest project, I found that on training, the training loss alway lower than the validation loss. I trained with ST-GCN (with some modified and model and hyperparameter but i dont think that the main cause). I pastes my training history here:

Image

My setting

Use focal loss, train for 160 epochs, use adam optimizer, use consine anealing scheduler (i know using adam + that scheduler is dumb, but it's too late)

Question

  • Does anyone have this issues?
  • Is it because of my setting?

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