Skip to content

Checkpointer Saves Model when it outputs NaN #64

Description

@sciapponi

When I'm training a model and gradients explode/vanish Micromind still saves it to the checkpoint folder.
Personally I'd like the framework to keep the last working version since it could be useful to identify issues about the weight computation.

image

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

bugSomething isn't working

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions