Skip to content

negative KL #28

@mjleone

Description

@mjleone

Hello,

I see your warning about negative KL divergence error for certain versions of tensorflow. Do you know of a version that avoids this, or of any other way to avoid it? I have the version in your yaml 2.7.0.

Also wondering how to overcome being stuck at a certain loss (in my case 86k, close to the starting 93k). This may be the averaging effect you note, but it occurs for any hyperparam combos I try for my data of ~6,000 sequences.

Thank you,

Mike

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