Skip to content

Why not detach the hidden state of GRU from the computational graph? #161

Description

@MejiroSilence

In RNNs, gradients accumulate over time steps. If the sequence is long, gradients can become very large (exploding gradients) or very small (vanishing gradients), leading to unstable training or difficulty in convergence. Detaching the hidden state can limit gradient propagation within each time step, preventing gradient accumulation over the entire sequence, thus mitigating exploding/vanishing gradient problems.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions