Skip to content

Batch Normalization Bug #28

@SpectralGuilem

Description

@SpectralGuilem

Hi I'm reimplementing the ladder and tagger networks in TensorFlow and have found what I believe to be a (minor) bug. Could you please clarify this?

Batch normalization (BN) parameters in the decoder (lines 488-494 in ladder.py) are not annotated as having role BNPARAM. As such, they are not replaced in the graph for the training set statistics by TestMonitoring._get_bn_params. If one attempts to evaluate the model after training with very small batch sizes performance is degraded because the mean and variance for the BN step is computed from a very small sample.

Thanks,
Guillem

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