Open
Description
Background and Feature Description
BatchNormalization output is sensitive to masking. Masking should be handled in the same way as RNN layers.
The mask tensor should be detected as an input of the BatchNormalization layer by the dependency detection mechanism to avoid the current exceptions raised when a Mask is declared for RNN layers. The same fix should be done for the RNN layers.
API Definition and Usage
public class BatchNormalizationOptionalArgs : IOptionalArgs
{
public string Identifier => "BatchNormalization";
public Tensor Mask { get; set; }
}
Alternatives
No response
Risks
Minimal