RuntimeError: "batch_norm" not implemented for 'Char'; error while trying to pass int8 input for model eval

[Atman-Kar]

Find a way to allow batch_norm layer to accept int8 type input.

Few options we have:

• Do some data conversion type work-around
• Remove the batch_norm layer altogether
• Create a custom batch_norm layer for binarized neural networks to replace with

[hades-rp2010]

The conversion sounds like the easiest fix.
We convert the inputs to the batchnorm layers to a float, reconvert them back to int8 post batch norm.
Tho I am not aware of what this means in terms of training or correctness.

[Atman-Kar]

If we go through the conversion route, we would have to make custom layer called BunnyBatchNorm2D which:

• First takes in int8 input and converts them to float32
• Does a normal batch normalization
• The output batchnorm is then sign clipped and pushed out as int8 for the next layer to process.

Should we do it this way?

Note: A similar error occurs with MaxPool2D. However, Conv2D and Linear layers work perfectly with int8 input types.