In h3, we should pass batch_first=True to TransformerEncoderLayer

[junjiechen-chris]

As described in the title.
Torch by default uses batch_first=False for the TransformerEncoderLayer, resulting in high training loss and test error.

Using default batch_first=False

Passing batch_first=True

[Exorust]

Totally valid. Would you be interested in raising a PR to patch this? :)