Gluon-nlp: ConvolutionalEncoder: contradictory description of parameter `conv_layer_activation`

[bikestra]

https://github.com/dmlc/gluon-nlp/blob/ca071a6dfd7ad729a50127d1e43afe417b3230b6/gluonnlp/model/convolutional_encoder.py#L63-L66

The docstring says that if nothing is specified for conv_layer_activation, no activation will be applied; however, the default is tanh, so tanh will be applied if nothing was specified. I don't think Kim et. al paper applies any activation, so the default value of this parameter should be linear in order to be consistent with the rest of the documentation.

Maybe I misinterpreted the docstring, but it's at least confusing.

[bikestra]

Sorry, Kim et al do use tanh.

[leezu]

https://github.com/yoonkim/lstm-char-cnn/blob/master/model/TDNN.lua#L27

[bikestra]

Yes, they use tanh, but then shouldn't the description of the default behavior be changed?

[leezu]

Thanks for pointing out the docstring issue and no worries regarding tanh. I agree we should change the docstring. Do you want to open a PR to change it?
If so, you may also wish to change

 seq.add(nn.Activation(conv_layer_activation))

to

if conv_layer_activation is not None:
    seq.add(nn.Activation(conv_layer_activation))

or similar in
https://github.com/dmlc/gluon-nlp/blob/ca071a6dfd7ad729a50127d1e43afe417b3230b6/gluonnlp/model/convolutional_encoder.py#L117

[bikestra]

OK, I will open a PR.