GitLab

Since the preloading of weights relies on the name of the class's
attributes changing the namespace breaks loading pretrained weights on
Bert and all related models. I reverted `self_attention` to `attention`
and us `crossattention` for the decoder instead.

In the seq2seq model we need to both load pretrained weights in the
encoder and initialize the decoder randomly. Because the
`from_pretrained` method defined in the base class relies on module
names to assign weights, it would also initialize the decoder with
pretrained weights. To avoid this we override the method to only
initialize the encoder with pretrained weights.

I am not sure what happens when the class is initialized with the
pretrained weights.

The modifications that I introduced in a previous commit did break
Bert's internal API. I reverted these changes and added more general
classes to handle the encoder-decoder attention case.

There may be a more elegant way to deal with retro-compatibility (I am
not comfortable with the current state of the code), but I cannot see it
right now.

There is currently no way to specify the quey, key and value separately
in the Attention module. However, the decoder's "encoder-decoder
attention" layers take the decoder's last output as a query, the
encoder's states as key and value. We thus modify the existing code so
query, key and value can be added separately.

This obviously poses some naming conventions; `BertSelfAttention` is not
a self-attention module anymore. The way the residual is forwarded is
now awkard, etc. We will need to do some refacto once the decoder is
fully implemented.

Several packages were imported but never used, indentation and line
spaces did not follow PEP8.

Fixed critical css font-family issues

Add option to use a 'stop token'

Added ValueError for duplicates in list of added tokens

add DistilBert model shortcut into ALL_MODELS