resize_token_embeddings not taken into account in save_pretrained for EncoderDecoderModel

[rahular]

Environment info

• transformers version: 4.5.0
• Platform: Darwin-17.7.0-x86_64-i386-64bit
• Python version: 3.7.2
• PyTorch version (GPU?): 1.4.0 (False)
• Tensorflow version (GPU?): not installed (NA)
• Using GPU in script?: no
• Using distributed or parallel set-up in script?: no

Who can help

@patrickvonplaten, @patil-suraj

Information

I am extending the embeddings of the decoder of an EncoderDecoderModel model. When I save it, the config does not reflect the new size. However, it works fine when I try doing the same for non EncoderDecoderModel models.

To reproduce

In [1]: model = t.EncoderDecoderModel.from_encoder_decoder_pretrained('bert-base-uncased', 'bert-base-uncased')

In [2]: model.decoder.bert.embeddings.word_embeddings
Out[2]: Embedding(30522, 768, padding_idx=0)

In [3]: model.decoder.resize_token_embeddings(30522+100)
Out[3]: Embedding(30622, 768)

In [4]: model.save_pretrained('test-bert')

Expected behavior

The updated embedding size should be saved in config.json

[cronoik]

This is caused by the EncoderDecoderConfig which initializes independent objects (link) instead of utilizing the already existing ones.

You can fix that for the moment by calling:

model.config.decoder = model.decoder.config
model.config.encoder = model.encoder.config

PR will follow.