We currently save the pretrained_weights of the encoder and decoder in
two separate directories `encoder` and `decoder`. However, for the
`from_pretrained` function to operate with automodels we need to
specify the type of model in the path to the weights.
The path to the encoder/decoder weights is handled by the
`PreTrainedEncoderDecoder` class in the `save_pretrained` function. Sice
there is no easy way to infer the type of model that was initialized for
the encoder and decoder we add a parameter `model_type` to the function.
This is not an ideal solution as it is error prone, and the model type
should be carried by the Model classes somehow.
This is a temporary fix that should be changed before merging.
Custom schedulers are currently initiated by wrapping Pytorch's LambdaLR
class and passing a method of the wrapping class to the __init__
function of LambdaLR. This approach is not appropriate for several
reasons:
1. one does not need to define a class when it only defines a
__init__() method;
2. instantiating the parent class by passing a method of the child class
creates a cyclical reference which leads to memory leaks. See issues #1742 and #1134.
In this commit we replace the wrapper classes with functions that
instantiate `LambdaLR` with a custom learning rate function. We use a
closure to specify the parameter of the latter. We also do a bit of
renaming within the function to explicit the behaviour and removed
docstrings that were subsequently not necessary.
We currently initialize `encoder_attention_mask` when it is `None`,
whether the stack is that of an encoder or a decoder. Since this
may lead to bugs that are difficult to tracks down, I added a condition
that assesses whether the current stack is a decoder.
the data provided by Li Dong et al. were already tokenized, which means
that they are not compatible with all the models in the library. We
thus process the raw data directly and tokenize them using the models'
tokenizers.
We write a function to load an preprocess the CNN/Daily Mail dataset as
provided by Li Dong et al. The issue is that this dataset has already
been tokenized by the authors, so we actually need to find the original,
plain-text dataset if we want to apply it to all models.
In Rothe et al.'s "Leveraging Pre-trained Checkpoints for Sequence
Generation Tasks", Bert2Bert is initialized with pre-trained weights for
the encoder, and only pre-trained embeddings for the decoder. The
current version of the code completely randomizes the weights of the
decoder.
We write a custom function to initiliaze the weights of the decoder; we
first initialize the decoder with the weights and then randomize
everything but the embeddings.
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.
Rémi Louf