Minor improvements to the Embedder

[LeonardoEmili]

The following changes have been applied:

• Add support to average the last four hidden layers of the transformer model
• Add a shape property to the TransformersEmbedderOutput class (referenced in the README)
• Add option to specify which hidden states to use for pooling
• Update the docs accordingly
• Run black in compliance with the project specifications
• Fix documentation issues

[Riccorl]

I suggest to change

def __init__(
        self,
        model: Union[str, tr.PreTrainedModel],
        return_words: bool = True,
        pooling_strategy: str = "last",
        output_layers: List[int] = [-4, -3, -2, -1],
        fine_tune: bool = True,
        return_all: bool = False,
    ) -> None:
        super().__init__()

to

def __init__(
        self,
        model: Union[str, tr.PreTrainedModel],
        return_words: bool = True,
        pooling_strategy: str = "last",
        output_layers: List[int] = None,
        fine_tune: bool = True,
        return_all: bool = False,
    ) -> None:
        super().__init__()
        if output_layers is None:
             output_layers  = [-4, -3, -2, -1]

to pass the checks.

[LeonardoEmili]

I suggest to change

def __init__(
        self,
        model: Union[str, tr.PreTrainedModel],
        return_words: bool = True,
        pooling_strategy: str = "last",
        output_layers: List[int] = [-4, -3, -2, -1],
        fine_tune: bool = True,
        return_all: bool = False,
    ) -> None:
        super().__init__()

to

def __init__(
        self,
        model: Union[str, tr.PreTrainedModel],
        return_words: bool = True,
        pooling_strategy: str = "last",
        output_layers: List[int] = None,
        fine_tune: bool = True,
        return_all: bool = False,
    ) -> None:
        super().__init__()
        if output_layers is None:
             output_layers  = [-4, -3, -2, -1]

to pass the checks.

Wouldn't it mean loosing default value suggestions on the IDE? What about passing the layers as a tuple instead of a list?

[Riccorl]

Let's try with the tuple first.