add keras callback

[lvwerra]

This PR adds a KerasCallback class which lets a user wrap an evaluation in a keras compatible wrapper. Following the discussions in #10 (comment).

import evaluate

recall = evaluate.load("recall")

transform = lambda x: np.argmax(x, axis-1) # labels are 1-hot encoded but recall needs integer representation

recall_callback = evaluate.KerasCallback(
    model, # model object
    x_test, # model input
    recall, # metric object
    {"references": transform(y_test), "average":"micro"}, # metric input
    predictions_processor=transform # transform between model ouput/metric inputs
    ) 

model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])

hist = model.fit(
    x_train,
    y_train,
    batch_size=batch_size,
    epochs=epochs,
    validation_split=0.1,
    callbacks=[recall_callback],
)

If there are multiple metrics we can either create several callbacks or combine them all with combine into a single metric. There are still a few namings/rough edges to iron out but I would be interested in some early feedback.

cc @lhoestq @sashavor

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[lhoestq]

Nice ! Cool that it's similar to https://keras.io/examples/keras_recipes/sklearn_metric_callbacks/ :)

[lvwerra]

Yep, I took that as inspiration :) I'll flesh out the PR for review.

[Rocketknight1]

We also have a KerasMetricCallback in transformers - but it might make sense to deprecate that in favour of a class in evaluate!

I definitely think having an easy way to wrap evaluate metrics in a Keras-compatible way would be really useful, and would let me get rid of a lot of boilerplate in all my example code.

[lvwerra]

Thanks @Rocketknight1! That looks pretty much like what we want. Maybe we can/need change two things:

1. Pass an evaluate metric instead of a callable and add use the prediction_processor as the piece between model and metric.
2. We want evaluate to be as framework (transformers) agnostic as possible. The main issue I see is integrating the generate related functions as they are not native to Keras and make quite a lot of use of transformers logic. Similar for padding across batches.

What are your thoughts on this?