[Question] Support for other loss functions.

[ydennisy]

This question is very broad and theoretical so apologies for that!

I would love to learn more about various loss functions which could be implemented in tfrs - when looking at other libs such lightfm, most will support a few loss functions.

I feel a tutorial on alternative loss functions would be excellent for example: https://github.com/tensorflow/addons/blob/master/tensorflow_addons/losses/triplet.py Using this function naively does not work - due to a data mismatch.

Thanks in advance!

[ydennisy]

@maciejkula sorry to bother - but I have seen you implemented other forms of loss in lightfm, would you be able to provide a little guidance how one could do the same for TFRS - also why does TFRS come with a single loss function out of the box, is this a design choice?

[maciejkula]

The Retrieval task accepts a loss function argument. You should be able to use your own loss function instead of the default softmax, as long as your happy accepting a num_queries x num_candidates logit and label matrix.

For more complex changes you could implement your own Task subclass, or even dispense with Task entirely.

The in-batch softmax setup is extremely robust, and works very well in most cases. This is why we don't provide extensive examples of other losses for this task.

[rlcauvin]

As an experiment implementing a custom loss function for a retrieval model, I implemented the following class, but I keep getting a "ValueError: No gradients provided for any variable” error when I fit the model.

import tensorflow as tf

class Top10CategoricalAccuracy(tf.keras.losses.Loss):

  def __init__(
    self, 
    name='top10_categorical_accuracy'):

    super().__init__(name=name)

  def call(
    self,
    y_true,
    y_pred,
    sample_weight = None):

    y_pred = tf.keras.backend.softmax(y_pred)
    top_10_accuracy = tf.keras.metrics.top_k_categorical_accuracy(y_true, y_pred, k=10)
    loss = 1.0 - top_10_accuracy

    return loss

Creation of the retrieval task looks like:

    loss_calculator = TopKCategoricalAccuracy(from_logits = True, top_k = 10) # Experimental custom loss function

    candidates = unique_candidate_ds.batch(128).map(lambda c: (c['item_id'], self.candidate_model(c)))
    metrics = tfrs.metrics.FactorizedTopK(candidates = candidates)
    batch_metrics = [tf.keras.metrics.AUC(from_logits = True, name = "retrieval_auc")]
    self.task: tf.keras.layers.Layer = tfrs.tasks.Retrieval(
      loss = loss_calculator,
      metrics = metrics,
      batch_metrics = batch_metrics)