Creates LogLossMetric and generalizes WandbLogger and ExamplesPerSecondCallback

[gabrielspmoreira]

Goals ⚽

This PR moves some utility classes from the Quick-start for Ranking scripts to Merlin Models, as they are generic and are useful in general for building training scripts with the library: LogLossMetric, WandbLogger, ExamplesPerSecondCallback

Implementation Details 🚧

• LogLossMetric: Log loss (binary cross entropy( is a common loss and metric for binary classification. This class is useful when your are using class weights for binary classification, as in that case the loss for training will be influenced by the class weights. Having this metric allows you to have the unweighted log loss metric.
• WandbLogger - This class manages the logging to Weights&Biases, providing methods to initialize wandb, configure the run, log metrics and to return the tf.keras.WandbCallback() that can be used with Keras fit() and evaluate().
• ExamplesPerSecondCallback - Logs the training or evaluation throughput, i.e. steps/sec (to wandb or to console)

Testing Details 🔍

• Adds a test for LogLossMetric: test_logloss_metric
• Changed retrieval integration tests to use the generalized WandbLogger and ExamplesPerSecondCallback

[github-actions]

Documentation preview

https://nvidia-merlin.github.io/models/review/pr-1085