Precision/recall scores not applicable to dropped samples

[bfhealy]

The Dataset class in scope/utils.py allows for samples to be dropped to balance positive and negative examples for a class. In the training process, the code attempts to evaluate an f1 score for the dropped samples (full code below):
https://github.com/ZwickyTransientFacility/scope/blob/ac22f211a188eb1736e35eb9c6c552a902735fb5/scope.py#L665

if datasets["dropped_samples"] is not None:
            # log model performance on the dropped samples
            if verbose:
                print("Evaluating on samples dropped from the training set:")
            stats = classifier.evaluate(datasets["dropped_samples"], verbose=verbose)
            if verbose:
                print(stats)

            if not kwargs.get("test", False):
                for param, value in zip(param_names, stats):
                    wandb.run.summary[f"dropped_samples_{param}"] = value
                p, r = (
                    wandb.run.summary["dropped_samples_precision"],
                    wandb.run.summary["dropped_samples_recall"],
                )
                wandb.run.summary["dropped_samples_f1"] = 2 * p * r / (p + r)

Since the dropped samples only contain labels from the underrepresented class, the precision p and recall r will be sometimes be zero, leading to a divide by zero error and a 'failed' run in wandb. A quick solution is to remove the f1 score from the metrics computed for the dropped samples.