BinaryClassificationEvaluator returns np.float32() and np.float64()

[chr-werner]

I came across the problem where

output_scores[similarity_fn_name] = {
  "accuracy": acc,
  "accuracy_threshold": acc_threshold,
  "f1": f1,
  "f1_threshold": f1_threshold,
  "precision": precision,
  "recall": recall,
  "ap": ap,
}

is returning outputs in np.float32() and np.float64() which cause problems during model saving, as the json/encoder.py sees those types as not JSON serializable.

By changing the code snippet to

output_scores[similarity_fn_name] = {
  "accuracy": acc.item(),
  "accuracy_threshold": acc_threshold.item(),
  "f1": f1.item(),
  "f1_threshold": f1_threshold.item(),
  "precision": precision,
  "recall": recall.item(),
  "ap": ap.item(),
}

the elements get copied to a standard Python scalar each (note: precision already is) and then behave like expected during saving.

[tomaarsen]

Hello!

Well spotted, this is suboptimal. I'd rather have all of these converted to floats as expected. I'm tackling some other PRs right now, but I'll pick this up in a few days if someone else hasn't beat me to it by then. Thanks for reporting!

• Tom Aarsen

[JINO-ROHIT]

ill be happy to work on this :)