Allow PredictionWriter to create different kinds of saved output

[sjfleming]

Perhaps write_prediction could be an input argument, and maybe you could specify a different writer function for different use cases? Not sure if this could work.

Now we write batches as separate CSV output files with write_prediction()

cellarium-ml/cellarium/ml/callbacks/prediction_writer.py

Lines 14 to 38 in ca4d6f1

	def write_prediction(
	prediction: torch.Tensor,
	ids: np.ndarray,
	output_dir: Path | str,
	postfix: int | str,
	) -> None:
	"""
	Write prediction to a CSV file.
	Args:
	prediction:
	The prediction to write.
	ids:
	The IDs of the cells.
	output_dir:
	The directory to write the prediction to.
	postfix:
	A postfix to add to the CSV file name.
	"""
	if not os.path.exists(output_dir):
	os.makedirs(output_dir, exist_ok=True)
	df = pd.DataFrame(prediction.cpu())
	df.insert(0, "db_ids", ids)
	output_path = os.path.join(output_dir, f"batch_{postfix}.csv")
	df.to_csv(output_path, header=False, index=False)

Is it possible to do something like keep appending to the same h5 file, for instance? Perhaps PredictionWriter could be made to accommodate other kinds of output writers.

[sjfleming]

Maybe the more sensible thing to do is have a separate PredictionWriter, like PredictionWriterH5, for that kind of thing