Description

The current implementation of AutoGluon's time series models, specifically those using the GluonTS torch backend, automatically deletes the lightning_logs directory after each training run. This directory contains logs that are essential for users who utilize TensorBoard to monitor and compare different training sessions. The automatic deletion of these logs makes it difficult to use TensorBoard effectively, as it relies on historical log data for comparison.

Proposal

I propose a feature for the timeseries module where the deletion of the lightning_logs directory is made optional. This could be implemented by adding a parameter to the model training functions and the fit() method, allowing users to choose whether to preserve the logs. By default, this parameter could be True to maintain the current behavior, but when set to False, it would keep the logs intact for further analysis with TensorBoard.

Code

this code is responsible for deletion of the logs:
if lightning_logs_dir.exists() and lightning_logs_dir.is_dir(): logger.debug(f"Removing lightning_logs directory {lightning_logs_dir}") shutil.rmtree(lightning_logs_dir)