Append, not overwrite `learning_stats.csv` when continuing training from checkpoint [minor request]

[guptadivyansh]

Is your feature request related to a problem? Please describe.
By looking at the (very useful) test metrics in learning_stats.csv, I decided to continue training my network for some more epochs. I wanted to plot the validation RMSE curves, but as soon as I restarted training, the csv only had the new epochs and the old metrics were gone.

Describe the solution you'd like
The CSV logger should just append more epochs as rows when training is continued, rather than creating (and thus overwriting) a 'new' file.

Describe alternatives you've considered
Parsing the lost metrics from the raw logs in 'train.txt'. Fortunately, 'train.txt' already opens the file in append mode, so the logging continues with retraining. The learning_stats.csv is obviously much nicer to parse though.

Additional context
dlc version 3.0.0rc13 using pytorch models.

[deruyter92]

@guptadivyansh, thanks for your request. We have been working on a fix for this issue (see PRs listed above this comment). Please try it out and let us know if it solves your problem.