About fine-tuning times

[dafmdev]

I did a test by running the fine-tuning process with 20 time series as in the example and with 1_000_000 time series. Both processes took less than 8 minutes.

Shouldn't the fine-tuning process take longer with more data?

[lostella]

The number of training steps, set here https://github.com/amazon-science/chronos-forecasting/blob/9d59057b72a1ddd52e163c536d0f99631eea4857/scripts/training/train.py#L510  is independent of the size of the dataset. For 1M series you probably want to set it to a value that, together with the batch size, allows the training loop to go over the dataset a sufficient number of times (“epochs”).