pytorch-forecasting icon indicating copy to clipboard operation
pytorch-forecasting copied to clipboard

Published 20 hours ago • verified publisher icon jdb78

Weights

Open polal2is opened this issue 4 years ago • 2 comments

Hi everyone, I believe it would be nice to clarify how to use weight in the TimeSeriesDataSet parameters.

I guess it is used to weight the output sequences and therefore give more importance to given prediction time steps ? Or is it weights to input samples e.g. to give more importance on earlier history ?

What is a good practice then to fill that 'weight' column ? Can it be monotonously increasing as e.g. data['weight'] = np.linspace(0, 1, len(data)) (case 2) or shall we treat every sequence/sample separately (case 1) ?

Thanks in advance!

polal2is avatar Sep 20 '21 10:09 polal2is

Is there any news? I would also be interested in the parameter 'weight' in TimeSeriesDataSet. Could it be used to give more weight to one class in an imbalanced dataset?

thommykuster avatar Apr 02 '22 16:04 thommykuster

An easy way to use sample weights would be very useful indeed (for example, maybe short horizons are more important in a real world use case).

GeorgeG92 avatar Sep 02 '22 14:09 GeorgeG92