Problem in reimplement m4 result in origin parper about n-beats HOT 10 CLOSED

Fixed in fb1f612

from n-beats.

[UPDATE] I find the origin paper mentioned that the stop gradient in denominator, but I dont see this trick in his code and that may result in relative bad performance. I will keep investigating it.

from n-beats.

The linear space function should use an adapted horizon,

def linear_space(backcast_length, forecast_length, is_forecast=True):
    # ls = K.arange(-float(backcast_length), float(forecast_length), 1) / forecast_length
    # return ls[backcast_length:] if is_forecast else K.abs(K.reverse(ls[:backcast_length], axes=0))
    horizon = forecast_length if is_forecast else backcast_length
    return K.arange(0,horizon)/horizon

from n-beats.

@fecet thank you very much!! Can you detail a bit more what you found? Is it just changing the linear_space and it works?

from n-beats.

And yes please post the full code it's very interesting!

from n-beats.

@fecet thank you very much!! Can you detail a bit more what you found? Is it just changing the linear_space and it works?

Yes, in former code, the denominator is always forecast_length, when backcast_lenghth is large, like 7*forecast_length the value in linear_space will reach up to 7, result in output explode in trend block since it will be multiplied by 7*p several times. And smape_loss's own property make it become 200 as y_pred dominate.

Thanks for your interest, I will create a repository for my reimplement later. BTW, I found this behaviour by your keract tool!

from n-beats.

@fecet Very happy to hear that!! Please post your results and code. Thank you for using keract. I'm happy it helps ;)

from n-beats.

Here is my repo, any comments would be highly appreciated.

from n-beats.

thank you so much I'm going to have a look shortly

from n-beats.

I added a link to your repo in the README btw

from n-beats.