TFT fixes - Githubissues

zalandoresearch / pytorch-ts

PyTorch based Probabilistic Time Series forecasting framework based on GluonTS backend

MIT License

1.24k stars 191 forks source link

TFT fixes #55

Closed aslinagy closed 3 years ago

aslinagy commented 3 years ago

I fixed a couple of issues regarding the TFT model, namely:

there was an issue around dynamic features as AgeFeature was not accounted for in the feature dimensions mapping
dynamic feature projection needed a separate parameter for feature slices
changed transformations so categorical features are converted to np.long

kashif commented 3 years ago

@aslinagy very cool! I'll review and merge shortly!

aslinagy commented 3 years ago

@kashif Thanks for merging my fix. May I ask when your planning to make a new release? I'm using your repo as a dependency for a project and an installable package with the fix would be greatly appreciated but I can manage if you are not planning to make a new release just now.

Thanks

kashif commented 3 years ago

@aslinagy yes sure I can make a new release... there is one annoying issue that the tqdm outputs are screwed up ever since I merged the PR to support evaluation datasets.... do you think you can have a quick look at that? If we can fix that I can make a release...

aslinagy commented 3 years ago

@kashif yea I noticed that at the beginning as well but I haven't touched that part and I'm actually using a custom trainer and with that, it worked just fine so I didn't give it a thought. I will have a look and try to find what could have caused it.

kashif commented 3 years ago

awesome!! thank you! Else we can even revert that PR...

aslinagy commented 3 years ago

Hi, I had a look but I haven't found anything obvious so I need a couple of days to find the time to properly trace it back. In the meanwhile feel free to revert my fix.

kashif commented 3 years ago

ok no problem... let me have a look...