Open TheaperDeng opened 2 years ago
"bigdl.chronos.forecaster.tf2": should we use "tf2" or just "tf" as nano ? For now we don't intend to support tf1. And even if we support tf1 in future we can make tf using tf 2.0 as default and use tf1 namespace for back compatibility.
"bigdl.chronos.forecaster.tf2": should we use "tf2" or just "tf" as nano ? For now we don't intend to support tf1. And even if we support tf1 in future we can make tf using tf 2.0 as default and use tf1 namespace for back compatibility.
Yeah, I have no strong opinion on tf or tf2. I choose tf2 just to align with orca. I think we can use tf for now.
"bigdl.chronos.forecaster.tf2": should we use "tf2" or just "tf" as nano ? For now we don't intend to support tf1. And even if we support tf1 in future we can make tf using tf 2.0 as default and use tf1 namespace for back compatibility.
Yeah, I have no strong opinion on tf or tf2. I choose tf2 just to align with orca. I think we can use tf for now.
@jason-dai what do you say?
- We need to ensure there won't be errors if user have only pytorch or tensorflow installed. - @jason-dai , do you still think a global flag is need to enable only pytorch or tensorflow models? maybe we don't need that anymore.
- We need install dependencies chronos[pytorch] chronos[tensorflow]. Together with lite and all, we may need pytorch-lite, tf-lite, pytorch-all, tf-all, etc.
I think we may have these installation option:
For bigdl-chronos
:
For bigdl-chronos-lite
:
Import
Here we take
LSTMForecaster
as an example, same things will happen onLSTMForecaster
,TCNForecaster
, andSeq2SeqForecaster
in the first round.Interface
I had a quick review of all APIs we have for pytorch and don't find any of them can not be supported in TF2. With the help of
bigdl.nano.tf
andbigdl.orca.learn.tf2
. We should be able to support exactly the same API as pytorch.