DeepAR - incremental learning

[NicolaiHarich]

Does DeepAR support incremental learning? Or in other words: Does it accept the 'model_uri' argument?

I want to update my model over time and I hope that training must not always start from scratch.

Update:

Apparently DeepAR does not support incremental training. Specifying a model as additional input channel leads to an error (shown below). This is really a pity and I can't understand why it's not supported. This is a standard scenario in deep learning....

Are there any plans to support incremental training for the future?

ValueError: Error for Training job DEMO-deepar-2019-03-24-17-01-52-934: Failed Reason: ClientError: Unable to initialize the algorithm. Failed to validate input data configuration. (caused by ValidationError)

Caused by: Additional properties are not allowed (u'model' was unexpected)

[lostella]

Hi @NicolaiHarich, as you noticed, incremental training is currently not supported. I will update this issue in case there are updates on supporting that training strategy.

[sujayramaiah]

Hi @lostella , Can you please advise if the incremental training supports DeepAR now?

[lostella]

@sujayramaiah incremental training is not supported unfortunately

[ShirleyMgit]

@lostella , is there any news in incremental learning in deepar? thanks

[lostella]

@ShirleyMgit there are no updates