search

wassname / attentive-neural-processes

implementing "recurrent attentive neural processes" to forecast power usage (w. LSTM baseline, MCDropout)
Apache License 2.0
90 stars 23 forks source link

ANP-RNN 'use_deterministic_path'= False? #2

Open xuesongwang opened 4 years ago

xuesongwang commented 4 years ago

Hi there,

I am just wondering if you forgot to set 'use_deterministic_path'= True for ANP-RNN, since in the paper the authors indicate so, and you obviously have already set cross-validations.

https://github.com/3springs/attentive-neural-processes/blob/af431a267bad309b2d5698f25551986e2c4e7815/neural_processes/models/neural_process/lightning.py#L189-L195

BTW, I was trying to replicate the experiment on GP dataset, and I have already implemented an ANP. Although the architecture of mine is slightly different from yours, I assume I will just need to replace all the MLPs with LSTMs plus sequential encodings of the input, output, right? However, mine result was really bad, when I sorted my x_context, x_target, the model only seems to be able to predict a few points and predicts flat curves afterwards. Could you please share with me any hints regarding this?

Your help is very much appreciated

wassname commented 4 years ago

Your right it probably should be true, although I'm not sure the Deterministic Path helps much.

Could you please share with me any hints regarding this?

Sure, I may need to see you code first though

plus sequential encodings

What do you mean by that? I put "time since now' as a feature in order to encode times.