search

Seanny123 / da-rnn

Dual-Stage Attention-Based Recurrent Neural Net for Time Series Prediction
333 stars 118 forks source link

Is there any room for gpu memory improvements? #17

Open realiti4 opened 4 years ago

realiti4 commented 4 years ago

Hi, it looks like the loop inside the decoder consumes too much gpu memory very quickly if you try to increase history size because we keep tracking hidden and cell units. I wonder if there is room for any improvements. Would it kill the purpose of the model if we detach some things somehow?

Seanny123 commented 4 years ago

I have no idea. I have no experience optimizing GPU execution. Are you using the jit branch?

On Sat, 15 Feb 2020 at 06:02, Onur notifications@github.com wrote:

Hi, it looks like the loop inside the decoder consumes too much gpu memory very quickly if you try to increase history size because we keep tracking hidden and cell units. I wonder if there is room for any improvements. Would it kill the purpose of the model if we detach some things somehow?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/Seanny123/da-rnn/issues/17?email_source=notifications&email_token=AAOUO2N7TYNQLXRFOLUSDXLRC7DUJA5CNFSM4KVXTJM2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4INYMT3A, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAOUO2N34DKS7JAAYH6CKGTRC7DUJANCNFSM4KVXTJMQ .

realiti4 commented 4 years ago

Hi, I was using master.