SivilTaram / Persona-Dialogue-Generation

The code of ACL 2020 paper "You Impress Me: Dialogue Generation via Mutual Persona Perception"
MIT License
309 stars 46 forks source link

Model weights for Transmitter and Receiver #10

Closed parthpatwa closed 4 years ago

parthpatwa commented 4 years ago

Can anyone please upload the model weights of trained Transmitter and Receiver? Thank you.

SivilTaram commented 4 years ago

@parthpatwa Hi, although I do not know why you need weights of Transmitter and Receiver (actually, we have uploaded the final fine-tuned model), I'm uploading it to Google drive. Please wait for minutes.

SivilTaram commented 4 years ago

@parthpatwa Now we only have Receiver for revised mode LINK and Transmitter for revised mode LINK. Hope it help you.

parthpatwa commented 4 years ago

@SivilTaram Thank you. I need it to retrain PSquare on a new data. I will use your existing model weights instead of starting from scratch

These are the steps I am following: 1) initialize transmitter to existing weights and retrain on new data 2) initialize receiver to existing weights and retrain on new data 3) initialize PSquare to existing weights and retrain by using above obtained transmitter and receiver.

Is this process correct?

SivilTaram commented 4 years ago

@parthpatwa Yes, it seems correct for me. And remember to set a small learning rate to avoid catastrophic forgetting (if your own data is relatively small).

parthpatwa commented 4 years ago

@SivilTaram help! the receiver file you uploaded seems to be corrupted. I tried downloading and using 3 times. pfa image

SivilTaram commented 4 years ago

@parthpatwa which version of pytorch do you use? I will recheck if it could be loaded in my environment.

parthpatwa commented 4 years ago

@SivilTaram I tried on 2 systems- torch==1.5.1 torchvision==0.6.1

and torch==1.6.0 torchvision==0.7.0

both gave same error

SivilTaram commented 4 years ago

@parthpatwa You're correct, and the file is corrupted. I have re-uploaded another weight file for receiver here. You could try it.