I am trying to train the encoder-decoder model (EncoderDecoderModel) using the REINFORCE algorithm. For that, I need to decode using the random sampling method (do_sample=True) and use the logits. I can’t find a way to get the logits for the sampled tokens as, in the training mode, the decoding used…

Implementing the REINFORCE algorithm for encoder-decoder model

NeoFelix March 14, 2022, 8:22pm 2

Hi, did you find a way to do this?

Topic		Replies	Views
Decoding strategy in training phase Beginners	2	708	November 25, 2022
Stochastic Sampling with Trainer.evaluate() Logits Intermediate	3	328	May 6, 2024
Custom Decoding Strategy Beginners	0	465	December 6, 2023
Carrying Gradients Through Generate Research	5	2736	January 29, 2023
How to generate without decoding? 🤗Transformers	1	371	December 13, 2023