Where your PRM?

[ZM-J]

In your paper you said PRM combines \alpha_{l-1} and \alpha_l' to generate a better \alpha_l.

The idea is cool. However, in the code, I didn't found something like PRN.

In trainer.py Line 185, you get prediction result alpha_pred_os1, alpha_pred_os4, and alpha_pred_os8 from the model. However, alpha_pred_os4 and alpha_pred_os8 are obtained from modules refine_OS4 and refine_OS8 in the decoder directly, which use ONE input.

So, where your g_l, i.e., guidance in your paper? I don't see anything like g_l except obtaining loss like utils.get_unknown_tensor_from_pred.

Where your PRM? WHERE YOUR PRM? HAIYAA...

[ZM-J]

So the PRM is only applied during inference, i.e., in infer.py Lines 25~30.

During training, nothing about PRM is used.

OK.