About mse

[TING-PAN]

Hello, the paper is very interesting. I tried to run the proposed code, but got MSE 39. After that, I tried to change the Lmoment to 0.1 and 0.01, but got MSE still around 40. What am I missing ?

[TING-PAN]

I set Lmoment as 0.1 and trained 1000 epochs, but MSE was 32, which still failed to achieve the effect described in the paper

[TING-PAN]

After experiments, I found that setting the batch size to 16 and using 1080ti can get good results. Thanks for author's help.

[eugenelet]

Hi @TING-PAN ,

Can you share what modifications you made to the existing code to obtain that result? As from issue #21 that I made, I couldn't even get close to the reported results using the default settings of the given code.

Eugene

[TING-PAN]

Can you share what modifications you made to the existing code to obtain that result? As from issue #21 that I made, I couldn't even get close to the reported results using the default settings of the given code.

Hi : I set lr 0.001 and batch size 16, teacher_forcing_ratio = np.maximum(0, 1 - epoch * 0.003), scheduler_enc = ReduceLROnPlateau(encoder_optimizer, mode = 'min', patience = 3, factor = 0.1, verbose = True)

[eugenelet]

@TING-PAN ,

Thanks for sharing your training hyperparameters. I'll update you if I can obtain the results shown in the paper.

[vincent-leguen]

Hi @TING-PAN @eugenelet , I confirm the modifications explained above by @TING-PAN. In addition, I added the separate encoders and decoders version explained in the CVPR OmniCV workshop paper. I have updated the Github repo and added a pretrained model that attains MSE=24.19. Best, Vincent

[eugenelet]

@vincent-leguen Thanks for the update!

Eugene