New Training fails to achieve reported performance

[caelumchz]

Hello, this work is full of inspiration, and we want to follow up this work and give stars. However, there's something wrong with my own train process. That is to say when I start a new training on Surreal with default paramaters and without using pretrained model, I cannot achieve the performance in the paper(5% lower). Is that because my hyperparamaters are incorrect or something else? And I wonder if the rotated_gt_2 is incorrect since it supervises the orient_2 which is the angle between original target pc and augmented tgt in student model, but seems to be designed to calculate the angle between two rotated point clouds?

[2019EPWL]

Hello, this work is full of inspiration, and we want to follow up this work and give stars. However, there's something wrong with my own train process. That is to say when I start a new training on Surreal with default paramaters and without using pretrained model, I cannot achieve the performance in the paper(5% lower). Is that because my hyperparamaters are incorrect or something else? And I wonder if the rotated_gt_2 is incorrect since it supervises the orient_2 which is the angle between original target pc and augmented tgt in student model, but seems to be designed to calculated the angle between two rotated point clouds?

After I retrained 300 epochs, there was also a decline in performance. I noticed that the maximum epoch in this repo was not 300 but 600, which was different from the baseline.

[caelumchz]

Hello, this work is full of inspiration, and we want to follow up this work and give stars. However, there's something wrong with my own train process. That is to say when I start a new training on Surreal with default paramaters and without using pretrained model, I cannot achieve the performance in the paper(5% lower). Is that because my hyperparamaters are incorrect or something else? And I wonder if the rotated_gt_2 is incorrect since it supervises the orient_2 which is the angle between original target pc and augmented tgt in student model, but seems to be designed to calculated the angle between two rotated point clouds?

After I retrained 300 epochs, there was also a decline in performance. I noticed that the maximum epoch in this repo was not 300 but 600, which was different from the baseline.

Can you successfully reproduce the performance through retraining? I still meet a failure after training 600 epochs.