The accuracy problem

[MARD1NO]

I wonder use your pytorch script to train RepVGGA0 can achieve which accuracy?

I try to reproduce RepVGGA0 with 0.1 labelsmooth, but get accuracy as 71.6%

[DingXiaoH]

Thanks for the feedback.

1. Did you use our training script (https://github.com/DingXiaoH/RepVGG/train.py) or your own code?
2. What is the final accuracy of RepVGG-A0 without labelsmooth? We did not use it for the 120-epoch training.

[MARD1NO]

Thanks for the feedback.

1. Did you use our training script (https://github.com/DingXiaoH/RepVGG/train.py) or your own code?

2. What is the final accuracy of RepVGG-A0 without labelsmooth? We did not use it for the 120-epoch training.

I just wonder whether other people can reproduce the accuracy with your training script. I use my own code(built in OneFlow framework) with the settings: 5 epochs warmup with cosine decay, 0.1 labelsmooth, 120 epochs.

I found the kernel initializer of conv between Pytorch and OneFlow is different, now i align the initializer and remove warmup, labelsmooth. I will reply the accuracy when the training is finished.

[MARD1NO]

Hi,I retrain my model without labelsmooth got accuracy 70.7%

[DingXiaoH]

I tested our train.py with RepVGG-A0 and the accuracy was 72.44. Maybe there are some other differences in your OneFlow implementation like data preprocessing (we used no color distortion for small models).

[MARD1NO]

I tested our train.py with RepVGG-A0 and the accuracy was 72.44. Maybe there are some other differences in your OneFlow implementation like data preprocessing (we used no color distortion for small models).

A good news is that I use the same data augumentation, add labelsmooth=0.1, warmup 5 epochs and set train epoch as 160(default is 120). I test on the validation set and get 72.6% accuracy!

[hopef]

So, to reproduce the accuracy of the paper, should I use labelsmooth=0.1? thanks!!!