Open XinGP opened 5 months ago
@XinGP Generally, if you reduce the batch size, you should also reduce the learning rate (ref: https://arxiv.org/pdf/1812.01187.pdf). The relatively large LR will lead to unstable training.
Hello author! I reproduced your AV1 training code, adjusted the epoch to 50, and adjusted the batch_size to 4, with no other adjustments made. What is the reason for the significant difference in the validation results obtained after training compared to your validation model? Hello, I've also encountered similar issues while tuning parameters, and I can't pinpoint the specific learning rate. Could you share what learning rate you ended up using? I'd like to discuss this issue further. My email is lipl23@mails.jlu.edu.cn.
Hello author! I reproduced your AV1 training code, adjusted the epoch to 50, and adjusted the batch_size to 4, with no other adjustments made. What is the reason for the significant difference in the validation results obtained after training compared to your validation model?
Have you found a suitable learning rate for batch_size 4 to close to the result which the author gives? @XinGP
@XinGP一般来说,如果减小batch size,也应该减小学习率(ref: https://arxiv.org/pdf/1812.01187.pdf)。LR过大会导致训练不稳定。
Yes, after reducing all learning rates by half, I can obtain validation set results similar to those in the paper
Hello author! I reproduced your AV1 training code, adjusted the epoch to 50, and adjusted the batch_size to 4, with no other adjustments made. What is the reason for the significant difference in the validation results obtained after training compared to your validation model?