Training setup issues

[xuxuduoduomeiGithub]

Hello, thank you for bringing wonderful work!
I have the following questions and I hope you can provide assistance:
I only have one RTX3090 graphics card with 24GB of memory, so in the training configuration file of RGT：
num_worker_per_gpu: 16
batch_size_per_gpu: 8
The configuration method should be reduced, and I will try to enable training correctly:
num_worker_per_gpu: 6
batch_size_per_gpu: 4
But will this affect the performance of the indicators after training, and should I adjust other training parameters to achieve balance, such as adjusting the learning rate, increasing the number of training rounds, etc.
Looking forward to your reply~

[zhengchen1999]

Yes, it will affect the training. Model training is based on steps (iterations) rather than epochs. Therefore, changes in batches and GPUs require adjustments to the iterations.

If your batch_size_per_gpu=4 and 1 GPU, compared to original setup of batch_size_per_gpu=8 and 4 GPUs, the amount of training data decreases by a factor of 8. You would need to multiply the iterations by 8 and adjust lr schedule accordingly. This might not guarantee exact reproduction of the results, but the difference shouldn't be significant.

PS: Considering your computing resources, you can consider using DAT-light.

[xuxuduoduomeiGithub]

是的，会影响训练。模型训练是基于步骤（迭代）而不是 epoch 的。因此，batch 和 GPU 的变化需要对迭代进行调整。

如果您的 batch_size_per_gpu=4 和 1 GPU，与原始设置的 batch_size_per_gpu=8 和 4 GPU 相比，训练数据量会减少 8 倍。您需要将迭代次数乘以 8，并相应地调整 lr 计划。这可能无法保证结果的准确再现，但差异应该不大。

PS：考虑到你的计算资源，你可以考虑使用DAT-light。

Okay, thank you for your reply~