question about the training time

[LGD2333]

Could you please let me know how much time it usually takes you to complete a result?
（When I run corrmatch on a single A100 GPU with a batch size set to 8, each epoch takes almost 22 hours.）

[BBBBchan]

Thanks for your attention. On the Pascal VOC dataset, we trained using 2x3090 GPUs. With a training size of 321x321 and the 92 split setting, the training took approximately 30 hours. On the Cityscapes dataset, we trained using 4xA40 GPUs. With a training size of 801x801 and the 1/8 split setting, the training took approximately 48 hours (the theoretically longest setting).

[Wang-zhenyan]

Thanks for your attention. On the Pascal VOC dataset, we trained using 2x3090 GPUs. With a training size of 321x321 and the 92 split setting, the training took approximately 30 hours. On the Cityscapes dataset, we trained using 4xA40 GPUs. With a training size of 801x801 and the 1/8 split setting, the training took approximately 48 hours (the theoretically longest setting).

作者你好，我想问一下你在实验的时候设置的batch是16吗（即两张卡，每张卡batch=8），我在2×4090上运行Pascal Voc数据集的时候，size=513×513，每张卡最大batch只能设置为2，这样还需要近两天的时间才能完成。我在2×A100上运行Pascal Voc数据集的时候，size=513×513，每张卡最大batch也只能设置为4。
Duplicate of #

[BBBBchan]

Thanks for your attention. On the Pascal VOC dataset, we trained using 2x3090 GPUs. With a training size of 321x321 and the 92 split setting, the training took approximately 30 hours. On the Cityscapes dataset, we trained using 4xA40 GPUs. With a training size of 801x801 and the 1/8 split setting, the training took approximately 48 hours (the theoretically longest setting).

作者你好，我想问一下你在实验的时候设置的batch是16吗（即两张卡，每张卡batch=8），我在2×4090上运行Pascal Voc数据集的时候，size=513×513，每张卡最大batch只能设置为2，这样还需要近两天的时间才能完成。我在2×A100上运行Pascal Voc数据集的时候，size=513×513，每张卡最大batch也只能设置为4。 Duplicate of #

您好，在513x513 size的情况下，您的训练时间是正常的。我们在训练时使用了4x3090，单卡batch size为2，训练时间为32h左右。