请教一下训练细节

[qizhou000]

您好！请问一下，对于每次的迭代，是从DIV2K中随机抽取16张图像，然后随机裁剪出48x48的patch，最后执行8个方向的变换，获得训练数据吗？DIV一共有800张图像，那么一个epoch应该是800/16=50次迭代，您论文中是每200000次迭代学习率下降一半，也就是说是每4000个epoch才下降一次学习率吗？但是您的代码中默认是每200次迭代下降一次学习率，并且整个训练也只有1000次迭代。而按照1000个epoch训练下来，set5的PSNR只能达到32.2。可能我在训练细节上的理解与您有出入，还请不吝赐教。 由于pytorch版本升级，我重写了您的数据集代码，因此细节未知。

[vvictoryuki]

you can check the code in this so in training process, the training data will be repeated 20 times. 800 images will be 800*20=16000 in one epoch, in other words, 16000/16=1000 iters. training of 200 epochs is exactly equal to 200000 iters

[qizhou000]

you can check the code in this so in training process, the training data will be repeated 20 times. 800 images will be 800*20=16000 in one epoch, in other words, 16000/16=1000 iters. training of 200 epochs is exactly equal to 200000 iters

Thank you very much~

[kongdebug]

you can check the code in this so in training process, the training data will be repeated 20 times. 800 images will be 800*20=16000 in one epoch, in other words, 16000/16=1000 iters. training of 200 epochs is exactly equal to 200000 iters

Thank you very much~

您好，请问您从头开始训练时，有遇到模型不收敛的问题吗？我的loss一直在1e29附近振荡

[Liiiiaictx]

@kongdebug 你大概训练了多少轮？