alven

Results 7 comments of alven

just remove parameter "out_grad" in SoftmaxOutput and LinearRegressionOutput

in my experience, pytorch is faster than tensorflow, about 200 pics/s(200 - 300 i think) in my k80. so one epoch which is 25w pics will take 250000 / 200...

dpn part missed something, could you pull and try again? the accuracy depend on your dataset

对数据进行整理可以增加精度; nsfw的数据比较敏感我个人提供不了不好意思~

现在准确度多少了啊?可以增加 data augmentation,dropout, 减少batch_size试下 其实主要还是脏数据太多了

等下,你这是准确率啊,那不是过拟合啊 继续训练呗,应该还没收敛吧

都可以的,只不过熟悉了pillow或者opencv就习惯用他们写了,而且改成tensorflow也容易 Gaussian,color,Blur等等,感觉都挺有用的啊,不仅相当于增加了数据集,还相当于regularization了,也就减少了过拟合。