liyuan97 / liyuan97.github.io

My Personal Blog
https://leeee.top/
1 stars 0 forks source link

pytorch再接触 | 元哥的日记 #50

Open liyuan97 opened 4 years ago

liyuan97 commented 4 years ago

https://leeee.top/2020/pytorch2/

skorch是基于pytorch的外部工具,集成了很多基础功能,使得训练过程变得异常简单。一个net.fit()就能搞定。本次是MSBD5002的第二次作业,作业要求用MLP实现二分类和多分类的任务。多分类其实是对于图片分类,所以我也照着MNIST的CNNdemo改了个CNN的模型,CNN效果会好一点。但是由于数据集本身的原因,准确率没能上90%。

wangcmino commented 4 years ago

踩踩。。。

liyuan97 commented 4 years ago

踩踩。。。

评论再看,养成习惯😇

amychu131 commented 4 years ago

about how the learning rate , just set to 0.1? no need to find the best?

amychu131 commented 4 years ago

also no need to 'You may also de- crease gradually during the learning process to enhance convergence.'???

liyuan97 commented 4 years ago

about how the learning rate , just set to 0.1? no need to find the best?

learning rate actually I always setting smaller than 0.001, it always works well. I think learning rate just is a speed to approach the convergence, it's no need to find the best if you do not care about speed.