xiaocmxiao

@shicai 你好。我用你提供的deploy改写的train_val进行训练，loss一直是2.5左右，不下降。我改train_val只改了开头的数据层和在结尾加了accuracy和softmaxwithloss，请问还有哪里需要修改？

@shicai I have trained the same data with googleNet and squeezeNet,the result is correct.So I think the data have no problem.After squeezeNet's Convolution,there are a ReLu and a Pooling Layer...

@shicai eh... I just want to test a single image,but the classify.py returns an error value always(for example, I have cat,dog,fish, it's return always is cat,although I input a fish...

It's work!thank you.爱你么么哒

@liuyiyiyiyi 训练的时候用了scale，测试的时候也必须用。你也可以训练的时候不用scale。

> > 我的样本是从10个字符里截取3到10个字符生成图片，训练没问题 > > 请问，训练图片尺寸是否要保持32x280 不需要

@hello2sys 不需要。比如训练集全是数字，能有什么语义信息？