search

leeyoshinari / YOLO_v2

The implementation of YOLO v2 with TensorFlow
GNU General Public License v3.0
80 stars 38 forks source link

train problem #6

Closed Nine9Nine closed 6 years ago

Nine9Nine commented 6 years ago

I use your program to train my dataset（8 classes，2500pictures）,the loss is too huge,should I change any parameters？

Nine9Nine commented 6 years ago

image

leeyoshinari commented 6 years ago

Are you train from scratch, or using pre-train model? Maybe you can use smaller learning rate. It maybe related to your dataset distribution.

Nine9Nine commented 6 years ago

我是从零开始训练的，把weights路径注释掉了，学习率调了还是爆炸sad

leeyoshinari commented 6 years ago

Because of the network is deep and difficult to train, my suggestion is using a pre-training model. If you want to start training from scratch, you can freeze some layers during training.

Nine9Nine commented 6 years ago

多谢大佬

Nine9Nine commented 6 years ago

hello my lord, I convert the dataset to C version Yolo's format,the training process is smoothly.I think maybe something is wrong in your train program...

ruikun commented 6 years ago

@Nine9Nine hello bro! I got the same problem, the loss is exploding to nan. How did you handle this problem? thanks a lot!

Nine9Nine commented 6 years ago

@ruikun 没解决，先用官网C版本对付了一下，之后还会继续研究

ruikun commented 6 years ago

@Nine9Nine 你好，感谢回复，能不能麻烦你看一看我新开的帖里提出的问题? https://github.com/leeyoshinari/YOLO_v2/issues/7 帮忙给个思路，现在也不知道是train的方法有问题还是数据集的毛病。多谢!