Caffe: Data layer prefetch queue empty

资源1

The problem was that the text files for training and test data were not shuffled, but sorted. Thus, test was done only on the images of same class, and parameters were learned from the images of that same class as well, leading to 100% accuracy very quickly.

资源2

So can you get the right result ( accuracy and loss) while finetune , in my result it seems only slow down the test process.

资源3

I think it is the problem of power supply, not the problem of Caffe.

总结：

原因：CPU数据读入慢于GPU计算，导致GPU处于空闲，使得计算速度下降，迭代时间间隔扩大。nvidia-smi显示为：GPU功率不断大幅度变化，GPU电压不稳定。

可能诱因：服务器同时在进行其他工作，如大规模读写数据（频繁使用IO），数据无法读入；CPU同时进行其他工作，负载极重，无法正常读取数据；CPU内存爆表，所有程序速度变慢。

解决办法：

1 使用lmdb替换txt文件，加快数据读取速度；

2 避免IO被频繁使用；

3 可以考虑不加入test.txt