What is the loss after training

[lanyufei]

I trained with my own data set, and loss felt high and converged slowly

[Ma-Dan]

如果是没有加载预训练权重的从头训练，请解除前面层的冻结。

[lanyufei]

我加载预训练权重了，最终earlystopping的loss30多，我在yolov3上训练loss训练后17多，这正常吗

[Augenstern-yzh]

我训练也遇到这种问题，loss很高，还没有我用yolov3得到的loss值低，而且loss值降低的很慢，请问这什么原因呢？

[robisen1]

I trained with my own data set, and loss felt high and converged slowly @lanyufei did you have to use a very small batch size when training on your own dataset? I had to use 4 even though I am using a RTX 2080 TI founders edition GPU with 11g's of ram. Doing the same on qqwweee's implementation, which this borrows from, I could do 16 when training from scratch. Any idea whats using up all the memory?

[iliask97]

I trained with my own data set, and loss felt high and converged slowly @lanyufei did you have to use a very small batch size when training on your own dataset? I had to use 4 even though I am using a RTX 2080 TI founders edition GPU with 11g's of ram. Doing the same on qqwweee's implementation, which this borrows from, I could do 16 when training from scratch. Any idea whats using up all the memory?

qqwweee's implemantation is on yolov3 not v4. yolov4 is a way bigger network and requiers a lot more vram to run. In the github repository of the original yolo their is a section witch tells you what changes you need to do according to your gpu memory to train it faster