When I train on my own Dataset, the following problem appears, how can I solve it？

[mia-zhu111]

----- TRAINING - EPOCH 40 -----
LEARNING RATE:  0.0033786452453014168
    =>The average detection loss of epoch 40 is 0.6146
    =>The average recognition loss of epoch %d is None, didn't train
----- VALIDATING - EPOCH 40 -----
The validation mode is: detection only.
=> Precision:  0.0
=> Recall:  0.0
=> mGauss: 0.000

The reason for this is that I can't directly use my own Dataset for training? What parameters need to be adjusted? Thank you for your reply！

[JackEasson]

Maybe you can supply more details like training loss. If the loss decreases normally and naturally? And you can use test_demo.py to evaluate if your model has been trained successfully. And please ensure your dataset (training and validation) are set up properly, especially labels. PS: For convenience, you can use Chinese to reply.

[JackEasson]

You can also decrease the threshold of detection. I set it as 0.7, maybe it's too large. You can adjust as described in this issue: https://github.com/JackEasson/SLPNet_pytorch/issues/10

[mia-zhu111]

You can also decrease the threshold of detection. I set it as 0.7, maybe it's too large. You can adjust as described in this issue: #10

loss可以降下来，数据集是按照角点 问题分析：检测部分应该没有问题，应该是车牌识别的训练部分出问题 1、train epoch: The average recognition loss of epoch %d is None, didn't train 2、validation epoch: The validation mode is:detection only 这种情况是因为某个参数没有调整吗？

[JackEasson]

detection only表示训练的仅仅是车牌检测部分。train.py中有两个参数reg_epochs和acc_threshold用来控制何时进行检测+识别的训练。你可以按以下三种种方式操作：

1. 改小acc_threshold，比如0.5，这表示检测精度达到多少时才会加入识别分支 example: python train.py --savedir SLPNetweight --acc_threshold 0.5
2. 直接赋值reg_epochs为一个整数（比如5），那么会从此epoch加入识别分支的训练 example: python train.py --savedir SLPNetweight --reg_epochs 5
3. 假如模型的检测部分已经训练好了，可以直接赋值reg_epochs为0，第一个epoch就会自动加入识别分支的训练，重点要记得要用之前训练过一定epoch，具有一定车牌检测能力的SLPNet预训练模型（你的情况比较符合此项，你手头的模型应该检测部分已经训练的比较好了） example: python train.py --savedir SLPNetweight --reg_epochs 0 --pretrained xxx.pth