About the total_loss and AP

[jianjian-123]

Hello, I trained the model for the first time according to the number of iterations you set 30,000 times, but the total_loss has been floating around 10 within 6 hours after training, and the results of the model evaluation are also very bad (Figure 1); The second time I commented out the gamma value in the config, I found that the total_loss began to decline, but it fluctuated around 6 within 10 hours of training, and the model evaluation effect was not satisfactory (as shown in Figure 2). Expect your advice.

[baiwhitedog]

excuse me, there is no .json file in datasets like the picture in Readme, what should I do?

[JoyeZLearning]

you can use the offered code to transform VOC to json(coco) style @baiwhitedog

It seems the network study few information from dataset, you can check you code (mainly in the section of Dataset processing) to address this issue @jianjian-123