[Deep learning] Lora model training summary

1. Lora training does not mean that the more training the better, there may be overfitting. The following are 20 pictures of insulator training

After testing, it is found that the first epoch of training is the most realistic and the situation is in line with my expectations, but the generation in the subsequent epochs is getting worse and worse, resulting in the generated images not meeting the actual requirements. In addition, after testing, I found that the URL generation model of vega ai is better than the model I trained myself, and he only trained for 15 minutes. I don't know why. Maybe the website training parameters are better than the default. But it is not that the less training, the better. The test found that if the number of iterations is too small, the shape will be out of shape.