245123030
commented
4 years ago Ok, just figured out that it is continuing training, using the --resume statement to add the model path, but I initially thought that using the --pretrain statement (feeling a bit reasonable), but the prompt is that you can not find the checkpoint, here is not resume and pretrain The definition is reversed -. -, am I correct? I would also like to ask you to use this code for the pre-trained model on 2017mlt. Do you need to train 600 epoch if you fine-tune it in the 2015 dataset? Also the model I trained in imageNet pre-trained in your code results recall: 0.71, precision: 0.82, hmean: 0.76, hmean does not reach the expected 0.8, TAT, is this a normal result? Looking forward to your reply！
Hello, first of all thank you very much for your contribution. I have used res_50 to complete the training and testing, and Hmean can reach 0.76. Your model on the 2017mlt dataset and the training model need to be fine-tuned on the ICDAR2015 dataset. You said it is used directly. Is it directly downloaded and then run the test file, or you need to continue to import into the train file for training. Because I used the test file directly, the result could not output the correct box. If you need to continue training, how do you need to modify the training code? Hope to reply -. -