Open 20231211 opened 4 weeks ago
似乎ema没有发挥作用?存储时ModelEMA这个类的state_dict中return的是dict(module=self.module.state_dict(),...),测试时是module=self.ema.module if self.ema else self.model,所以使用的还是self.module,并不是平均的结果?
测试时是module=self.ema.module if self.ema else self.model,所以使用的还是self.module,并不是平均的结果?
self.ema
不是None
的话 测试时候也是用self.ema.module
的
@lyuwenyu 但self.ema.module存的还是当前epoch的self.module?似乎没发挥ema的作用?
What was the design of the multi_scale before entering the backbone? It seems to work better without using multi_scale on my dataset. In addition, I would like to ask how many rounds of general training, the training dataset is obtained by slicing, there are about 16,000 pieces, when the training is 72 rounds best_stat_epoch=71, do you need to continue training? When training for 120 rounds, the AP@0.5:0.95 improved, but the AP@0.5 remained almost the same. There is also a question about the selection of test results, because EMA is used, should the test results be the last round of test results? Which PTH file should I choose as the final model result? How much is it reasonable to have about a difference in APs between epochs in the later stages of training?