ZhaoZhibin
commented
2 years ago 当我在用CWRU数据集上由0到2迁移时，在middle epoch时报错： Traceback (most recent call last): File "E:\UDTL-master\train_advanced.py", line 88, in trainer.train() File "E:\UDTL-master\utils\train_utils_combines.py", line 286, in train distance_loss = self.distance_loss([features.narrow(0, 0, labels.size(0)), File "E:\UDTL-master\loss\JAN.py", line 44, in JAN loss = torch.mean(XX + YY - XY - YX) RuntimeError: The size of tensor a (20) must match the size of tensor b (64) at non-singleton dimension 1 在1迁移到2时则不会报错。 相应参数设置为： 04-09 11:14:17 model_name: resnet_mix_features_1d 04-09 11:14:17 data_name: CWRU 04-09 11:14:17 data_dir: E:\Data\CWRU 04-09 11:14:17 transfer_task: [[0], [2]] 04-09 11:14:17 normlizetype: mean-std 04-09 11:14:17 cuda_device: 0 04-09 11:14:17 checkpoint_dir: ./checkpoint 04-09 11:14:17 pretrained: False 04-09 11:14:17 batch_size: 64 04-09 11:14:17 num_workers: 0 04-09 11:14:17 bottleneck: True 04-09 11:14:17 bottleneck_num: 256 04-09 11:14:17 last_batch: False 04-09 11:14:17 distance_metric: True 04-09 11:14:17 distance_loss: JMMD 04-09 11:14:17 trade_off_distance: Step 04-09 11:14:17 lam_distance: 1 04-09 11:14:17 domain_adversarial: True 04-09 11:14:17 adversarial_loss: CDA 04-09 11:14:17 hidden_size: 1024 04-09 11:14:17 trade_off_adversarial: Step 04-09 11:14:17 lam_adversarial: 1 04-09 11:14:17 opt: adam 04-09 11:14:17 lr: 0.001 04-09 11:14:17 momentum: 0.9 04-09 11:14:17 weight_decay: 1e-05 04-09 11:14:17 lr_scheduler: step 04-09 11:14:17 gamma: 0.1 04-09 11:14:17 steps: 150, 250 04-09 11:14:17 middle_epoch: 5 04-09 11:14:17 max_epoch: 300 04-09 11:14:17 print_step: 50 04-09 11:14:17 using 1 gpus 麻烦赵老师有空帮我解答一下。非常感谢
从问题来看应该是你同时使用了距离和对抗，因为距离的时候我们是丢掉最后一个batch的，不然维度不匹配没法计算距离，对抗的话不用丢弃最后一个batch。所以有可能这里面出现了矛盾，你可以仔细看一下。
yang301301
seamoonlight-YBY
当我在用CWRU数据集上由0到2迁移时，在middle epoch时报错： Traceback (most recent call last): File "E:\UDTL-master\train_advanced.py", line 88, in
trainer.train()
File "E:\UDTL-master\utils\train_utils_combines.py", line 286, in train
distance_loss = self.distance_loss([features.narrow(0, 0, labels.size(0)),
File "E:\UDTL-master\loss\JAN.py", line 44, in JAN
loss = torch.mean(XX + YY - XY - YX)
RuntimeError: The size of tensor a (20) must match the size of tensor b (64) at non-singleton dimension 1
在1迁移到2时则不会报错。
相应参数设置为：
04-09 11:14:17 model_name: resnet_mix_features_1d
04-09 11:14:17 data_name: CWRU
04-09 11:14:17 data_dir: E:\Data\CWRU
04-09 11:14:17 transfer_task: [[0], [2]]
04-09 11:14:17 normlizetype: mean-std
04-09 11:14:17 cuda_device: 0
04-09 11:14:17 checkpoint_dir: ./checkpoint
04-09 11:14:17 pretrained: False
04-09 11:14:17 batch_size: 64
04-09 11:14:17 num_workers: 0
04-09 11:14:17 bottleneck: True
04-09 11:14:17 bottleneck_num: 256
04-09 11:14:17 last_batch: False
04-09 11:14:17 distance_metric: True
04-09 11:14:17 distance_loss: JMMD
04-09 11:14:17 trade_off_distance: Step
04-09 11:14:17 lam_distance: 1
04-09 11:14:17 domain_adversarial: True
04-09 11:14:17 adversarial_loss: CDA
04-09 11:14:17 hidden_size: 1024
04-09 11:14:17 trade_off_adversarial: Step
04-09 11:14:17 lam_adversarial: 1
04-09 11:14:17 opt: adam
04-09 11:14:17 lr: 0.001
04-09 11:14:17 momentum: 0.9
04-09 11:14:17 weight_decay: 1e-05
04-09 11:14:17 lr_scheduler: step
04-09 11:14:17 gamma: 0.1
04-09 11:14:17 steps: 150, 250
04-09 11:14:17 middle_epoch: 5
04-09 11:14:17 max_epoch: 300
04-09 11:14:17 print_step: 50
04-09 11:14:17 using 1 gpus
麻烦赵老师有空帮我解答一下。非常感谢