search

yue-zhongqi / ifsl

[NeurIPS 2020] Released code for Interventional Few-Shot Learning
164 stars 23 forks source link

some question about MAML #16

Open ljx970128 opened 1 year ago

ljx970128 commented 1 year ago

1.--dataset miniImagenet --method metatrain --train_aug --test maml5_resnet use this command, the accuracy is 92%

--dataset miniImagenet --model na --method metatrain --train_aug --test maml5_ifsl_resnet use this command, the accuracy is 89.33%

This is not consistent with the results in the paper. In particular, ifsl does not improve the accuracy of maml

2.--method metatrain --train_aug --test maml5_ifsl_resnet_tiered use this command, the following error occurs: size mismatch for fc.weight: copying a param with shape torch.Size([64, 512]) from checkpoint, the shape in current model is torch.Size([351, 512]). size mismatch for fc.bias: copying a param with shape torch.Size([64]) from checkpoint, the shape in current model is torch.Size([351]).

The parameters of the pre-training model are the same as those of miniImagenet.

3.In the MateTrain.py, I find some commands about new_test.py, but I couldn't find such a document.

ljx970128 commented 1 year ago

image The accuracy reaches 100 in the first epoch. Then the subsequent training is not meaningful image