accuracy about miniImagenet protonet ResNet18 train_aug

[ChaiXingliang]

Hello, I have run your code python ./train.py --dataset miniImagenet --method protonet --model ResNet18 --train_aug python ./save_features --dataset miniImagenet --method protonet --model ResNet18 --train_aug python ./test.py --dataset miniImagenet --method protonet --model ResNet18 --train_aug but I got the max test accuracy is 72.30%, while the reported is 73.68. the version of my pytorch is 1.0.0. could you give me some explains and advice? Thank you!

[wyharveychen]

Hello, in terms of the 1% error, it may happen if the "best model" selected by validation is not the best model. If you want to ensure it is the best model, you can have a more strict validation. For example, change the range(100) in line 112 meta_template.py to the range(600), so your validation would be based on 600 times test now.

[ChaiXingliang]

Oh, thanks for your answer, I have tried it but the test accuracy still has no change, I run it with python 2.7.15 and pytorch0.3, and it has no improvement.

[wyharveychen]

Hello, also notes that even with the same model, the accuracy reported by the test.py each time would be slightly different. Can you check it?

[ashok-arjun]

If the seed parameter is set, it wouldn't be different. It would benefit for the train code also, since the sampling of episodes is random at every run