High training performance but low evaluation performance

[evan69]

Your work is excellent and thanks for releasing the code. However when excuting your code I face with a problem. I trained 1 mode 5w1s MAML on cifar by:
python main.py --dataset cifar --maml-model True --num-batches 600000 --output-folder maml_5w1s_cifar
The training performance is fine with "accuracy meta_val after" > 80%, but I cannot obtain similar result in --eval mode:
python main.py --dataset cifar --maml-model True --num-batches 60000 --checkpoint ./train_dir/maml_5w1s_cifar/maml_conv_60000.pt --eval
The "accuracy meta_val after" is about just 30%, and confidence interval gives similar result on "post_val_accuracy":

Evaluation results:
num_batches: 6000
post_train_accuracy: 0.9916700004965068 +- 0.0003492838883183404
post_train_loss: 0.4392507508163651 +- 0.0006438929325898017
post_val_accuracy: 0.28511111702825875 +- 0.0006488388304810349
post_val_loss: 1.6567623161733154 +- 0.0012289125473787653
pre_train_accuracy: 0.20061333698555828 +- 0.0013524214528860354
pre_train_loss: 1.6136694465855759 +- 0.0002620773080873398
pre_val_accuracy: 0.20086889288052914 +- 0.0005964522753135613
pre_val_loss: 1.6139000959157945 +- 0.00012502413922341256

Is 28.51%+- 0.06% the accuracy of MAML on cifar? I am new to meta learning and I doubt whether I use your code properly. I would really appreciate it if you could help me evaluate the model in correct way. Thanks!

---

In addition, problem also occurs in 2w1s training on cifar, but does not occur on omniglot dataset.