About training and testing

[zkk-web]

Hello! When the code is running, training and testing display 6 and 11 numbers respectively, such as training acc: [0.18333333 0.37 0.51333333 0.54666667 0.55333333 0.55 ] Test acc: [0.2974 0.531 0.6133 0.658 0.6636 0.664 0.67 0.6694 0.6714 0.672 0.6714] What do these numbers mean? thanks!

[juliusyang97]

看最后一个数，是最后累计的结果，前面是每步更新过程中的值

[YangJae96]

@zkk-web Hi, did you figure out what it means?

[juliusyang97]

@zkk-web Hi, did you figure out what it means?

Hello, you can check the meta.py. Training acc has 6 values because task level inner update steps are 5. The acc will be calculated once for each update. The same is true for test acc。

[YangJae96]

@zkk-web Hi, did you figure out what it means?

Hello, you can check the meta.py. Training acc has 6 values because task level inner update steps are 5. The acc will be calculated once for each update. The same is true for test acc。

Thank for the explanation. In the original paper, the 5-way 1shot accuracy for Mini-ImageNet is 48.70.

But when I ran the code, the accuracy is lower than original paper.

Which value in Test acc: [0.1996 0.4207 0.4316 0.437 0.4373 0.4377 0.438 0.4382 0.438 0.4387 0.4385] should I look at to compare with the original paper?

[juliusyang97]

@zkk-web Hi, did you figure out what it means?

Hello, you can check the meta.py. Training acc has 6 values because task level inner update steps are 5. The acc will be calculated once for each update. The same is true for test acc。

Thank for the explanation. In the original paper, the 5-way 1shot accuracy for Mini-ImageNet is 48.70.

But when I ran the code, the accuracy is lower than original paper.

Which value in Test acc: [0.1996 0.4207 0.4316 0.437 0.4373 0.4377 0.438 0.4382 0.438 0.4387 0.4385] should I look at to compare with the original paper?

The last number. That is 43.85

[juliusyang97]

@zkk-web Hi, did you figure out what it means?

Hello, you can check the meta.py. Training acc has 6 values because task level inner update steps are 5. The acc will be calculated once for each update. The same is true for test acc。

Thank for the explanation. In the original paper, the 5-way 1shot accuracy for Mini-ImageNet is 48.70. But when I ran the code, the accuracy is lower than original paper. Which value in Test acc: [0.1996 0.4207 0.4316 0.437 0.4373 0.4377 0.438 0.4382 0.438 0.4387 0.4385] should I look at to compare with the original paper?

The last number. That is 43.85

Maml training is not stable and very slow. Keep on training and try to lengthen the training time