sungyubkim / gbml Goto Github PK

As mentioned in Section 2.2 of "Meta-Learning with Implicit Gradients" paper, regularization term should be added in inner-loop objective function like figure below. however, your code (line 25 of gbml/imaml.py) is missing this regularization term.

Hi Sungyub Kim,

Thanks for the great implementation.

I have a question about your code here:

Here, you are first training on all the batches in the meta-train set, then you are doing validation and testing.
However, the original algorithm seems to record the testing accuracy after training on every batch of tasks in the meta-train set. Do you observe the difference? I know it is a matter of implementation, we could have done the testing simultaneously, but as a matter of keeping consistent with the original implementation and the way others report their accuracy, would it make sense to observe the testing accuracy immediately after training on 1 meta-batch?

Thanks

Thank you for sharing your codes.

python main.py --alg "iMAML" --net "ResNet"

I have some problem, when I train imaml with ResNet.

ConVNet case is Ok for me, but ResNet case makes some problem.

Did you ever face like this problem?

Hi @sungyubkim, thank you for publishing this great repository.

I found that your MAML algorithm on ResNet increases the GPU memory usage every iteration and causes a CUDA out of memory error.
I ran it by python main.py --alg MAML --net ResNet. I use PyTorch 1.3.1 with CUDA 10.1 and the latest dependencies.

Thank you.

Thank you for your excellent work.

I run python3 main.py --alg=MAML without any changes for 5way1shot tasks on miniImagenet, and the test accuracy seems to be 63.85%. But the official result of MAML is 49%. Is my understanding wrong?

Hello, This seems like an error in my torchmeta setup, but wanted to get your perspective if otherwise. thanks in advance - VV

(dsai20) dg8965@mb-qs-pp-v100:~/GBML$ python3 main.py --alg=Reptile
using gpu: 0
Traceback (most recent call last):
File "main.py", line 225, in
main(args)
File "main.py", line 115, in main
target_transform=Categorical(num_classes=args.num_way)
File "/home/dg8965/.conda/envs/dsai20/lib/python3.6/site-packages/torchmeta/datasets/miniimagenet.py", line 89, in init
download=download)
File "/home/dg8965/.conda/envs/dsai20/lib/python3.6/site-packages/torchmeta/datasets/miniimagenet.py", line 127, in init
raise RuntimeError('MiniImagenet integrity check failed')
RuntimeError: MiniImagenet integrity check failed

When codes to add meta curvature and iMAML are planned. Any updates.

Thanks so much for the great codes!

I was checking the reptile code, and it appears I need to set track_higher_grads=True in the context for this to run.

Is there something I am missing here? Thanks!