Mengye Ren

Hi, there shouldn't be problem with compute the gradient after the dummy GD and in fact it is by design. It is backproping through the dummy GD step multiple times...

Hi, thank you for the interest in our work. I found that data augmentation and early stopping is important for training ResNet.

The Acc means top-1 accuracy. Train-A-Val top-1 is around 62.4%.

Hi, You need to make sure that the mini-imagenet pickle files are in the right place. Sorry about the messy code but line 97 shouldn't be reached if the pickle...

You can use the following command: ``` ./run.sh {GPUID} python run_exp.py --config {CONFIG_FILE} \ --dataset {DATASET} \ --data_folder {DATASET_FOLDER} \ --pretrain {PRETRAIN_CKPT_FOLDER} \ --nshot {NUMBER_OF_SHOTS} \ --nclasses_b {NUMBER_OF_FEWSHOT_WAYS} \ --results...

Hi Wihan, Thanks for reporting the bug. I have fixed the doc using tf.sparse_softmax_cross_entropy_with_logits instead. I will work on an update that includes the gradient for softmax_cross_entropy_with_logits. Mengye

Thanks for reporting! I fixed in the latest commit.

Which dataset are you referring to?

Thanks a lot for the post. I like your idea, which is way more elegant, in terms of using the built-in reverse-mode AD in TensorFlow (although from implementation perspective, forward-mode...

I can take a look, but this one of the core module of tensorflow, so better be changed by some of their core developers. In the mean time, you can...