Sorry to bother again. I wonder if you can provide the argument setting for reproducing outcomes of the other datasets like CIFAR-FS or Tiered-ImageNet. Hope this won't bring you much trouble. Thx a lot.

Hello, thank you very much for your work, I have some questions.

1. I noticed that you added a argument "batch_fs_size" to args.py and its default value is set to 20, does this argument affect the test performance? From my understanding, it seems to affect only the GPU memory usage and speed during testing.
2. How to reproduce the results on CUB and tiered-ImageNet?
3. According to the command you gave for mini-ImageNet, the computation and back-propagation of a batch's loss is performed twice, where the first time is the classification loss with manifold mixup and the second time is the original classification loss and the rotation loss, i.e., a batch contains two gradient updates. Why does it need to be done in this way? Does this mean that the training of 500 epochs is approximately equivalent to the training of 1000 epochs for the methods with one-time gradient update for a batch? I personally tried to perform the manifold mixup classification and the roation loss together, (according to the code: if mixup and args.mm: ......) https://github.com/ybendou/easy/blob/85937e0d2d67a801dba7a96974a79c2d6cad86b7/main.py#L84-L96 but this resulted in a slight performance degradation, is this because there is a potential conflict between manifold mixup and rotation?
4. Have you studied the compatibility of the different augmentation methods with the self-supervised rotation loss? For example, does random horizontal flipping break the rotation loss? I'd like to hear your opinion.

Sorry for my poor English, I hope I expressed my question clearly. Best,

Excuse me, I have followed the command of training a ResNet12 with the followings: python main.py --dataset-path ' ' --dataset miniimagenet --model resnet12 --epochs 0 --manifold-mixup 500 --rotations --cosine --gamma 0.9 --milestones 100 --skip-epochs 450 --batch-size 128 --preprocessing ME --save-model "result/official/mini1.pt" --n-shot [1,5] which is the same as the one you provided. But I can only reach the accuracy of 64 of each backbone, and after I ensembled the features the accuracy can only reach 65. The features I downloaded from OneDrive can have an accuracy of around 70 for each, which is far beyond mine. Hope you can solve my problem, thx!

Hello,

In this issue https://github.com/ybendou/easy/issues/11#issuecomment-1116237889 you mentioned changing the argument dataset to ‘tieredimagenet’ can reproduce the results on tieredImageNet.
My device is an RTX 3090 with 24GB memory. Even if I reduce the batch size from 128 to 64, there is still the problem of out of GPU memory during training. I noticed that the memory usage is only about 6GB during training, but once the training of a epoch is finished and the validation starts, the GPU memory is immediately exhausted. What should I do to complete the training on tieredImageNet?

Best,

Hello,

May I know how to run test for tieredimagenet using ASY? I used the following command line, but it seems not working for multiple GPU. python main.py --dataset-path 'my data path' --dataset tieredimagenet --model resnet12 --test-features tieredfeaturesAS1.pt11 --preprocessing ME --n-shots 1 --device cuda:0123 If I just used --device cuda:0, I got out of memory error as the feature is too big for tieredimagenet. I am using 4 2080ti GPU with 12GB memory per device.

Are we also saving the few-shot head here?

I am concerned with deploying the model in production. Looks like we can't just do a normal prediction on the loaded model but need to take care of the logic of few_shot_eval.eval_few_shot?

First of all, thx for your work! Since I want to reproduce the outcome of 2x_ResNet12_√2 which has an accuracy of 70 in 1-shot case on mini-ImageNet, I want to know what is the training setting of the main.py? I want to extract the features from my own-trained model, hope you can help me.

Hello, I have a question about the feature.

Why do the inductive and transductive use the same feature?

I thought they were using different ways to train cause the testing set had only been used in transductive learning, not inductive learning.

Thanks.

Hi,

I try to test features on cifarfs by this command: """ $ python main.py --dataset-path "" --dataset cifarfs --model resnet12 --test-features "[/AS1.pt11, /AS2.pt11, /AS3.pt11]" --preprocessing ME --n-shots 1 """ and it output "RuntimeError: CUDA out of memory".

I am using 3080 GPU with 8GB memory. I also tried colab and got the same problem. May I know how to fix it?

Thanks

Hi, your work is wonderful! Thank you for the detailed and neat code! I have some questions about the code.

About the network:

1. I'm a bit confused about the network structure of ResNet12, it seems to have 3 convolutional layers in each BasicBlock? But the official ResNet18 provided by Pytorch has only two convolutional layers in each BasicBlock.
2. Why is the MaxPooling operation in ResNet12 performed outside the BasicBlock and not inside the BasicBlock as in ResNet18?
3. ResNet18 has a total of 17 convolutional layers and one FC layer, so it contains a total of 18 layers with trainable weights. But ResNet12 in this code has 12 convolutional layers and one FC layer, that is, it has 12 layers with trainable weights, should it be called ResNet13?
4. The number of parameters in ResNet12 (12.4M) is even more than that of ResNet18 (11.2M), which makes me confused. Why do we design a network that seems lighter from the name but is actually heavier? I replaced ResNet12 with ResNet18 for training and found that the performance dropped.

About training and testing:

1. When centering the feature vectors, each vector should minus an average feature vector. The paper says that “\overline{z} is the average feature vector of the base dataset if in inductive setting or of the few-shot considered problem if in transductive setting.” But the code seems to show that \overline{z} is still used as the average vector of the base dataset in transductive setting?
2. If I want to reproduce a Y ResNet12(√2) model, should the script be “python main.py --dataset-path "" --dataset miniimagenet --model resnet12 --epochs 0 --manifold-mixup 500 --rotations --cosine --gamma 0.9 --milestones 100 --batch-size 128 --preprocessing ME --feature_maps 45”
3. Why set forceCPU=True for val_loader and novel_loader?
4. What does the .pt55 suffix file represent in the pre-trained weights and features provided? I tried to test the features in minifeatures1.pt55 and found its performance inferior to minifeatures1.pt11. By the way, can Pytorch read and write files in .pt11 or .pt55 format? I know the common formats include .pt, .pkl, and .pth, but what format is .pt11?

Hi,

Great paper ! I have a question regarding the training procedure with multiples datasets. If I have lets say 6 datasets with 5 class each, with the old training procedure of MAML or ProtoNet, where the model is trained with episodes and tasks, this seems simple enough since each data set would be a task. With your training procedure (which seems to be used in all recent papers), how would you do it ? Concatenate all the classes to have 30 classes and train with batches made of samples of all datasets ?

Thanks a lot if you have the time to answer :)