chrysts / dsn_fewshot Goto Github PK

Hi, thanks for your work, it's very inspiring to me.
i used the dataset and pretrained model provided in attached files, but i can't reproduce the results by running the test_subspace_discirminative.py file. Also in the line 95 and 96, the trlog doesn't have the attribute 'val_loss' and 'val_acc'

No where has the CIFAR_FS_train.pickle, only "train", "test", and "meta"

Hey，
The work of your team is very impressive, and i want to do some contrast experiments with your approach, could you please give out the link to the pre-trained conv4 model of miniImageNet?

Hope to receive your reply soon, thank you.

1 As the tilte , I run the program for the tieredImagenet ,but I met the problem at the data_train['labels'] , then i debug and find that the data_train don't have labels attribute, but it have label_general and label_specific, do you meet the problem or you have any ideas about this? by the way ,the data you provide is about 12GB, I found another 23GB data and it works. Why the data you give have the label_general and label_specific problem?

2 thank you for your work, the mini-imagenet and the tieredImagenet you give the learning rate, could you give the learning rate for cifar fs.
@chrysts look forward you reply sincerly!

Thanks for sharing your excellent works！
I have some questions when I reproduce the results of ProtoNet using ResNet-12：

1. In mini-Imagenet experiments of Resnet12, why you load the val dataset in train.py (line 84) use “phase=test", which I think it should be “phase=val", or the val dataset will be same with the test dataset in test.py (line 91).
2. In your code, it seems that the test dataset and val dataset loaded from the pickle file include some base categories (training data). Dose it cause the data leakage problem？

Thank you very much！

Thank you for your reply for the two issues.
I have another question, I used the 10-shot 5-way to train, and tested the 5-shot, but only get 78.4,could you tell me the paramerts, like the learning rates, my learning rates is lambda_epoch = lambda e: 1.0 if e < 20 else (0.012 if e < 45 else 0.0052 if e < 59 else (0.00054 if e < 68 else (0.00012))),
Are there any other paramerts need to ajust? or how to train and test to get the 80.00 accuracy for the tired-imagnet.
Thank you very much.

Hello, the updated file link is still incorrect and cannot be downloaded after opening it. Can you publish it again?

Fisrtly, thanks for your research, which inspire me a lot.
While I'm tryin to reproduce the result of miniImageNet on ResNet12, the links seems fail and cannot downloads it.
Is there any solution for this, thanks.

Hi!
Thanks for sharing the code.
I noticed that you converted the tensor to double before fedding to torch.svd
May I ask why you have this implementation?
Thanks!

Hi!
Thanks for sharing the code.
When running the conv4 experiment using the python train_subspace_discriminative.py --data-path /home/dsn_fewshot/miniImageNet command, I encountered an error as follows：

acc = count_acc(logits, label)
File "/home/dsn_fewshot/Conv4/utils.py", line 86, in count_acc
return (pred == label).type(torch.cuda.FloatTensor).mean().item()
RuntimeError: CUDA error: an illegal memory access was encountered

In debug mode, I locate this error appearing in line 43 in the subspace_projection.py file，that is,
projected_query_j = torch.bmm(h_plane_j, torch.bmm(torch.transpose(h_plane_j, 1, 2), target_features_expanded) ,
and the error is Unable to get repr for＜class‘torch.Tensor‘＞

Could you tell me how to solve this problem?

Hi,

After downloading the datasets as suggested in the README file, I cannot run the ResNet-12 experiments (Conv4 experiments do run however).

Here are the commands and the output I get:

python3 train.py --train-shot 15 --head Subspace --network ResNet --dataset miniImageNet --eps 0.1

Loading mini ImageNet dataset - phase trainval
Traceback (most recent call last):
File "/home/src/dsn_fewshot/Resnet12/data/mini_imagenet.py", line 45, in load_data
with open(file, 'rb') as fo:
FileNotFoundError: [Errno 2] No such file or directory: '/home/mini-imagenet/miniImageNet_category_split_train_phase_train.pickle'

python3 train.py --train-shot 15 --head Subspace --network ResNet --dataset tieredImageNet --eps 0.1
Loading tiered ImageNet dataset - phase train
Traceback (most recent call last):
File "train.py", line 135, in
(dataset_train, dataset_val, data_loader) = get_dataset(opt)
File "train.py", line 83, in get_dataset
dataset_train = tieredImageNet(phase='train')
File "/home/src/dsn_fewshot/Resnet12/data/tiered_imagenet.py", line 100, in init
self.labels = data_train['labels']
KeyError: 'labels'

Which copies of mini-imagenet and tiered-imagenet have you used for your ResNet-12 experiments? The Mini-Imagenet dataset that you have pointed in the README does not contain any pickle files.

Hi, thanks for your work.

I notice that there is a discriminative loss in section 4.4 of your paper, but i didn't find it in your code when using resnet12 as the backbone.

Should I add this loss if I want to reproduce the results of resnet12 in your paper?

Hi, thanks for your work, it's very inspiring to me.
But I‘m confused about how Equation2 is derived. Could you please give me some promts about how this formula is derived?
And I notice that there is an identity matrix I in this formula, but it seems that the code implementation does not have it. I also don't quite understand how it's implemented here.
Thanks.

could you tell me the config for the protonet for the tiredimagenet and cifar fs ,I can't the the 5 way 1 shot accuracy as the papers,use the protonet (backbone is resnet 12)?

Hello, thanks for your great work first.
When I run the test_subspace_discriminative in Conv4 on the miniimagenet dataset. The result is that:

{'data_path': '/media/ubuntu/460CFEF50CFEDEBD/wl/CovaMNet-master/dataset/mini-imagenet/For_FewShot/',
 'gpu': '0',
 'lamb': 5,
 'max_epoch': 600,
 'query': 15,
 'save_epoch': 100,
 'save_path': './save/subspace-5w5sdiscriminative/max-acc.pth',
 'shot': 1,
 'subspace_dim': 0,
 'test_way': 5,
 'train_way': 5}
using gpu: 0
 TEST loss=1.2516 acc=0.4917 maxacc=0.0000

I want to know if this result is correct, or what other parameters I need to add during the run.
Thanks in advance for your help.

Hi! Thanks for the great work.

Does the parameter '--train-shot' in ResNet12/train.py mean K-shot during meta-training? And when I set this parameter as 1, the network seems to do not converge and the ACC always is 20%, however, the network can be trained successfully when it is 15.

I am trying to reproduce the results in the paper, and I tried to run the scripts in the Conv4 dir with default parameters, but the results are not as good as the reported ones in the paper. Could you share the parameters and more training details? That can be very helpful for me and I really appreciate it.

Hi,
Using your code, i can reproduce the results on your paper.
Then i try to use DSN to train on the CUB-200-2011 dataset, but I got the error "The algorithm failed to converge because the input matrix is ill-conditioned or has too many repeated singular values".
Are there any solutions to this issue? BTW, I used the common data format of CUB, which contains the raw images with different resolutions.

Thank you in advance.

Hi. In Conv4/train_subspace_discriminative.py, Subspace_Projection is initialized before the args.subspace_dim is assigned, which in the case of 1-shot results in the subspace dimension being set to 0. This seems to be inconsistent with the paper.

Looking forward to your reply！

Sorry to bother you, is your email still [email protected]? I can get in touch with you via email, but it keeps failing, I don’t know why。