search

zhangdoudou123 / SemFew

[CVPR2024] Simple Semantic-Aided Few-Shot Learning
18 stars 5 forks source link

Hello, first of all, thank you very much for your excellent work. I am very interested in your work, so I tried to reproduce your work, but I encountered some problems after replicating the result. If you can see it, I hope you can help me solve my doubts. #7

Open myl666666 opened 1 month ago

myl666666 commented 1 month ago

SemFew.txt

DruryXu commented 1 month ago

Please refer to my replies in #5

myl666666 commented 1 month ago

Thank you very much for your reply. I want to conduct research on the experiment on resnet network. May I ask how to obtain the checkpoint files of the training four data sets on resnet network? I saw that the checkpoint files of imagenet and cifar data sets you uploaded on github seem to be the only ones. Does the FC100 also use cifar100_semantic_clip_gpt.pth file? Finally, I wish you good health and success in your work

DruryXu commented 1 month ago

What is the exact meaning of "checkpoint" in your question? If you are talking about semantic features, then there are only two semantic features files, because MiniImageNet and TieredImageNet both come from ImageNet, while CIFAR-FS and FC100 both come from CIFAR-100. Therefore, it is true that FC100 also use cifar100_semantic_clip_gpt.pth file.

DruryXu commented 1 month ago

Sorry, I made a mistake with the reference in my reply yesterday. Please see my edited reply.

myl666666 commented 1 month ago

Thank you very much for taking time out of your busy schedule to solve my doubts. The accuracy achieved by my FC100 dataset on the Resnet network is as follows: 1-shot: 7-28 23:24:47, 026-meta_test. py[line:49] -INFO: {'test_batch': 600, 'shot': 1, 'query': 15, 'test_way': 5, 'center': 'mean', 'seed': 13, 'feat_size': 640, 'semantic_size': 512, 'num_workers': 8, 'mode': 'clip', 'text_type': 'gpt', 'dataset': 'FC100', 'backbone': 'resnet', 'model_path': 'E:/miaodashuai/SemFew-main/SemFew-main/ResNet-FC100.pth', 'work_dir': 'resnet_FC100_clip_gpt_mean_1'} 2024-07-28 23:24:48, 495-meta_test. py[line:103] -INFO: best epoch: 45 65.284446 2024-07-28 23:24:48, 496-meta_test. py[line:104] -INFO: best k: 0.400000 2024-07-28 23:25:20,453 - meta_test.py[line:141] -info: max |k: 0.4 |mix acc: 48.33+0.75% |gap: 4.48 2024-07-28 23:25:20,453 - meta_test.py[line:143] -info: ACC:|proto acc: 43.85+0.74% |gen acc: 42.66+0.73% 5-shot: 7-29 00:48:20, 416-meta_test. py[line:49] -INFO: {'test_batch': 600, 'shot': 5, 'query': 15, 'test_way': 5, 'center': 'mean', 'seed': 13, 'feat_size': 640, 'semantic_size': 512, 'num_workers': 8, 'mode': 'clip', 'text_type': 'gpt', 'dataset': 'FC100', 'backbone': 'resnet', 'model_path': 'E:/miaodashuai/SemFew-main/SemFew-main/ResNet-FC100.pth', 'work_dir': 'resnet_FC100_clip_gpt_mean_5'} 2024-07-29 00:48:21,636 - meta_test.py[line:103] -info: best epoch: 32 82.213335 2024-07-29 00:48:21,636 -meta_test. py[line:104] -INFO: best k: 0.790000 2024-07-29 00:48:56, 614-meta_test. py[line:141] -INFO: max |k: 0.79 |mix acc: 60.42+0.78% |gap: 0.66 2024-07-29 00:48:56,614 -meta_test. py[line:143] -INFO: ACC:|proto acc: 59.77+0.78% |gen acc: 43.02+0.73% Is the above result normal? I conducted the test according to your experimental Settings, the pth file should be corresponding and the path is correct. If there are other neglected problems, please point out. Thanks again for your work and guidance!

KZF-kzf commented 3 weeks ago

@myl666666 can you reproduce their results on FC100 or CIFAR_FS now?

myl666666 commented 2 weeks ago

您现在可以在 FC100 或 CIFAR_FS 上重现他们的结果吗？ i can not ,sorry

DruryXu commented 2 weeks ago

@myl666666 can you reproduce their results on FC100 or CIFAR_FS now?

Could you please share your reproduction results for helping us locating where the problem comes from?