loss is nan

[YvoGao]

Thanks for your work, I have a problem with running default params, loss is nan like:
Session 0, epo 1, lrc=0.0002,total loss=nan, loss_CE=nan, loss_ED=nan, loss_SKD=nan, acc=0.0160: 86%|██████████████████████████████████████████████▍ | 202/235 [04:08<00:40, 1.22s/it

[rongtongxueya]

same question.can give me some advice?

[zichengpan]

Same question and the accuracy is quite low.

[pgh2874]

As other FSCIL methods do, we used FSCIL Benchmark Settings proposed in FSCIL. If you have an issue, It would be helpful to use the data split index file from the [link](https://github.com/xyutao/fscil, which is widely used in FSCIL to benchmark scenarios.

[YvoGao]

As other FSCIL methods do, we used FSCIL Benchmark Settings proposed in FSCIL. If you have an issue, It would be helpful to use the data split index file from the [link](https://github.com/xyutao/fscil, which is widely used in FSCIL to benchmark scenarios.

We mean there are some wrong of your code. The loss is nan and the ACC is quite low, only 1-2%.

[pgh2874]

I've found that it is due to the Torch version. Try using the Torch version 1.12 or 1.11. I conducted all the experiments using Torch version 1.12.
Please let me know if the loss still contains NaN values after updating the Torch version.

[YvoGao]

Slove the problem that loss is nan, but the ACC is still low, I use the default parameters.
Namespace(Dataset=<module 'dataloader.cifar100.cifar' from '/data/gaoyunlong/FSCIL/PriViLege/dataloader/cifar100/cifar.py'>, ED=True, ED_hp=0.1, LT=True, MP=False, PKT_tune_way=1, SKD=True, WC=False, base_class=60, base_mode='ft_dot', baseline=False, batch_size_base=128, batch_size_new=0, clip=False, comp_out=1, dataroot='/data/gaoyunlong/dataset/', dataset='cifar100', debug=False, decay=0.0005, dp=False, episode_query=15, episode_shot=1, episode_way=15, epochs_base=5, epochs_new=3, fraction_to_keep=0.1, ft=False, gamma=0.1, gpu='0', l2p=False, low_shot=1, low_way=15, lp=False, lr_base=0.0002, lr_new=0.0002, lrg=0.1, milestones=[20, 30, 45], model_dir=None, momentum=0.9, new_mode='avg_cos', not_data_init=False, num_classes=100, num_gpu=1, num_workers=4, out='PriViLege', prefix=False, pret_clip=False, project='base', rotation=False, save_path='checkpoint/PriViLege/cifar100/base_ViT_Ours/ft_dot-avg_cos-data_init-start_0/Epo_5-Lr_0.0002-COS_80-Gam_0.10-Bs_128-Mom_0.90-Wd_0.00050-seed_1-T_16.00', schedule='Cosine', scratch=False, seed=1, sessions=9, set_no_val=False, shot=5, start_session=0, step=80, taskblock=2, temperature=16, test_batch_size=128, train_episode=50, vit=True, way=5)
epoch:000,lr:0.0002,training_loss:3.63284,training_acc:0.13897,test_loss:3.12547,test_acc:0.19867
epoch:001,lr:0.0002,training_loss:3.06128,training_acc:0.23677,test_loss:2.86275,test_acc:0.25467
epoch:002,lr:0.0001,training_loss:2.77224,training_acc:0.30250,test_loss:2.65879,test_acc:0.29983
epoch:003,lr:0.0001,training_loss:2.52510,training_acc:0.35493,test_loss:2.53084,test_acc:0.33217
epoch:004,lr:0.0000,training_loss:2.32806,training_acc:0.40283,test_loss:2.40860,test_acc:0.36167
Session 0, Test Best Epoch 4,
best test Acc 36.1670

Session 1, test Acc 23.831

Session 2, test Acc 22.143

Session 3, test Acc 19.693

Session 4, test Acc 18.600

Session 5, test Acc 16.918

Session 6, test Acc 15.978

Session 7, test Acc 15.095

Session 8, test Acc 13.840

Base Session Best Epoch 4

[25.567, 23.831, 22.143, 19.693, 18.6, 16.918, 15.978, 15.095, 13.84]

[YvoGao]

FileNotFoundError: [Errno 2] No such file or directory: '/data/pgh2874/FSCIL/Ours/dataloader/miniimagenet/map_clsloc.txt'

[pgh2874]

Low accuracy can result from the different TiMM versions. I recommend using timm==0.6.7.

For the FileNotFoundError, I've uploaded the map_clsloc.txt file to this repository. Make sure to change the directory path to match your directory path.