bug information :

Epoch: 0, LR: [0.1]
Train set: 196, Train Loss: 3.4775 Acc: 9.5960
/opt/conda/conda-bld/pytorch_1556653215914/work/aten/src/THCUNN/ClassNLLCriterion.cu:105: void cunn_ClassNLLCriterion_updateOutput_kernel(Dtype *, Dtype *, Dtype *, long *, Dtype *, int, int, int, int, long) [with Dtype = float, Acctype = float]: block: [0,0,0], thread: [4,0,0] Assertion t >= 0 && t < n_classes failed.

I solved this by modified

to

tg_model = resnet_cifar.resnet32(num_classes=args.num_classes)

I'm not sure if I am correct。？？？？

Where can we find the supplementary material?
Thanks.

Hi,

I was able to successfully trained the model and want to test it on a separate test set. However, when I load the saved checkpoint, the output layer only has nb_cl_fg units, instead of the total number of classes. I have a total of 114 classes and loading the saved model only outputs 40 units (my nb_cl_fg).

How will I be able to recover all 114 classes in the output? Thank you.

Hello, Thank you for your work.

cifar100-class-incremental/class_incremental_cosine_cifar100.py mentioned that the class_means[:,current_cl[iter_dico],0] is for iCaRL (mean of exemplars) and class_means[:,current_cl[iter_dico],1] is for NCM (theoretical mean).

But in cifar100-class-incremental/eval_cumul_acc.py, you use the class_means[:,current_cl[iter_dico],1] as the final result of your method, does it mean that you use all the historical exemplars to calculate the mean of the classes?

Can you please share the 'list of the class directories' used for the ImageNet-100 experiment. Or the original sequence in ImageNet-1000 which was used for sampling and shuffling.

Could you tell me what is the '--nb_protos' meaning, i think each class should have a prototype, why the number is fixed here?

Hello,

I want to change the number of classes in first group. Use cifar100, let the number of classes in first group is 10, and then train the 90 classes in the incremental way. Can I just modify the '--nb_cl_fg' in the argparse, or need to do something else?

Thank you.

Hello,
I found the calculation of the adaptive weight of less-forget constraint different from the description in the paper. Did I misunderstand this part or miss some details?

_, term_width = os.popen('stty size', 'r').read().split()
ValueError: not enough values to unpack (expected 2, got 0)