==> Current Class: [50, 51, 52, 53, 54, 55, 56, 57, 58, 59] ==> Building model.. in_features: 512 out_features: 50 current net output dim: 60 old net output dim: 50 [50, 51, 52, 53, 54, 55, 56, 57, 58, 59] [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59] Constructing exemplar set for class-50... exemplar set shape: 33 Done Constructing exemplar set for class-51... exemplar set shape: 33 Done Constructing exemplar set for class-52... exemplar set shape: 33 Done Constructing exemplar set for class-53... exemplar set shape: 33 Done Constructing exemplar set for class-54... exemplar set shape: 33 Done Constructing exemplar set for class-55... exemplar set shape: 33 Done Constructing exemplar set for class-56... exemplar set shape: 33 Done Constructing exemplar set for class-57... exemplar set shape: 33 Done Constructing exemplar set for class-58... exemplar set shape: 33 Done Constructing exemplar set for class-59... exemplar set shape: 33 Done start self-distillation for original model..... setting optimizer and scheduler................. Traceback (most recent call last): File "main_imagenet.py", line 472, in train(model=net, old_model=old_net, epoch=args.epochs, optimizer=optimizer, scheduler=scheduler, lamda=args.lamda, train_loader=trainLoader, use_sd=False, checkPoint=50) File "main_imagenet.py", line 151, in train exemplar_set = ExemplarDataset(exemplar_sets, transform=transform_ori) File "/home/ubuntu/Desktop/Alex/IL/essentials_for_CIL/data/data_loader_imagenet.py", line 17, in init self.data = np.concatenate(data, axis=0) File "<array_function internals>", line 6, in concatenate ValueError: all the input arrays must have same number of dimensions, but the array at index 0 has 4 dimension(s) and the array at index 1 has 1 dimension(s)

I expect to get the result of CCIL+H-Aug 71.66, but I get 68.25.

I conduct the following command

python main_cifar.py  \
        --new-classes 10 \
        --start-classes 50 \
        --cosine \
        --kd \
        --w-kd 1 \
        --epochs 120 \
        --exp-name 'kd_ep120_50_5' \
        --save \
        --num-sd 0\
        --aug  

Results: [79.56, 74.01666666666667, 69.98571428571428, 65.1375, 61.63333333333333, 59.16] Avg accuracy: 68.24886904761904

Do I miss some details?

Expect your reply. Thanks

Hi Sudhanshu,

Thank you for your nice work and codebase.

If possible, can you please share the version of ImageNet 1000 that you have used? Is it ImageNet 2017? If you have the download link handy, it would be very helpful if you could share them.

Thanks!

Hi, I think the following codes seem not proper: https://github.com/sud0301/essentials_for_CIL/blob/c7eb378c822792977d96daa5dd98c0202263ec44/main_cifar.py#L321-L337 After debugging I find that when features[idx, :] = 0.0 implemented, exemplar_features is also set to 0, thus S is always full of 0 at any time, making it meaningless. I am not sure whether it is a common problem or it is triggered due to my incompatible version of packages. Besides, what if the random saving strategy get the same index? Is it right the purpose of setting the corresponding feature to 0, so that in the next phase the repetitive sample has a large distance and will be deleted? But still, what if class_mean itself is close to 0?

In 'main_imagenet.py' , line 321:

x = Variable(transform(Image.fromarray(img))).cuda()

You use Image.fromarray(img) to load a array type image object. However, channel order in img is BGR (opencv) while the channel order in Image is RGB, which means this operation will confuse the channel order.