sud0301 / essentials_for_cil Goto Github PK

I can't find where is "coarse_label_names" used in the code?

I want to use the checkpoint.pth , load it and and want to do predition , can you please share a small snippet for this

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, I think the following codes seem not proper:

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.

你好，请问你尝试使用常规fc层做实验而非cosinelinear吗？如果尝试过，在论文算法下，两种分类器对比怎么样？

还有在使用cosinelinear分类器，我训练自己数据集，发现sigma固定增加，模型很快（10个epoch样子）就达到最佳，继续性能就会降得很快，你在训练中是否有该现象。

==> 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)

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!

I am currently using 128 gigabytes of RAM, but out of memory comes out when I train with Imagenet. How many gigabytes of ram did you experiment with?