xyutao / bocl Goto Github PK

Hi again,

I trained session S1 of cifar10 on ResNet18, using SGD optimizer for 20 epochs with the learning rate provided in the paper. I am able to obtain only 45% at max. Can you please suggest anything I might have missed?

When I ran your code, I received an error: ModuleNotFoundError: No module named '_init_paths'

Is it a custom python file or part of any packages to be installed?

Please provide details on how the Cifar10/100 datasets were split into different sessions, and the code to replicate the illumination and saturation transformations.

Hello! Thank you for opening your implementation.
Is your experimental setting as same as described in the paper? I did not reach your experimental result according to your settings.

When I enable plc-loss, I get the following error:

line 330

Traceback (most recent call last):
  File "tools/train_cifar10_naive.py", line 375, in <module>
    main()
  File "tools/train_cifar10_naive.py", line 372, in main
    train(net, context)
  File "tools/train_cifar10_naive.py", line 330, in train
    train_plc_loss += sum([al.mean().asscalar() for al in plc_loss])
NameError: name 'plc_loss' is not defined

Thanks.