maple-research-lab / cappronet-pytorch Goto Github PK

Thanks for sharing of your wonderful work.

I tried the following script to run experiment with resnet (pre-activation):

python main.py data_dir --dataset cifar10 --arch preresnet164 --save_path ./test --epochs 200 --schedule 60 120 160 --gammas 0.2 0.2 0.2 --learning_rate 0.1 --decay 0.0001 --batch_size 128 --Ddim 4

The best test error is 6.44% (mean of 3 runs), worse than the reported 4.88% in the paper, also worse than the resnet baseline as 5.46%.

Training original resnet (pre-activation) with the same configuration achieves a best test error as 5.4%, better than reported 5.46%.

Is there any mistake in my script?

why it always interrupt?

why it happen:
Process finished with exit code 139 (interrupted by signal 11: SIGSEGV)?
thanks

When the last layer of the fully connected layer becomes CapProNet, how do you set up the initialization of the network? Or just change the CapProNet part, do not change the initialization of the network?
thanks.