JiyueWang

Results 15 comments of JiyueWang

How to early stop the update of branch weights as the paper said?

Can you show some results about how serious the overfitting problem is?

Model Accuracy

I trained seresnet32 and got 24.5 error rate, far below the 22.07. On serenet152, the error rate is 21.8

resnet50 accuracy is a little bit worse than resnet18

PyTorch 1.4.0

resnet50 accuracy is a little bit worse than resnet18

You need to adjust the learning rate function. if epoch > args.warm: train_scheduler.step(epoch) to if epoch > args.warm: train_scheduler.step() print the learning rate by print('lr:{:5.4f}'.format(optimizer.param_groups[0]['lr'])) to check out By the...

Can not reproduce the result (both paper & yours)

I got similar results on CIFAR10. And on CIFAR100, the accuracy also 1% below the claimed 83.3%. Anyone success? "model": { "type": "wresnet28_10" }, "dataset": "cifar100", "aug": "randaugment", "randaug": {...

The prob of applying augmentation?

In addition, the Identity operation is not included

Why not updating batch norm parameters?

This line means BN layer is excluded for weight decay, not the whole BP. However, I tried the regular way and got similar results on CIFAR100. I think the link...

The Volatile GPU-Util is always 0, in examples/imagenet

Does anyone solve this problem?

validate loss cannot update

I have no problem on the cifar10 dataset, but I got the same problem as you when transfer to my own local data. Have you solved the problem?

validate loss cannot update

I have solved my problem. The reason is the trainset is not shuffled before split it. Shuffle the indices before split indices = list(range(n_train)) shuffle = 1 if shuffle: np.random.seed(42)...