junyuseu / pytorch-cifar-models Goto Github PK

💯

Thanks for sharing such a good framework. I'm curious that when I'm using WideResNet with depth-k (26-10), it actually has less GPU usage than DenseNet-BC(k=40) even though WideResNet has more trainable params. Am I missing something? Hope for your clarification.

That's a great framework of different networks.
When I use resnet20, I input my own dataset through ImageFolder. When I resized the image as (32,32), it worked but when I set the size of input images as (64,64), I got an error at: output = model(input_var).

The error is size mismatched. m1:[128,5148],m2:[64,42].
128 is my batch size and 42 is the number of classes. Any help?

Hey, I really like this repository and would like to use your model implementations for a research paper. For this I need a licence. Could you add a licence to you repository (for example a MIT Licence). Thanks a lot.

The original paper of wideresnet use 40-4, 16-8, 28-10, while your code use 20-10,26-10

In models/wide_resnet_cifar.py

from resnet_cifar import BasicBlock

should be

from .resnet_cifar import BasicBlock

How many epoch you need to get these results?
I rand resneXt 200 epochs but didn't achieve your results.

Any ideas?

Can you please share which models needs 2 GPUs to train, and which can train on only one?

Thanks

I ran directly in the author code and only achieved 94.95 accuracy. I wonder how wrn-10's 3.89% accuracy is achieved