scheckmedia/keras-shufflenet

ShuffleNet Implementation using Keras Functional Framework 2.0

keras-shufflenet's Issues

Group Convolution

@scheckmedia ,

How fast is your grouped Convolution code? I've seen the lambda implementation elsewhere and it was pretty slow.

Do you have similar results?

Thanks for this great implementation

Execution time for forward pass

Hey! Can you do me a favor and check the execution time of a single forward pass for batch size = 1? I am not getting nearly what is stated in the paper of about 37.8ms for ShuffleNet 1x (g=3).
Even though I an using ShuffleNet 1x (g=8), I am getting over 1.35s per inference iteration.

Edit x1: Thanks!
Edit x2: I am using an Intel Xeon E5-2680 v3 @ 2.50GHz which is a fairly decent processor. The paper states the benchmarks were performed on a Qualcomm Snapdragon 820 (single thread) which we should easily be able to surpass given we are using desktop/server CPUs.

What `ShuffleNet_1X_g3_br_0.25_373.hdf5` file is trained from?

is there pre-trained model to share？

hi，thx for sharing your work。
i want to try the shuffle-net on my own dataset(small dataset), so i am searching for pre-trained model. would you kindly enough to share your model file with its acc and configuration (like shuffle 1x g3)??

BN after first Conv2D

Hey! There is supposed to be a batch-norm between the first Conv2D and the 3x3 max-pool in stage1. The pre-trained weight file that I'm using (ShuffleNetV1-1x-g=8.npz) contains these parameters (gamma, beta, mean and variance for BN for conv1).
Could you please confirm if there is a BN in your pretrained model? I can't check since the link to your model is dead.

Thanks

Dont use GPU

when i do training, it only use 1% GPU, and it says
Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX

Do you know how to fix it

License

Hi!

The project has the MIT license, but no copyright notice is given, which I find slightly confusing.

In addition, would you consider adding the "LICENSE.txt" file? While not needed, it's helpful: it specifies which MIT license in particular (there are variations) and it can be parsed automatically, which is useful.

Pre-trained weights link leads to 404

Hi Tobias,

I tried to download your weights file but got 404 page.
Would you mind checking it out?

Appreciate it!

Validation Accuracy fluctuating alot

Validation accuracy is fluctualtuating alot between epochs. What can be the reason?