Dynamic Convolution (training optimization)

Paper: Dynamic Convolution: Attention over Convolution Kernels

Implementation with reference to [1] https://github.com/kaijieshi7/Dynamic-convolution-Pytorch

The training time is about 7 times faster than [1] upper link on the cifar10 dataset.

Check

python dyconv2d.py

Training

python train.py 
    --device 0 #'cuda device, i.e. 0 or 0,1,2,3 or cpu'
    --training_optim #training more faster

Inference

just call model.inference_mode()

model = DyMobileNetV2(num_classes=opt.num_classes, input_size=32, width_mult=1.)
model.inference_mode()