There is a technique called Network Pruning, which is set Model parameters to zero, but won't lose a lot accuracy.
It will let the network lightweight, and decrease training time.
- Model : RESNET18
- Dataset : CIFAR10
To set up enviroment
$pip install -r ./requirements.txtThere are two method below
We set the parameters which small than the threshold to 0, let the numbers of zero is x%(x is a hyper-parameter)
We will set some row of RESNET model to 0, and retrain the model.
| Methods | Accuracy | Pruning rate |
|---|---|---|
| Ratio | 98.2% | 90% |
| Coarse | 70% | 85% |