RandWireNN

Results

In small regime, C=78, WS(4,0.75)

datasets	top-1 acc.	top-5 acc.	epoch
MNIST	99.60	100.	100
CIFAR-10	91.71	99.75	250
CIFAR-100	72.49	92.15	250
ImageNet-12	56.70	78.95	81

In regular regime, C=109, WS(4,0.75)

datasets	top-1 acc.	top-5 acc.	epoch
ImageNet-12	work in progress

- Because my computer does not have powerful computing power, it will take some time to update.

Running the example

Setup

Please prepare:

Clone the repository and install the following additional packages:

git clone https://github.com/hebo1221/RandWireNN.git
pip install -r requirements.txt

Running the demo

Just

python run_RandWireNN.py

• If you want to change dataset, see run_RandWireNN.py, get_configuration(). MNIST,CIFAR,ImageNet available
• You don't have to prepare a dataset. The code will automatically download it.
• But if you have it already, set your dataset directory in RandWireNN_config.py, __C.DATASET_DIR
• If you want to see a train-loss graph, see RandWireNN_config.py, __C.VISDOM
• You can change the hyperparameters and dataset settings from *_config.py files. Look it up.

Reference

All details regarding the Randomly Wired Neural Networks can be found in the original research paper: https://arxiv.org/pdf/1904.01569v2.

License

All materials in this repository are released under the Apache License 2.0.