This repo implements LDMNet presented in the following work:
LDMNet: Low Dimensional Manifold Regularized Neural Networks Wei Zhu, Qiang Qiu, Jiaji Huang, Robert Calderbank, Guillermo Sapiro, Ingrid Daubechies
This implementation uses:
- pytorch > 0.4.1 for network models
- skorch >= 0.4.0 as a high level model training API
- sacred > 0.7.4 for argument and experiment management
To run use:
python main.py with mnist
python main.py with cifar10
python main.py with svhn
You can change parameters as follows:
python main.py with mnist train_size=1000 device=cuda dropout=0.5 alphaupdate.lambda_bar=0.01
Call python main.py print_config to see all parameters available.
Results, arguments, run info and network weights for each run will be stored in ldmnet_runs, under a directory corresponding to the run's id number.
testcommand to load previous run and evaluate on test set