Codes for Stability and Generalization of Bilevel Programming in Hyperparameter Optimization.
The codes are implemented on
- Python=3.7
- PyTorch=1.4.0
python fl_omniglot.py # run feature learning experiments
python rw_mnist.py # run data reweighting experiments
The experiment to run is controlled by the tag
variable in the above *.py
files. For instance,
tag = ablo_K
will run the UD algorithm with differentK
tag = rs_K
will run the random search algorithm with differentK
tag = wdh
will run the UD algorithm with different weight decay in the outer leveltag = wdl
will run the UD algorithm with different weight decay in the inner level
You can also specify other settings by modify the args
variable in the above *.py
files.