Optimizer for configuration of hyperparameters in neural networks.
What does this library do? - Module can optimize hyperparameters of a neural network for a pre-defined architecture.
What deep learning libraries can this module work with? - PyTorch.
What algorithm is used for optimization? - An evolutionary algorithm with mutation and crossover operators is used. The neural network is continuously trained in the process of evolution.
'python>=3.7',
'numpy',
'cudatoolkit==10.2',
'torchvision==0.7.0',
'pytorch==1.6.0'
Description of the submodules:
For now all the necessary description can be found in docstring.
How to run the algorithm can be seen in the examples:
- FNN classification task - MNIST classification (The effectiveness of MIHA is compared with the optuna framework)
- CNN regression task - gap-filling in remote sensing data (The effectiveness of MIHA is compared with init neural network training without hyperparameters search) (in russian)
Feel free to contact us:
-
Mikhail Sarafanov | mik_sar@mail.ru