| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546 |
- https://bitbucket.org/rshegde/deepmie-deep-learning-electromagnetic-scattering
- Abstract
- ===============================================================
- We propose the use of Deep Convolutional Neural Networks (DCNN)
- in concert with Differential Evolution for solving inverse problems in
- electromagnetics. We show that networks built using innovations like
- convolutional layers, batch normalization, parametric REctified Linear Unit
- (ReLU) and residual blocks achieve comparable performance to previously
- reported dense fully-connected networks but achieve drastic reduction in
- parameter weights and training epoch lengths. The proposed lightweight CNNs
- are used in concert with Differential Evolution global optimization to achieve
- 8x speedup in convergence time. Systematic hyperparameter grid searches are
- reported to permit faster model discovery for application to other problems.
- See the following arxiv paper for details
- Requirements
- ================================================================
- 1) CUDA capable GPU
- 2) Python 3.5
- 3) Numpy, Matplotlib, Pandas, Scipy, Scikitlearn
- 4) Cudatooklit 9.0, CuDNN 7.1
- 5) Mxnet-GPU, Mxnet-cumkl
- 6) Keras, Keras-mxnet
- 7) Scattnlay -- needed if you want to generate other datasets
- 8) Jupyter notebook
- 9) ipyparallel for calling multiple runs at the same time
- Getting started
- ==================================
- Most of the code is placed in Jupyter notebooks for interactivity.
- PyMIEtest.ipynb -- familiarizes you with generating datasets
- Scatternets.ipynb -- allows building and visualizing models
- hypoptim.ipynb -- hyperparameter grid search using scikit CV grid search
- scipydiffe -- differential evolution using compiled models.
|
