This repository contains Python notebooks with code for optimizing AlexNet, VGG, and U-Net neural networks using the LCRM optimization method based on our research work "LCRM: Layer-wise Complexity Reduction Method for CNN Model Optimization on End Devices, IEEE Access, 2023 ". We also include the source code for the ablation study to demonstrate the effectiveness of the optimization method and an appendix section with additional code examples.
To get started, first clone this repository:
git clone https://github.com/han-an/LCRM.git
The code in this repository requires Python 3.6 or later and the following packages:
- Tensorflow-gpu >2.0
- NumPy
- Matplotlib
- TensorBoard (Profiler plugin)
Required packages are given at the beginning of each notebook.
- EMNIST used in LCRM-VGG
- CIFAR from tensorflow dataset used in LCRM-AlexNet
- Kaggle datascience Bowl dataset used in LCRM-UNet
To use the optimization method, open the appropriate Python notebook for the network you want to optimize (e.g., LCRMVGG9.ipynb or LCRMUnet.ipynb). Run the notebook cells in order to train the model with the optimization method. We have included both the parent model and the LCRM-generated model for comparison.
The code and results for ablation study are presented in the ablation_study.ipynb notebook.
The appendix Python notebook contains code examples for using the optimization method with other set of convolutions exist in the literature: The substitution layers studied include standard convolution (Std Conv), transpose convolution (TrC), Groupwise convolution (GrC), depthwise convolution (DWC), pointwise convolution (PWC), and dilated convolution (DiC) and their combinations for checking the number of parameters.
A portion of U-Net parent model code is referred from https://github.com/bnsreenu/python_for_microscopists.