This project aims to train CNN models for image classification tasks and reduces the computational overhead of large models by compressing the weights of deep neural networks using Keras.
- Basic CIFAR10 CNN classifier for Benchmarking
- Parameter pruning
- Paper - Learning both Weights and Connections for Efficient Neural Networks
- Compresses the weights of CNN by removing weights with small absolute value
- Tensor decomposition
- Paper - Compression of Deep Convolutional Neural Networks for Fast and Low Power Mobile Applications
- Applies Tucker decomposition to the weights of a pre-trained CNN to reduce computations
- Decomposes a single Convolution layer into a number of convolutional layers where total number of computations are lesser.
Use the package manager pip to install keras and related dependencies.
- tensorly - used for Tensor Decompositions
pip install keras
pip install tensorly- Learning both Weights and Connections for Efficient Neural Networks - Song Han, Jeff Pool, John Tran, William J. Dally
- Compression of Deep Convolutional Neural Networks for Fast and Low Power Mobile Applications - Yong-Deok Kim, Eunhyeok Park, Sungjoo Yoo, Taelim Choi, Lu Yang, Dongjun Shin
- Review on tensor decompositions