/Quantized_CNN

Training code for quantized CNN using qKeras

Primary LanguagePython

Quantized CNN training

Training of quantized CNN on MNIST,fashion-MNIST and SVHN datasets. Quantization provided by the Google Keras extension provided at github.com/google/qkeras

Dependencies

For training: Python 3.6, TensorFlow version >= 2.1.0, Keras version: 2.2.4-tf, QKeras (https://github.com/google/qkeras). QKeras is a sub-module of this repositry. To build and install:

cd qkeras/
python3 setup.py build
python3 setup.py install --user
cd ..

Training

To train choose one of the provided *.yaml config files or create a new one. Specific architecture (see models.py), number of filters,kernel size, strides, loss function, L1 regularization can be defined in the config.

To train use the command:

python3 train.py -c quantized_cnn.yml

W.I.P: To evaluate performance:

python3 compare_models.py -c quantized_cnn/

Hyperparameter scan (Bayesian Optimisation)

This method depends on the GPy and GPyOpt packages and currently only support Python 2.7 (?). First, install these with

pip install gpy --user
pip install GPyOpt --user

Then you can do

python2 hyperParamScan.py

The hyperparameters to be optimized are listed in the dictionary "bounds" and can be configured as desired