network-compression

The project is dedicated to neural networks compression. A convolutional network for classification of faces by gender is used as an example. You can get Celeba dataset for experiments here.

How to run the experiments

There are considered 3 cases:

• initial - a convolutional network (2 Conv + 4 Dense);
• cut - initial version without dense layers with both dimensionalities > 1 (2 Conv + 1 Dense);
• compressed - initial version where dense layers with both dimensionalities > 1 are compressed (2 Conv + 3 compressed Dense + 1 Dense).

1. Building a model

In oreder to build initial or cut version, you can use according functions for their construction from classifier/tf/models.py. To get compressed model, you should first train initial, save it and use classifier/tf/Compress.ipynb for compression.

2. Training

Use classifier/tf/Train.ipynb. Don't forget to specify necessary paths to downloaded Celeba dataset. Once you trained the model and saved it, you can use classifier/tf/convert_to_tflite.sh script in order to get the model in .tflite format.

3. Measure accuracy

On PC you can do it with classifier/tf/Test.ipynb. For devices see instructions below.

3.1 Android

Build in Android Studio and install on your device the application in android-app/ project. Put your trained model to assets/.

In order to measure accuracy on the phone, it has to be connected via adb (type $ adb devices to ensure that the connection is established).

Then we need to make a simple server for taking test images from PC. We can use Python for that:

from http.server import SimpleHTTPRequestHandler, HTTPServer


def run(name, port):
    server_address = (name, port)
    httpd = HTTPServer(server_address, SimpleHTTPRequestHandler)
    httpd.serve_forever()


if __name__ == "__main__":
    run('', 8000)

Run the server in the same directory as img_align_celeba/ (the directory with Celeba images). In order to allow the phone to access the server, type:

$ adb reverse tcp:8000 tcp:8000

Finally, press START button in the application. After the process is complete, you can find a celeba_test.csv file with resutls in the phone's Document/ folder.

3.2 Raspberry Pi

Connect to Raspberry Pi via ssh.
Upgrade pip and install TensorFlow and Keras.

pip install --upgrade pip  
pip install tensorflow
pip install keras

For testing model run rpi_test.py. In this file you should specify a path to dataset and saved model and choose a name of output csv-file.

For getting results run results.py. In this file you should specify a path to csv-file mentioned above and to list_attr_celeba.csv.
This script prints ROC AUC score of the model, mean of classification time and standard deviation of this time.