Convolutional Feature Extractor (Autoencoder Model)

This repository is intended to serve as a reference to a front-end convolutional auto-encoder implementation. Specifically, it contains the complete pipeline for training, inference testing and weights extraction functions.

This repository also includes extraction functions that can be used to pickle the trained weights, biases, input and output activations for validation purposes.

How to Use

# Clone repository
>> git clone https://github.com/lchingen/conv-feature-extract
>> cd conv-feature-extract

# Download dataset (TFRecords)
>> https://tinyurl.com/y5sqxybu
>> mkdir db
>> cp celeb-face/* ./db

# Train feature extractor
>> mkdir logs
>> python3 auto-encoder/train.py

# Extract and dump trained weight, bias, ia, oa
>> mkdir dump
>> python3 auto-encoder/extract-weight.py
>> python3 auto-encoder/extract-input.py

References & Notes

• To generate TFRecords from scratch, please reference the following repository:

# Clone repository
>> git clone https://github.com/lchingen/tf-data-pipe
>> rm db/*

# Download dataset
>> wget http://www.kaggle.com/jessicali9530/celeba-dataset/kernels ./db

• Source description:
  • config.py: training configuration
  • model_fn.py: DNN model
  • train.py: train the DNN model
  • utils.py: utility functions
  • test-img.py: test single image inference on trained network
  • test-webcam.py: test webcam streaming and DNN inference results
  • extract-weight.py: extract W and B based on restored DNN model
  • extract-input.py: extract normalized IA and W after inference