Skip to content

Latest commit

 

History

History
27 lines (26 loc) · 2.45 KB

README.md

File metadata and controls

27 lines (26 loc) · 2.45 KB

How to install:

  • Add the path to the src directory of this project in PYTHONPATH environment variable.
    • Example .bashrc: export PYTHONPATH="$PYTHONPATH:/path/to/neural-wrappers/src"

Structure of this project:

  • examples/
    • mnist-cifar10-classifier - Simple MNIST and Cifar10 classifier using 2 networks (FC and Conv). See this for basic usage of the framework
    • char-rnn - Implementation of a simple recurrent network that predicts one character after another. Inspired by: http://karpathy.github.io/2015/05/21/rnn-effectiveness/
    • video-autoencoder - Implementation of a simple video auto-encoder, that takes two videos (say 240p and 120p) and creates a convolutional + LSTM network that tries to compress the given video. Can be used to generate a new video of a smaller dimension (or even increase resolution, if a better model would be implemented)
    • variational-autoencoder - Implementation of a standard VAE, a reconstruction CVAE and a class-based CVAE, tested on MNIST
    • generative-adversian-network - Implementation of the standard GAN, tested on MNIST
  • neural_wrappers/
    • dataset_reader/ - Base class for all readers and various readers for known datasets.
    • models/ - Various models from different papers implemented. May contain additional or missing features from original articles (just PyTorch for now).
    • transforms/ - Basic transforms and some built-in transforms for data augmentation (mirror, cropping)
    • wrappers/ - Main wrapper directory
      • pytorch/ - Files that implement various features on top of PyTorch framework
    • callbacks.py - Basic callback class and some built-in callbacks for training (history.txt, model saving)
    • metrics.py - Basic metrics and some built-in metrics for training (accuracy/loss)
    • utils.py - Various functions
  • reader_converters - Various converters from the form the dataset is offered online (usually a big archive of images and labels, textual or not) to the form that is accepted by an implemented reader (under neural_wrappers/readers), which uses the DatasetReader class API (compatible with Keras fit_generator method and NeuralNetworkPyTorch train_generator). Generally, these converters will generate one h5py file that is used in the reader.
  • test/
    • Unit tests for all the implemented modules (WIP)
    • To run tests, go in the tests directory and type 'pytest' in the console. Requires the pytest module to be installed, which can be done using pip install pytest
  • README.md - this file