Tensorflow Wrappers for Neural Networks
Neural Networks can handle:
- Points represented by a fixed-dimensionality vector;
- 2D-Space maps representing, for example, images;
- Time-depedent functions representing, for example, time series, sound signals (not yet in this toolbox!).
Python 2.7 or 3.x and Tensorflow (1.1 or later)
In this toolbox, we provide wrappers for:
Tensorboard visualizations: use
sh tensorboard.shto print an URL that you should copy-paste in your favorite web browser (Chrome for example).
Linear Units with activations: see example.py (useful for vectors)
Convolutions with activations: see conv_example.py (useful for images)
Dropout: see both examples with the keep_prob and dropout variables (useful for regularization)
Batch-Normalization: activate (or not) the batch_normalization boolean at the beginning of each code examples (it dramatically improves the training convergence speed)
We provide 3 examples:
- example.py: Basic Neural Network (MLP = Multi-Layered Perceptron) applied for MNIST handwritten digits image classification
- conv_example.py: Convolutional Neural Network (CNN) applied for MNIST handwritten digits image classification
- rnn_example.py: Recurrent Neural Network (RNN) for pseudo-periodic time series classification -- not yet.
Warith HARCHAOUI, Astrid MERCKLING -- MAP5 -- Université Paris Descartes -- 2017