Deeposlandia

*Train convolutional neural networks with Keras API*

This new release is characterized by the transition from the TensorFlow library to the
Keras library so as to train neural networks and predict image labels.

Additionally, the code has been structured in a production-like purpose:

- the program modules have been moved to a deeposlandia repository;
- a tests repository contains a bunch of tests that guarantee the code validity;
- a setup.py file summarizes the project description and target. Some complements may
arise in order to publish the project on Pypi.

*Validate and test the trained a wider range of TensorFlow models*

In this patch, a more mature code is provided:

- Dataset handling is factorized, we can now consider Mapillary or shape datasets
starting from a common Dataset basis
- Model handling is factorized, we can generate feature detection models or semantic
segmentation models, with common behaviors (basic layer creation, for instance)
- Some state-of-the-art models have been implemented (VGG, Inception)
- A base of code has been deployed for considering Keras API (the switch from TensorFlow
to Keras will be the object of a next minor release)

*Validate and test the trained model (Minor README fixes)*

Fix the 0.3 release with minor changes around README.md file (picture updates,
essentially).

*Validate and test the trained model*

- Add a single-batched validation phase during training process, the corresponding
metrics are logged onto Tensorboard so as to be compared with training metrics (same
graphs) ;
- Add a model inference module, that call the test() method of
ConvolutionalNeuralNetwork: it takes a trained model as an input, and infer label
occurrences on a image testing set ;
- Manage the Tensorboard monitoring in a more clever way ;
- Add the possibility to gather similar labels for Mapillary dataset: by aggregating
them, the number of labels decreases and the model may become easier to
train. :warning: With this new feature, the dataset structure in json files has been
modified: the labels keys are now dictionaries (instead of a lists) that link class ids
(keys) and label occurrences (values), for each image.

*Object-oriented convolutional neural network*

This new release provide an improved version of the project by considering
object-oriented programming.

- The project is structured around Dataset and ConvolutionalNeuralNetwork classes. These
classes are written in dedicated modules.
- As a consequence, the main module contains only program-specific code (argument
handling).
- A second dataset has been introduced to the project (geometric shapes), so as to make
development easier and more reliable.

*Street-scene object detection*

This repository runs a convolutional neural network on Mapillary Vistas Dataset, so as to
detect a range of street-scene objects (car, pedestrian, street-lights, pole,
... ). Developments are still under progress, as the model is unable to provide a
satisfying detection yet.