Neuvol

Genetic algorithm to find the best neural network architecture with Keras

Neuvol is a genetic algorithm API for generating neural networks based on Keras. The main idea is to work with data only, without direct architecture constructing.

Grow your own neural network!

• Data in -> Neural Network Architecture out
• A large number of allowed layers types
• Complex crossing approaches
• Modular structure
• Keep the whole mutation history, resolve "dead" mutations such as cycling connections
• All possible architecture manipulations: add/remove layer/connection
• All possible structures: branches, skip-connections, combination of layers with different dimensions
• On the fly shape analysis: add reshape for concatenations or for connecting layer with 3 output dimensions and layer with 1 input (like dense after convolution)

Features:

• Supported data types: texts, images
• The list of supported layers is constantly expanding and contains most popular of them

TODO

• [x] Architectures distribution generation
• [x] Images support
• [x] More available layers
• [x] Logo
• [x] Serialiser
• [x] Complex layer generation
• [ ] Experimental study
• [ ] Visualization
• [x] Pytorch support