NeuralFit is a simple neuro-evolution library for Python with an API that is similar to the Keras library. You can train models with evolution instead of backpropagation in just a few lines of code. You can also export evolved models to Keras so you can use them in your existing environments without changing a thing. For more information, please visit https://neuralfit.net/. NeuralFit is currently in the alpha phase and being developed rapidly, expect many bugs and crashes 🐛.

Installation

NeuralFit is tested and supported on 64-bit machines running Windows or Ubuntu with Python 3.7-3.10 installed. You can install NeuralFit via pip with version 19.3 or higher.

pip install neuralfit

If you want to export models to Keras, make sure you also have tensorflow installed!

Get started

As an example, you can evolve a model to learn the XOR gate as follows.

import neuralfit as nf
import numpy as np

# Define dataset
x = np.asarray([[0,0], [0,1], [1,0], [1,1]])
y = np.asarray([[0], [1], [1], [0]])

# Define model (2 inputs, 1 output)
model = nf.Model(2, 1)

# Compile and evolve
model.compile(loss='mse', monitors=['size'])
model.evolve(x,y)

To verify that your model has fitted the dataset correctly, we can check its predictions.

print(model.predict(x))

Done with evolving your model? Simply save the model or export it to Keras!

# Save model
model.save('model.nf')

# Export model to Keras
keras_model = model.to_keras()

There is much more to discover about NeuralFit, and new features are being added rapidly. Please visit the official documentation for more information or follow one of our examples. One of the examples shows a method to visualize the evolved model, which results in the pictures below.

Recurrent models

It is also possible to evolve recurrent models, although they cannot (yet) be exported to Keras. These models can evolve backwards connections that allow models to have memory, making them ideal for timeseries prediction. The above XOR dataset can be transformed to a timeseries, where each input feature is inputted sequentially instead of all at once.

import neuralfit as nf
import numpy as np

# Define dataset
x = np.asarray([[0,0], [0,1], [1,0], [1,1]])
y = np.asarray([[0], [1], [1], [0]])

# Reshape to timeseries format
# 4 samples of length 2 with 1 feature as input
# 4 samples of length 1 with 1 feature as output
x = np.reshape(x, (4,2,1))
y = np.reshape(y, (4,1,1)) 

# Define model (1 input, 1 output)
model = nf.Model(1,1, recurrent = True)

# Compile and evolve
model.compile(loss='mse', monitors=['size'])
model.evolve(x, y, epochs=500)

After the model has evolved, we can give it timeseries and extract the last output to see what the model predicts.

# Make predictions
print(model.predict(x)[:,-1,0])

It is also possible to define a target output series that has the length of each input series, which is useful for stock market prediction for example. Please keep an eye on our examples page, as more recurrent examples will be added soon. Again, it is possible to visualize the model, which will now likely include backwards and self-connections.

Custom evaluation functions

If your task does not have a supervised dataset, you can pass your own evaluation function to be used in the evolution process. The above feedforward XOR example can be for example adapted as follows:

# Define dataset
x = np.asarray([[0,0], [0,1], [1,0], [1,1]])
y = np.asarray([[0], [1], [1], [0]])

# Define evaluation function (MSE)
def evaluate (genomes):
    losses = np.zeros(len(genomes))

    for i in range(len(genomes)):
        for j in range(x.shape[0]):
            result = genomes[i].predict(x[j])
            losses[i] += (result - y[j])**2
    
    return losses/x.shape[0]

# Define model (2 inputs, 1 output)
model = nf.Model(2, 1)

# Compile and evolve
model.compile(monitors=['size'])
model.func_evolve(evaluate)

This can also be done for tasks where recurrency (i.e. memory) is required. Just make sure to clear() the genomes before each series, as is done in the recurrent XOR example.

Issues?

Found a bug? Want to suggest a feature? Or have a general question about the workings of NeuralFit? Head over to the issues section and feel free to create an issue. Note that any questions about licenses or payments should be sent to info@neuralfit.net.

Become a supporter

Unfortunately, it is not possible to sustain the research and development for NeuralFit withour your help. Please consider supporting us by buying a license subscription, which also unlocks various extra features. We promise that all NeuralFit features will be available for everyone if other sources of income allow for it 💚.