Okrolearn

Checkout my machine learning library, which is a raw implementation of combining pytorch with scikit-learn.
https://github.com/Okerew/okrolearn
Why did I make this project? I made it as I saw problems with pytorch, there weren't any data analasys featurues, some more algortihms could be implemented, better support for sparse tensors with scipy, use of cupy, easier creation of cuda kernels. A view to simplify, use a lot more of python, create better support for cpus and MacOS.

Can be installed with pip install okrolearn

---

Example usage

from okrolearn.okrolearn import *


def print_epoch_start(epoch, total_epochs):
    print(f"Starting epoch {epoch + 1}/{total_epochs}")


network = NeuralNetwork(temperature=0.5)
network.add(DenseLayer(3, 4))
network.add_hook('pre_epoch', print_epoch_start)
network.add(ReLUActivationLayer())
network.add(DenseLayer(4, 4))
network.add(LinearActivationLayer())
network.add(LeakyReLUActivationLayer(alpha=0.1))
network.add(DenseLayer(4, 3))
network.add(ELUActivationLayer())
network.add(SoftsignActivationLayer())
network.add(HardTanhActivationLayer())
network.remove(2)
network.add(SoftmaxActivationLayer())

inputs = Tensor(np.random.rand(100, 3))
targets = Tensor(np.random.randint(0, 3, size=(100,)))
loss_function = CrossEntropyLoss()
optimizer = SGDOptimizer(lr=0.01, momentum=0.9)

losses = network.train(inputs, targets, epochs=100, lr=0.01, batch_size=10, loss_function=loss_function)

# Plot the training loss
network.plot_loss(losses)

network.save('model.pt')

test_network = NeuralNetwork()
test_network.add(DenseLayer(3, 4))
test_network.add_hook('pre_epoch', print_epoch_start)
test_network.add(ReLUActivationLayer())
test_network.add(DenseLayer(4, 4))
test_network.add(LinearActivationLayer())
test_network.add(LeakyReLUActivationLayer(alpha=0.1))
test_network.add(DenseLayer(4, 3))
test_network.add(ELUActivationLayer())
test_network.add(SoftsignActivationLayer())
test_network.add(HardTanhActivationLayer())
test_network.remove(2)
test_network.add(SoftmaxActivationLayer())

test_network.load('model.pt')

test_inputs = Tensor(np.random.rand(10, 3))
test_outputs = test_network.forward(test_inputs)
print(test_outputs)