Suggest Architecture

Information

From update 0.2.6 the neural network can suggest architecture for you which might speed up getting the best possible architecture, note this is not perfect.

To suggest, prepare input data, select how much lr do you want batch sizes, epochs, depth and temperature than call .train_with_suggested_architecture().

Example

from okrolearn.okrolearn import *
from okrolearn.optimizers import *

# Generate some dummy data
np.random.seed(0)
input_size = 20
output_size = 3
num_samples = 1000

X = np.random.randn(num_samples, input_size)
y = np.random.randint(0, output_size, size=num_samples)

assert X.shape[1] == input_size, f"Expected input size {input_size}, but got {X.shape[1]}"

# One-hot encode the targets for classification
y_one_hot = np.eye(output_size)[y]
print(f"Shape of y_one_hot: {y_one_hot.shape}")

# Convert one-hot encoded targets back to class indices
y_class_indices = np.argmax(y_one_hot, axis=1)

# Create and train the network
network = NeuralNetwork(temperature=0.8)
losses = network.train_with_suggested_architecture(
    inputs=X,
    targets=y_class_indices,
    input_size=input_size,
    output_size=output_size,
    task_type='classification',
    data_type='tabular',
    depth=5,
    epochs=5,
    lr=0.2,
    batch_size=32,
)

print(f"Final loss: {losses[-1]}")
outputs = network.forward(Tensor(X))
print(outputs)