A simple fully connected feed forward neural network written in python from scratch using numpy. It is possible to have multiple hidden layers, change the amount of neurons per layer & have a different activation function per layer.
Written in python 3.7.7
If you have any tips on how to imporve performace, let me know!
import numpy as np
from numba.experimental import jitclass
from numba import types, typed
data_input = np.load("data/ci_inputs.npy")
data_output = np.load("data/ci_outputs.npy")
print("Begin compiling!")
begin_time = time.time_ns()
compile_nn = make_neural_network(layer_sizes=[data_input.shape[1], data_output.shape[1]], layer_activations=["sigmoid"])
compile_nn.train(data_input[:1], data_output[:1], data_input[1: 2], data_output[1: 2])
end_time = time.time_ns()
print("Compile time:", (end_time-begin_time) / 1e9)
for i in range(10):
random_seed = np.random.randint(10, 1010)
np.random.seed(random_seed)
train_input, validate_input, test_input = h.kfold(4, data_input, random_seed)
train_output, validate_output, test_output = h.kfold(4, data_output, random_seed)
nn = make_neural_network(layer_sizes=[train_input.shape[1], 20, train_output.shape[1]], layer_activations=["sigmoid", "sigmoid"])
begin_time = time.time_ns()
epochs, current_mse = nn.train(train_input, train_output, validate_input, validate_output)
end_time = time.time_ns()
train_mse = nn.calculate_MSE(train_input, train_output)
test_mse = nn.calculate_MSE(test_input, test_output)
accuracy_test = nn.evaluate(test_input, test_output)
print("Seed:", random_seed, "Epochs:", epochs, "Time:", (end_time-begin_time)/1e9, "Accuracy:", accuracy_test, "Tr:", train_mse, "V:", current_mse, "T:", test_mse)