Pure-numpy version of Keras for educational purposes.

• Keras-like model fitting procedure in nn.py
• Activation functions (ReLU, ELU, Softmax) in activations.py
• Early Stopping and ModelCheckpoint callbacks in callbacks.py
• Multiple layers available (Dense, BatchNorm, Dropout) in layers.py
• Different Optimizers (Momentum SGD, RMSProp, Adam) and learning-rate schedulers in optim.py

import mlp
from mlp.nn import NeuralNetwork, Dense
from mlp.tools import one_hot_encoder
from mlp.metrics import cv_score, accuracy
from sklearn.datasets import load_iris

X, y = load_iris(return_X_y=True)

y_enc, _ = one_hot_encoder(y)

model = NeuralNetwork(loss='crossentropy')
model.add(Dense(3, activation='softmax'))
model.fit(X, y_enc, n_epochs=50)