mlpclassifier 0.1.4

An implementation of an MLP classifier (with an interface of that of scikit-learn's MLPClassifier class.

Neural-Network-From-Scratch-COSC-221-CSB

Links:

• github link
• pypi link

Documentation and derivation:

• Multi-Layered perceptron derivation from scratch

Neural Network to classify handwritten digits (a rite of passage project at this point lol).

We will try to re-implement a stripped down version of the MLPClassifier class from scikit-learn from first principles. With this, we can then train a general classifier using the Multi-Layered Perceptron model.

To run

So since we've re-implemented an MLP using scikit-learn's MLPClassifier as a template, the API should be familiar.

Downloading the package

pip install mlpclassifier

Or create a virtual environment to download it.

Using uv:

uv add mlpclassifier

To use the library:

from classifier.classifier import *
# to load the MNIST dataset.
images = get_images_fast("dataset/train-images.idx3-ubyte")
labels = get_labels_fast("dataset/train-labels.idx1-ubyte")

test_images = get_images_fast("dataset/t10k-images.idx3-ubyte")
test_labels = get_labels_fast("dataset/t10k-labels.idx1-ubyte")

# pre-processing
X = images.reshape(images.shape[0], -1)/255
X_test = test_images.reshape(test_images.shape[0], -1)/255

model = MLPClassifier(
        hidden_layer_sizes=(128, 64),
        activation='relu',
        max_iter=1000,
        alpha=1e-3,
        batch_size=200,
        verbose=True,
)

SAVE_PATH = "weights/self_trained.pkl"

# if Ctrl-C, then it will save to that file path
model.fit(X, labels, save_path=SAVE_PATH)

N = 10_000
score, incorrect_indicies = model.score(X_test[:N], test_labels[:N])
print("score: ", score)
model.save(SAVE_PATH)

Loading the saved model:

with open("weights/sklearn_weights_and_biases.pkl", 'rb') as file:
    weights_and_biases = pickle.load(file)
    model.load_weights(weights_and_biases["weights"],
                       weights_and_biases["biases"],
                       weights_and_biases["classes"])

Results

Digits MNIST

Accuracy: 97.47%

Code:

model = cl.MLPClassifier(
    hidden_layer_sizes=(512, 256, 128),
    activation="relu",
    learning_rate="adaptive",
    max_iter=200,
    alpha=1e-3,
    verbose=True
)

model.fit(X, Y, save_path="./saved_models/digit_classifier.pkl")
model.score(X_test, Y_test)

Fashion MNIST

Accuracy: 86.61%

code:

model = cl.MLPClassifier(
    hidden_layer_sizes=(512, 256, 128),
    max_iter=1_000,
    activation="relu",
    learning_rate="adaptive",
    alpha=10e-3,
    verbose=True
)

X = images.reshape(len(images), -1) / 255
X_test = test_images.reshape(len(test_images), -1)/255

Y = labels
Y_test = test_labels

model.fit(X, Y, save_path="./saved_models/mnist_fashion.pkl")
model.score(X_test, Y_test)

Dataset

Download the dataset from Kaggle

curl -L https://www.kaggle.com/api/v1/datasets/download/hojjatk/mnist-dataset -o ./dataset.zip

Then just unzip it into a directory called ./dataset

unzip -d dataset ./dataset.zip

Optional, but clean redundancy:

rm -r *-idx*-ubyte

I've removed some duplicates, so currently I have:

$ ls ./dataset/
 t10k-images.idx3-ubyte   train-images.idx3-ubyte
 t10k-labels.idx1-ubyte   train-labels.idx1-ubyte

So it seems like by convention:

• we divide our dataset into training data, and then testing data
• currently, it seems like we have 60k training examples and 10k testing examples
• they do this to see how well the model has generalized

TODO

• debug all the row vector stuff
• package it in pip
• document the API

Forward propagation

• variable L for layer
• a list $n^{[l]}$ for the size at each layer
• initialize using He's initalization
• forward propagation step using that forward propagation formula

Backward propagation

• He's initialization
• back propagation
• scoring
• saving
• make the learn rate $\alpha$ more adjustable