NumpyNet

High level neural network API implementated using numpy.

The project is my homework in deeplearning course at NCTU. Anyone can trace the code to learn how to do the backpropagation on sequential model and how to build a basic deeplearning API with python.

Requirements

• python3

Install

$ git clone https://github.com/w86763777/numpynet
$ cd numpynet
$ python setup.py install

Example

from numpynet.dataset import iris, split_dataset
from numpynet.models import SequentialModel
from numpynet.optimizers import Adam
from numpynet.loss import CrossEntropy
from numpynet.metrics import categorical_accuracy
from numpynet.layers import Input, Dense, ReLU, Softmax, Dropout


if __name__ == "__main__":
    # load iris dataset
    iris = iris.read_data_sets()
    # split dataset
    train, test = split_dataset(iris, test_size=0.33)

    # build model
    model = SequentialModel()
    model.add(Input((4,)))
    model.add(Dense(10))
    model.add(ReLU())
    model.add(Dropout(0.3))
    model.add(Dense(10))
    model.add(ReLU())
    model.add(Dropout(0.3))
    model.add(Dense(3))
    model.add(Softmax())

    # assign objective, optimizer and metrics which is going to be shown on
    # progress bar
    model.compile(
        objective=CrossEntropy(),
        optimizer=Adam(learning_rate=0.001),
        metric=[categorical_accuracy])
    
    # fit on data
    model.fit(
        x=train.X, y=train.y, val_x=test.X, val_y=test.y,
        epochs=500, batch_size=8)

output

Epoch 1/500
100%|█████████████| 13/13 [00:00<00:00, 1441.88it/s, categorical_accuracy=0.2900, cross_entropy=1.0986, val_categorical_accuracy=0.3600, val_cross_entropy=1.0982]
Epoch 2/500
100%|█████████████| 13/13 [00:00<00:00, 1288.82it/s, categorical_accuracy=0.4200, cross_entropy=1.0968, val_categorical_accuracy=0.3600, val_cross_entropy=1.0982]
...
Epoch 500/500
100%|█████████████| 13/13 [00:00<00:00, 1296.02it/s, categorical_accuracy=0.6900, cross_entropy=0.8285, val_categorical_accuracy=0.9600, val_cross_entropy=0.2492]

more examples

How it work

• TODO

Issues

• regularization deos not work