alonsopg/NumpyDL

Deep Learning Library. For education. Based on pure Numpy. Support CNN, RNN, LSTM, GRU etc.

NumpyDL: Numpy Deep Learning Library

Descriptions

NumpyDL is:

1. Based on Pure Numpy/Python
2. For DL Education
3. And for My Homework

Features

Its main features are:

1. Pure in Numpy
2. Native to Python
3. Automatic differentiations are basically supported
4. Commonly used models are provided: MLP, RNNs, LSTMs and CNNs
5. API like Keras library
6. Examples for several AI tasks
7. Application for a toy chatbot

Documentation

Available online documents:

1. latest docs
2. development docs
3. stable docs

Available offline PDF:

1. latest PDF

Installation

Install NumpyDL using pip:

$> pip install npdl

Install from source code:

$> python setup.py install

Examples

NumpyDL provides several examples of AI tasks:

• sentence classification

  • LSTM in examples/lstm_sentence_classification.py
  • CNN in examples/cnn_sentence_classification.py

• mnist handwritten recognition

  • MLP in examples/mlp-mnist.py
  • MLP in examples/mlp-digits.py
  • CNN in examples/cnn-minist.py

• language modeling

  • RNN in examples/rnn-character-lm.py
  • LSTM in examples/lstm-character-lm.py

One concrete code example in examples/mlp-digits.py:

import numpy as np
from sklearn.datasets import load_digits
import npdl

# prepare
npdl.utils.random.set_seed(1234)

# data
digits = load_digits()
X_train = digits.data
X_train /= np.max(X_train)
Y_train = digits.target
n_classes = np.unique(Y_train).size

# model
model = npdl.model.Model()
model.add(npdl.layers.Dense(n_out=500, n_in=64, activation=npdl.activation.ReLU()))
model.add(npdl.layers.Dense(n_out=n_classes, activation=npdl.activation.Softmax()))
model.compile(loss=npdl.objectives.SCCE(), optimizer=npdl.optimizers.SGD(lr=0.005))

# train
model.fit(X_train, npdl.utils.data.one_hot(Y_train), max_iter=150, validation_split=0.1)

Applications

NumpyDL provides one toy application:

• Chatbot

  • seq2seq in applications/chatbot/model.py

And its final result:

Supports

NumpyDL supports following deep learning techniques:

• Layers

  • Linear
  • Dense
  • Softmax
  • Dropout
  • Convolution
  • Embedding
  • BatchNormal
  • MeanPooling
  • MaxPooling
  • SimpleRNN
  • GRU
  • LSTM
  • Flatten
  • DimShuffle

• Optimizers

  • SGD
  • Momentum
  • NesterovMomentum
  • Adagrad
  • RMSprop
  • Adadelta
  • Adam
  • Adamax

• Objectives

  • MeanSquaredError
  • HellingerDistance
  • BinaryCrossEntropy
  • SoftmaxCategoricalCrossEntropy

• Initializations

  • Zero
  • One
  • Uniform
  • Normal
  • LecunUniform
  • GlorotUniform
  • GlorotNormal
  • HeNormal
  • HeUniform
  • Orthogonal

• Activations

  • Sigmoid
  • Tanh
  • ReLU
  • Linear
  • Softmax
  • Elliot
  • SymmetricElliot
  • SoftPlus
  • SoftSign