pydnn 0.0.dev

deep neural network library in Python

pydnn is a deep neural network library written in Python using Theano (symbolic math and optimizing compiler package). I wrote it as a learning project while competing in Kaggle’s National Data Science Bowl in March 2015 (where it produced an entry finishing in the top 6%) and plan to continue developing it by adding support for the most important deep learning techniques (including RNNs).

Design Goals

• Simplicity

  Wherever possible simplify code to make it a clear expression of underlying deep learning algorithms. Minimize cognitive overhead, so that it is easy for someone who has completed the deeplearning.net tutorials to pickup this library as a next step and easily start learning about, using, and coding more advanced techniques.

• Completeness

  Include all the important and popular techniques for effective deep learning and not techniques with more marginal or ambiguous benefit.

• Ease of use

  Make preparing a dataset, building a model and training a deep network only a few lines of code; enable users to work with NumPy rather than Theano.

• Performance

  Should be roughly on par with other Theano neural net libraries so that pydnn is a viable choice for computationally intensive deep learning.

Features

• High performance GPU training (courtesy of Theano)

• Quick start tools to instantly get started training on inexpensive Amazon EC2 GPU instances.

• Implementations of important new techniques recently reported in the literature:

  • Batch Normalization

  • Parametric ReLU activation function,

  • Adam optimization

  • AdaDelta optimization

  • etc.

• Implementations of standard deep learning techniques:

  • Stochastic Gradient Descent with Momentum

  • Dropout

  • convolutions with max-pooling using overlapping windows

  • ReLU/Tanh/sigmoid activation functions

  • etc.

Usage

First download and unzip raw image data from somewhere (e.g. Kaggle). Then:

import pydnn
import numpy as np
rng = np.random.RandomState(e.rng_seed)

# build data, split into training/validation sets, preprocess
train_dir = 'home\ubuntu\train'
data = pydnn.data.DirectoryLabeledImageSet(train_dir).build()
data = pydnn.preprocess.split_training_data(data, 64, 80, 15, 5)
resizer = pydnn.preprocess.StretchResizer()
pre = pydnn.preprocess.Rotator360(data, (64, 64), resizer, rng)

# build the neural network
net = pydnn.nn.NN(pre, 'images', 121, 64, rng, pydnn.nn.relu)
net.add_convolution(72, (7, 7), (2, 2))
net.add_dropout()
net.add_convolution(128, (5, 5), (2, 2))
net.add_dropout()
net.add_convolution(128, (3, 3), (2, 2))
net.add_dropout()
net.add_hidden(3072)
net.add_dropout()
net.add_hidden(3072)
net.add_dropout()
net.add_logistic()

# train the network
lr = pydnn.nn.Adam(learning_rate=pydnn.nn.LearningRateDecay(
            learning_rate=0.006,
            decay=.1))
net.train(lr)

From raw data to trained network (including specifying network architecture) in 25 lines of code.

Short Term Goals

• Implement popular RNN techniques.

• Integrate with Amazon EC2 clustering software (such as StarCluster).

• Integrate with hyper-parameter optimization frameworks (such as Spearmint and hyperopt).

Authors

Isaac Kriegman