deepy: Highly extensible deep learning framework based on Theano

Dependencies

• Python 2.7 (Better on Linux)
• numpy
• theano
• scipy for L-BFGS and CG optimization

Tutorials (Work in progress)

http://deepy.readthedocs.org/en/latest/

Clean interface

# MNIST Multi-layer model with dropout.
from deepy.dataset import MnistDataset, MiniBatches
from deepy.networks import NeuralClassifier
from deepy.layers import Dense, Softmax, Dropout
from deepy.trainers import MomentumTrainer, LearningRateAnnealer

model = NeuralClassifier(input_dim=28*28)
model.stack(Dense(256, 'relu'),
            Dropout(0.2),
            Dense(256, 'relu'),
            Dropout(0.2),
            Dense(10, 'linear'),
            Softmax())

trainer = MomentumTrainer(model)

annealer = LearningRateAnnealer(trainer)

mnist = MiniBatches(MnistDataset(), batch_size=20)

trainer.run(mnist, controllers=[annealer])

Examples

Enviroment setting

• CPU

source bin/cpu_env.sh

• GPU

source bin/gpu_env.sh

MNIST Handwriting task

• Simple MLP

python experiments/mnist/mlp.py

• MLP with dropout

python experiments/mnist/mlp_dropout.py

• MLP with PReLU and dropout

python experiments/mnist/mlp_prelu_dropout.py

• Deep convolution

python experiments/mnist/deep_convolution.py

• Elastic distortion

python experiments/mnist/mlp_elastic_distortion.py

• Recurrent visual attention model
  • Result visualization

python experiments/attention_models/baseline.py

Language model

Penn Treebank benchmark

• Baseline RNNLM (Full-output layer)

python experiments/lm/baseline_rnnlm.py

• Class-based RNNLM

python experiments/lm/class_based_rnnlm.py

• LSTM based LM (Full-output layer)

python experiments/lm/lstm_rnnlm.py

Char-based language models

• Char-based LM with LSTM

python experiments/lm/char_lstm.py

• Char-based LM with Deep RNN

python experiments/lm/char_rnn.py

Deep Q learning

• Start server

pip install Flask-SocketIO
python experiments/deep_qlearning/server.py

• Open this address in browser

http://localhost:5003

Auto encoders

• Recurrent NN based auto-encoder

python experiments/auto_encoders/rnn_auto_encoder.py

• Recursive auto-encoder

python experiments/auto_encoders/recursive_auto_encoder.py

Train with CG and L-BFGS

• CG

python experiments/scipy_training/mnist_cg.py

• L-BFGS

python experiments/scipy_training/mnist_lbfgs.py

Other experiments

DRAW

See https://github.com/uaca/deepy-draw

# Train the model
python mnist_training.py
# Create animation
python animation.py experiments/draw/mnist1.gz

Highway networks

• http://arxiv.org/abs/1505.00387

python experiments/highway_networks/mnist_baseline.py
python experiments/highway_networks/mnist_highway.py

Effect of different initialization schemes

python experiments/initialization_schemes/gaussian.py
python experiments/initialization_schemes/uniform.py
python experiments/initialization_schemes/xavier_glorot.py
python experiments/initialization_schemes/kaiming_he.py

Other features

• Auto gradient correction

---

Sorry for that deepy is not well documented currently, but the framework is designed in the spirit of simplicity and readability. This will be improved if someone requires.

Raphael Shu, 2015