Theano, Pylearn2, libgpuarray: Sharing and Futurenyoml/theano_2.pdf · Theano, Pylearn2,...

Post on 15-Sep-2018

224 views 0 download

transcript

Theano, Pylearn2, libgpuarray: Sharing and Future

Frédéric Bastien, Bart van Merriënboer

Département d'Informatique et de Recherche Opérationnelle

Université de Montréal

Montréal, Canada

{bastienf, vanmerb}@iro.umontreal.ca

OML Workshop 2014

Future

Future

2 / 8

Future

Theano

I Easier C code development and better documentation of that

I Faster compilation

I Multi-GPU

I Better looping (update to scan)

I Allow checkpoint with GPU to reload without GPU

I Less memory allocation(lower Theano overhead)

I Faster convolution

1 / 8

Future

libgpuarray

I Find other projects to use it?

I More functionality as NumPy

I Move some of the functionality from Python/Theano to the Clevel

I Optimize the kernel selection and parametrization based onthe GPU

2 / 8

Future

Pylearn2

I RNN

I Better hyperparameter search support, using e.g. Hyperopt

I Documentation

I Checkpoint

I Better support for sparse dataset

I Machine translation examples

I Gated activations for conditional computation

I Variational Auto-Encoders

3 / 8

Future

4 / 8

Future

Simplifying code sharing between

1. License: Suggest BSD as it is used by many software in our�eld.

I Common license help share code.I When reusing code, don't forget to keep the license and the

copyright notice

2. Common base object! libgpuarray

3. Otherwise: put important implementation(e.g. convolution) inseparate �le and use raw ptr/shape/strides as inputs.Document that interface.

4. Acknowledge reuse in section on web site AND in papersabout the software we reuse! (and use too)

5 / 8

Future

Simplifying code sharing between

1. License: Suggest BSD as it is used by many software in our�eld.

I Common license help share code.I When reusing code, don't forget to keep the license and the

copyright notice

2. Common base object! libgpuarray

3. Otherwise: put important implementation(e.g. convolution) inseparate �le and use raw ptr/shape/strides as inputs.Document that interface.

4. Acknowledge reuse in section on web site AND in papersabout the software we reuse! (and use too)

6 / 8

Future

Simplifying code sharing between

1. License: Suggest BSD as it is used by many software in our�eld.

I Common license help share code.I When reusing code, don't forget to keep the license and the

copyright notice

2. Common base object! libgpuarray

3. Otherwise: put important implementation(e.g. convolution) inseparate �le and use raw ptr/shape/strides as inputs.Document that interface.

4. Acknowledge reuse in section on web site AND in papersabout the software we reuse! (and use too)

7 / 8

Future

Simplifying code sharing between

1. License: Suggest BSD as it is used by many software in our�eld.

I Common license help share code.I When reusing code, don't forget to keep the license and the

copyright notice

2. Common base object! libgpuarray

3. Otherwise: put important implementation(e.g. convolution) inseparate �le and use raw ptr/shape/strides as inputs.Document that interface.

4. Acknowledge reuse in section on web site AND in papersabout the software we reuse! (and use too)

8 / 8