clnn

OpenCL backend for Torch nn neural networks library.

Installation

Please see distro-cl for installation instructions.

What works

Parameterized Modules

• nn.Linear

Basic Tensor methods

These mostly 'just work', since based on underlying tensor methods, already implemented in cltorch. Tested with:

• nn.Narrow

Miscellaneous modules

• nn.Identity
• nn.Dropout

Convolution layers

• nn.SpatialConvolutionMM
• nn.SpatialMaxPooling (including ceil mode)
• nn.SpatialAveragePooling
• nn.TemporalConvolution2 This is specific to clnn. It works on cpu and cuda too, not just on OpenCL. It is API-compatible with TemporalConvolution, and faster than TemporalConvolution, on both CUDA and OpenCL.

Transfer function layers

• nn.Tanh
• nn.Sigmoid
• nn.ReLU
• nn.ELU
• nn.Exp
• nn.Sqrt
• nn.Square
• nn.Abs
• nn.LogSigmoid
• nn.HardTanh
• nn.LogSoftMax
• nn.SoftMax (including spatial mode)

Table layers

These 'just work', since they are based on underlying torch operations, which are already implemented in cltorch. Tested with:

• nn.CMulTable
• nn.CAddTable

Criterions

• nn.MSECriterion
• nn.ClassNLLCriterion

Containers:

Containers 'just work', since they just call standard operations on the contained modules. Tested with:

• nn.Sequential
• nngraph

Trainers

In theory, trainers 'just work', since they just call standard torch methods on the network. The following are good first choices:

• nn.StochasticGradient
• optim.lbfgs
• optim.adam

Timings

Soumith benchmark layers

Please see https://github.com/soumith/convnet-benchmarks#imagenet-winners-benchmarking

• On a Titan X, OpenCL torch is about 3 times slower than CUDA torch
  • eg for VGG, cutorch takes 1100ms, and cltorch takes 3400ms

Example networks

• Andrej's char-rnn is OpenCL-enabled, simple add option -opencl 1
• Justin's neural-style has an OpenCL port in progress by Shubhanshu napsternxg/neural-style

Porting guidelines

Porting guidelines, for project maintainers, available here: porting-guidelines.md.

Recent changes

• 2nd May:
  • Re-applied:
    • 26th March:
      • add TemporalConvolution2: same API and usage as TemporalConvolution, but faster on GPUs
• 31st April:
  • Re-applied:
    • 10th March:
      • @pawni (Nick Pawlowski) added SpatialUpSamplingNearest. Thank you Nick
    • 20th February:
      • @gloine (Jaehyung Lee) added support for non-batched input to ClassNLLCriterion. Thank you Jaehyung
• 30th April:
  • rolled back to as-of 21st February, prior to lots of THNN changes in upstream Torch
  • additionally, installation procedure is now to use a specific torch distro, for stability
• 1st Feb:
  • merged/ported THNN phase 3. Any weird build issues, please update both nn and clnn.
• 2nd January, 2016:
  • merged/ported THNN architecture across, and the implementation of Abs, so the unit-tests pass again now
• 15th December:
  • merged Sergey's SpatialAveragePadding and ceil kernels into master branch
• 29th November:
  • added ELU
• 25th September:
  • ported Sergey's not-yet-merged SpatialAveragePadding and ceil kernels, into clnn-avgpool branch
  • ported latest version of SoftMax, ie essentially Jonghoon's Update SoftMax to work in spatial mode
• 23rd September:
  • ported latest cunn implementation of SpatialMaxPooling across, ie approximately Sergey's Deterministic max-pooling PR
    • this includes :ceil() implementation
• 22nd September:
  • added non-batch implementation of LogSoftMax (previously only handled batched input)
  • added SoftMax, for both batched and non-batched
• 20th September:
  • added non-batch implementation for SpatialMaxPooling (previously only handled batched input), for contiguous pools

Older changes