cudnn.torch

Torch7 FFI bindings for NVIDIA cuDNN (R5) kernels!

Modules are API compatible their nn equivalents. Fully unit-tested against nn implementations. Conversion between nn and cudnn is available through cudnn.convert function.

Installation

• Install cuDNN (version R5 EA)

• Have at least CUDA 7.0

• Have libcudnn.so in your library path ($LD_LIBRARY_PATH) (Install cuDNN it from https://developer.nvidia.com/cuDNN )

• Instead of the previous step, you can copy the library files into /usr/local/cuda/lib64/ or to the corresponding folders in CUDA directory

Modules

-- All inputs have to be 3D or 4D(batch-mode), except ReLU, Tanh, Sigmoid, and BatchNormalizationcudnn.SpatialConvolution(nInputPlane, nOutputPlane, kW, kH, [dW = 1], [dH = 1], [padW = 0], [padH = 0], [groups = 1])
cudnn.SpatialMaxPooling(kW, kH, dW, dH, padW, padH)
cudnn.SpatialAveragePooling(kW, kH, dW, dH, padW, padH)-- the pointwise functions take an additional optional argument. if inplace=true then they do operations in-place without using any extra memory for themselvescudnn.ReLU(inplace[=false])
cudnn.ClippedReLU(ceiling, inplace[=false])
cudnn.Tanh(inplace[=false])
cudnn.Sigmoid(inplace[=false])-- SoftMax can be run in fast mode or accurate mode. Default is accurate mode.cudnn.SoftMax(fastMode [= false])          -- SoftMax across each image (just like nn.SoftMax)cudnn.LogSoftMax()                         -- LogSoftMax across each image (just like nn.LogSoftMax)cudnn.SpatialSoftMax(fastMode [= false])   -- SoftMax across feature-maps (per spatial location)cudnn.SpatialLogSoftMax()                  -- LogSoftMax across feature-maps (per spatial location)cudnn.VolumetricSoftMax(fastMode [= false])   -- SoftMax across feature-maps (per spatial location)cudnn.VolumetricLogSoftMax()                  -- LogSoftMax across feature-maps (per spatial location)cudnn.SpatialCrossEntropyCriterion()       -- A spatial version of LogSoftMax + ClassNLLCriterion in one shotcudnn.VolumetricCrossEntropyCriterion()       -- A volumetric version of LogSoftMax + ClassNLLCriterion in one shot-- Batch Normalizationcudnn.BatchNormalization(nFeature, eps, momentum, affine) -- same arguments as https://github.com/torch/nn/blob/master/doc/simple.md#nn.BatchNormalizationcudnn.SpatialBatchNormalization(nFeature, eps, momentum, affine)
cudnn.VolumetricBatchNormalization(nFeature, eps, momentum, affine)-- Volumetric inputs (4D or 5D batched mode)cudnn.VolumetricConvolution(nInputPlane, nOutputPlane, kT, kW, kH, dT, dW, dH, padT, padW, padH)
cudnn.VolumetricMaxPooling(kT, kW, kH, dT, dW, dH, padT, padW, padH)
cudnn.VolumetricAveragePooling(kT, kW, kH, dT, dW, dH, padT, padW, padH)-- Recurrent Modules-- All inputs have to be 3D. Accepts input of seqLength x batch x inputDim, or batch x seqLength x inputDim if batchFirst set to true.cudnn.RNNReLU(inputDim, outputDim, numberOfLayers, [batchFirst = false])
cudnn.RNNTanh(inputDim, outputDim, numberOfLayers, [batchFirst = false])
cudnn.LSTM(inputDim, outputDim, numberOfLayers, [batchFirst = false])
cudnn.GRU(inputDim, outputDim, numberOfLayers, [batchFirst = false])
cudnn.BLSTM(inputDim, outputDim, numberOfLayers, [batchFirst = false])

Modes

There are two globally availabe modes useful for tuning performance:

require 'cudnn'cudnn.benchmark = true -- uses the inbuilt cudnn auto-tuner to find the fastest convolution algorithms.
                       -- If this is set to false, uses some in-built heuristics that might not always be fastest.

by default cudnn.benchmark is set to false.  Setting to true will improve performance, at the expense of using more memory.  The input shape should be the same for each batch, otherwise autotune will re-run for each batch, causing a huge slow-down.

cudnn.fastest = true -- this is like the :fastest() mode for the Convolution modules,
                     -- simply picks the fastest convolution algorithm, rather than tuning for workspace size

by default, cudnn.fastest is set to false.  You should set to true if memory is not an issue, and you want the fastest performance

cudnn.verbose = true -- this prints out some more verbose information useful for debugging

by default, cudnn.verbose is set to false.

Conversion between cudnn and nn

Conversion is done by cudnn.convert function which takes a network and backend arguments and goes over network modules recursively substituting equivalents. No memory copy is done, just metatables are swapped. If you don't want to convert all modules you can pass a function as the third argument to cudnn.convert. It will be called at each step, with a module that is currently converted.  It is meant to exclude modules i.e. if it returns true, they will be left untouched, otherwise they will be subject to conversion.

Note that you cannot do backward pass when using cuDNN and when your model has batch normalization layers and is in evaluate mode.

net = nn.Sequential()
net:add(nn.SpatialConvolution(3,96,11,11,3,3))
net:add(nn.ReLU())
cudnn.convert(net, cudnn)print(net)

net = nn.Sequential()
net:add(nn.SpatialConvolution(3,96,11,11,3,3))
net:add(nn.ReLU())
cudnn.convert(net, cudnn, function(module)   return torch.type(module):find('ReLU')end)print(net)

will result in:

nn.Sequential {
  [input -> (1) -> (2) -> output]
  (1): cudnn.SpatialConvolution(3 -> 96, 11x11, 3,3)
  (2): cudnn.ReLU
}
nn.Sequential {
  [input -> (1) -> (2) -> output]
  (1): cudnn.SpatialConvolution(3 -> 96, 11x11, 3,3)
  (2): nn.ReLU
}

Older versions

For version CuDNN R1, checkout the branch R1For version CuDNN R2, checkout the branch R2For version CuDNN R3, checkout the branch R3For version CuDNN R4, checkout the branch R4