Skip to content

soumith/cuda-convnet2.torch

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

cuda-convnet2.torch

Torch7 bindings for cuda-convnet2 kernels! Kept as a separate repo because of the License, and because the codebase is not small.

**This is a Work IN PROGRESS! ** DONT USE any modules which are not listed below

####Modules that are usable:

ccn2.SpatialConvolution(nInputPlane, nOutputPlane, kH, [dH = 1], [padding = 0], [groups = 1], [partialSum = oH * oH])
ccn2.SpatialConvolutionLocal(nInputPlane, nOutputPlane, inputHeight, kH, [dH = 1], [padding = 0])
ccn2.SpatialMaxPooling(kW, [dW = kW])
ccn2.SpatialAvgPooling(kW, [dW = kW])
ccn2.SpatialCrossResponseNormalization(nCrossFeaturemaps, [addScale = 0.0001], [powScale = 0.75], [minDiv = 1])

####What's left to do?

All the modules from here: https://code.google.com/p/cuda-convnet/wiki/LayerParams

####How to do it? it is pretty simple,

  • Add the function signature from cudaconv3/include into ffi.lua
  • Call the function in your lua module

For an example, look at SpatialConvolution.lua,

How to use them?

Either send in an input of layout Depth x Height x Width x Batch, or wrap around nn.Transpose modules

Example

fSize = {3, 96, 128, 128, 384}
features = nn.Sequential()
features:add(nn.Transpose({1,4},{1,3},{1,2}))
features:add(ccn2.SpatialConvolution(fSize[1], fSize[2], 9))
features:add(nn.ReLU())
features:add(ccn2.SpatialMaxPooling(2,2))
features:add(ccn2.SpatialConvolution(fSize[2], fSize[3], 5))
features:add(nn.ReLU())
features:add(ccn2.SpatialMaxPooling(2,2))
features:add(ccn2.SpatialConvolution(fSize[3], fSize[4], 4))
features:add(nn.ReLU())
features:add(ccn2.SpatialConvolution(fSize[4], fSize[5], 3))
features:add(nn.ReLU())
features:add(ccn2.SpatialMaxPooling(2,2))
features:add(nn.Transpose({4,1},{4,2},{4,3}))
features:add(nn.View(featuresOut))

###NVMatrix to THTensor cheatsheet

NVMatrix THCudaTensor
.getNumCols() .size[1]
.getNumRows() .size[0]
.getNumElements() THCudaTensor_nElement()
.getNumDataBytes() THCudaTensor_nElement() * 4
.getStride() .stride[0]
.isTrans() N/A
.getDevData() THCudaTensor_data()
.resize() THCudaTensor_resizeXd where X = dims
.getTextureObject() THCudaTensor_getTextureObject
.isContiguous THCudaTensor_isContiguous
.isSameDims THCudaTensor_isSameSizeAs
.apply THCudaTensor_fill()
  • check contiguity of all tensors, if not, make contiguous
  • ignore/remove assertions (because you are doing contiguous checks anyways)
  • harmonize getTextureObject. destroy all the texture objects after usage, treat them like pointers. NVMatrix does it in it's destructor, but since the object is not a member of the THCudaTensor structure, we have to destroy it manually after use.
  • double-check places where strides are allowed (especially conv) Agg = ?, Agg.getBaseValue, Agg.output(.., ..)
  • Remember that NVMatrix only supports 2D tensors!