Torch has undergone some changes in it's backend since I last updated this repository. Hence, the installation now need to be performed inside nnx, instead of nn directly.
I have updated the installation instructions accordingly.
In this repository I will be maintaining a list of modules that I had to come up with to ease my work. While starting out to work with Volumetric (4D or batchmode 5D) data I found a lot of implementations missing, even though their Spatial versions existed. This repository is the collection of the codes that I wrote then.
Hopefully it will be of help to someone, someday, somewhere.
Most of these are simple extensions of their Spatial counter-parts. So there's a high chance that even though it wasn't there when I wrote it, it may have been incorporated in the official library now. So it is highly recommended that you cross-check with Torch's main repository to see if it already exists. I will notify the same, below each module, if I happen to come across one.
To use this modules along with the exisiting modules of the nn and cunn package requires installing them in a bit hackish (read dirty) way.
It is explained over here in detail.
It is important to note that there should not be any name conflicts with any of the existing modules. For example, suppose you want to use a module abc.lua. Then it is important to make sure that there is no other module called abc.lua present in the nn (or cunn package).
Performs Volumetric Convolution that supports padding
nn.VolumetricConvolution(nInputPlane, nOutputPlane, kT, kW, kH, [dT], [dW], [dH], [padT], [padW], [padH])This module is already merged into the main Torch repository.
Associated files
Performs Batch Normalization over 5D (batch mode) data
nn.VolumetricBatchNormalization(N [,eps] [, momentum] [,affine])where N = Number of input features. Details regarding the other parameters could be found over
here
This module is already exists in the main Torch repository.
Associated files
Performs 3D upsampling on input videos containing any number of input planes.
nn.VolumetricUpSamplingNearest(scale_t, scale_xy)where scale_t = upsample ratio along time domain. scale_xy = upsample ratio along height, width dimension. Must be positive integers. I have personally used this to perform unpooling. Use case in Spatial domain could be found in the dc-ign paper.
Associated files
Performs 3D Volumetric Max Unpooling using indices previously computed with Volumetric Max Pooling.
nn.VolumetricMaxUnpooling(max_module)Example usage
max_module = nn.VoulmetricMaxPooling(kT, kW, kH)
model = nn.Sequential()
model:add(max_module)
model:add(nn.VolumetricMaxUnpooling(max_module))Currently this is only the CPU version. I will update the CUDA version soon.
Note: It requires kT==dT, kW==dW, kH==dH.
This module is already merged into the main Torch repository.
Associated files
Performs Volumetric Max Pooling that supports padding
nn.VolumetricMaxPooling(nInputPlane, nOutputPlane, kT, kW, kH, [dT], [dW], [dH], [padT], [padW], [padH])This module is already merged into the main Torch repository.
Associated files
I will keep on adding stuffs as and when required. If you need anything in particular, you are most welcome to ask about it. Also feel free to give suggestions, comments. It will be much appreciated.
And finally I would like to thank all the wonderful Torch's contributors for actively maintaining such a wonderful and easy-to-use library. Really appreciate their efforts and hard work.