DynamicView module for Torch that views an input tensor dynamically at runtime with user-provided function of its size. See an example of total variation computation below.

DynamicView.lua

function DynamicView(getSizeTable)
    local module = nn.View(-1)
    module.updateOutput = function(self, input) return nn.View.updateOutput(self:resetSize(unpack(getSizeTable(input:size()))), input) end
    return module
end

-- function TotalVariation() --accepts 4D tensors
--   local sk = torch.Tensor(2, 1, 2,2)
--     sk[1][1]:copy(torch.Tensor(2,2):set(torch.Storage({-1, 1, 0, 0})))
--     sk[2][1]:copy(torch.Tensor(2,2):set(torch.Storage({-1, 0, 1, 0})))
--     local conv = cudnn.SpatialConvolution(1,2, 2,2, 1, 1,0, 0):noBias()
--     conv.weight:copy(sk)
--     local batchSize
--     return nn.Sequential()
--         :add(DynamicView(function(sz) batchSize = sz[1]; return {sz[1] * sz[2], 1, sz[3], sz[4]} end))
--         :add(conv)
--         :add(nn.Abs())
--         :add(DynamicView(function(sz) return {batchSize, -1} end))
--         :add(nn.Mean(2))
-- end