local rnnlib = require 'rnnlib'
local lstm = nn.LSTM{ inputsize = 256, hidsize = 512, nlayer = 3, usecudnn = false }

local rnnlib = require 'rnnlib'
local lstm   = rnnlib.makeRecurrent{
    cellfn    = rnnlib.cell.LSTM,
    inputsize = 256,
    hids      = { 512, 512, 512 },
}

local rnnlib = require 'rnnlib'
local lstm   = rnnlib.makeCudnnRecurrent{
    cellstring = 'LSTM',
    inputsize  = 256,
    hids       = { 512, 512, 512 },
}

local rnnlib = require 'rnnlib'
-- The table of cells is fed to each level of the recurrent network to construct each layer.
-- The table of initialization functions helps with the construction of the hidden inputs.
local cells, initfunctions = {}, {}
cells[1], initfunctions[1] = rnnlib.cell.LSTM(256, 512)
cells[2], initfunctions[2] = rnnlib.cell.LSTM(512, 512)
cells[3], initfunctions[3] = rnnlib.cell.LSTM(512, 512)
local lstm = nn.SequenceTable{
    dim = 1,
    modules = {
        nn.RecurrentTable{ dim = 2, module = rnnlib.cell.gModule(cells[1]) },
        nn.RecurrentTable{ dim = 2, module = rnnlib.cell.gModule(cells[2]) },
        nn.RecurrentTable{ dim = 2, module = rnnlib.cell.gModule(cells[3]) },
    },
}
rnnlib.setupRecurrent(lstm, initfunctions)

local mutils    = require 'rnnlib.mutils'
-- The vocabulary size.
local vocabsize = 10000
-- The dimensionality of the last hidden layer.
local lasthid   = 512
-- The dimensionality of the input embeddings.
local insize    = 256
-- The sequence length.
local seqlen    = 32
-- The batch size.
local bsz       = 32
-- The lookup table.
local lut       = nn.LookupTable(vocabsize, insize)
-- The decoder.
local decoder   = nn.Linear(lasthid, vocabsize)
-- The full model.
local model     = nn.Sequential()
    :add(mutils.batchedinmodule(lstm, lut))
    :add(nn.SelectTable(2))
    :add(nn.SelectTable(-1))
    :add(nn.JoinTable(1))
    :add(decoder)
model:cuda()
-- This returns a flattened view of the output tensor.
-- If you want this to be of a different shape, you can add an nn.View at the end.
-- Generate the input to the model.
local input = torch.range(1, seqlen * bsz)
    :resize(seqlen, bsz)
    :cudaLong()
-- Create and initialize hiddens to zero.
lstm:initializeHidden(bsz)
-- Perform the forward pass.
local output = model:forward{ lstm.hiddenbuffer, input }
-- This is just an example, normally you would not use the output as the gradOutput.
-- But the gradOutput should have the same shape as the model output.
model:backward({ lstm.hiddenbuffer, input }, output)
model:updateParameters(0.1)