recurrent benchmark

torch/benchmark_recurrent.md

@@ -0,0 +1,48 @@
+## Results - Recurrent package
+Tested on a Titan-x
+```
+th recurrent.lua -networkType rnn -hiddenSize 100
+Setup : compile + forward/backward x 1
+--- 0.036673069000244 seconds ---
+Forward:
+--- 100000 samples in 7.1649069786072 seconds (13956.907428054 samples/s) ---
+Forward + Backward:
+--- 100000 samples in 18.566176176071 seconds (5386.1375926401 samples/s) ---
+
+
+th recurrent.lua -networkType rnn -hiddenSize 500
+Setup : compile + forward/backward x 1
+--- 0.085377931594849 seconds ---
+Forward:
+--- 100000 samples in 12.131023168564 seconds (8243.316253702 samples/s) ---
+Forward + Backward:
+--- 100000 samples in 36.012835979462 seconds (2776.7874612969 samples/s) ---
+
+
+th recurrent.lua -networkType rnn -hiddenSize 1000
+Setup : compile + forward/backward x 1
+--- 0.20405006408691 seconds ---
+
+Forward:
+--- 100000 samples in 21.969285011292 seconds (4551.8086501038 samples/s) ---
+Forward + Backward:
+--- 100000 samples in 64.306167125702 seconds (1555.0607355506 samples/s) ---
+
+
+th recurrent.lua -networkType lstm -hiddenSize 100
+Setup : compile + forward/backward x 1
+--- 0.34961485862732 seconds ---
+Forward:
+--- 100000 samples in 26.887173891068 seconds (3719.2429748535 samples/s) ---
+Forward + Backward:
+--- 100000 samples in 67.390839099884 seconds (1483.881143735 samples/s) ---
+
+
+th recurrent.lua -networkType lstm -hiddenSize 500
+Setup : compile + forward/backward x 1
+--- 0.51357793807983 seconds ---
+Forward:
+--- 100000 samples in 29.653849840164 seconds (3372.2413593724 samples/s) ---
+Forward + Backward:
+--- 100000 samples in 102.79220104218 seconds (972.83641213161 samples/s) ---
+```

torch/recurrent.lua

@@ -0,0 +1,83 @@
+require('torch')
+require('cutorch')
+require('nn')
+require('cunn')
+require('recurrent')
+
+
+-- cutorch.setDevice(2)
+
+cmd = torch.CmdLine()
+cmd:text()
+cmd:text('Options')
+cmd:option('-nSamples', 100000, 'Number of samples')
+cmd:option('-networkType', 'lstm', 'Network type')
+cmd:option('-inputSize', 100, 'Neural network input size')
+cmd:option('-hiddenSize', 100, 'Neural network hidden layer size')
+cmd:option('-seqLength', 30, 'Sequence length')
+cmd:option('-batchSize', 20, 'Batch size')
+cmd:option('-cpu', false, 'Run on CPU')
+cmd:text()
+for k, v in pairs(cmd:parse(arg)) do _G[k] = v end
+
+local xValues = torch.rand(nSamples, seqLength * inputSize)
+local yValues = torch.rand(nSamples, hiddenSize)
+if cpu ~= true then
+   xValues = xValues:cuda()
+   yValues = yValues:cuda()
+end
+local xBatches = xValues:split(batchSize, 1)
+local yBatches = yValues:split(batchSize, 1)
+
+local a = torch.Timer()
+local rnn
+if networkType == 'rnn' then
+   rnn = nn.RNN(inputSize, hiddenSize)
+elseif networkType == 'lstm' then
+   rnn = nn.LSTM(inputSize, hiddenSize)
+else
+   error('Unkown network type!')
+end
+local rnn = nn.Sequential():add(rnn):add(nn.SelectTable(-1))
+local criterion = nn.MSECriterion()
+if cpu ~= true then
+   rnn:cuda()
+   criterion:cuda()
+end
+
+rnn:sequence()
+rnn:training()
+local input = xBatches[1]:split(inputSize, 2)
+criterion:forward(rnn:forward(input), yBatches[1])
+rnn:backward(input, criterion:backward(rnn.output, yBatches[1]))
+if cpu ~= true then cutorch.synchronize() end
+print("Setup : compile + forward/backward x 1")
+print("--- " .. a:time().real .. " seconds ---")
+rnn:evaluate()
+
+a:reset()
+for i = 1, #xBatches do
+   if (i % 1000 == 0) then
+      print(i)
+   end
+   rnn:forward(xBatches[i]:split(inputSize, 2))
+end
+if cpu ~= true then cutorch.synchronize() end
+print("Forward:")
+print("--- " .. nSamples .. " samples in " .. a:time().real .. " seconds (" .. nSamples / a:time().real .. " samples/s) ---")
+
+rnn:training()
+a:reset()
+for i = 1, #xBatches do
+   if (i % 1000 == 0) then
+      print(i)
+   end
+   local input = xBatches[i]:split(inputSize, 2)
+   criterion:forward(rnn:forward(input), yBatches[i])
+   rnn:zeroGradParameters()
+   rnn:backward(input, criterion:backward(rnn.output, yBatches[i]))
+   rnn:updateParameters(0.01)
+end
+if cpu ~= true then cutorch.synchronize() end
+print("Forward + Backward:")
+print("--- " .. nSamples .. " samples in " .. a:time().real .. " seconds (" .. nSamples / a:time().real .. " samples/s) ---")