removed some benchmarking code to DeepLearningBenchmarks on github

benchmark/README.txt

-The benchmarking folder contains efforts to benchmark Theano against various
-other systems.  Each subfolder corresponds to a particular type of computation,
-and each sub-subfolder corresponds to the implementation of that computation in
-with a particular software package.
-
-Since there is a variety of benchmark problems and of software systems, there
-isn't a standard for how to run the benchmark suite.
-There is however a standard for how each benchmark should produce results.
-Every benchmark run should produce one or more files with the results of
-benchmarking. These files must end with extension '.bmark'.  These files must
-have at least three lines each:
-
-1) line 1 - description of computation/problem
-2) line 2 - description of implementation/platform
-3) line 3 - time required (in seconds)
-4) line 4 - [optional] an estimated number of FLOPS performed (not necessarily same for all implementations of problem)
-

benchmark/mlp1/torch5/mlp_sgd.lua

-
-require "lab"
-require "os"
-dataset={};
-n_examples=1000;
-inputs=1000;
-outputs=12;
-HUs=500;
-
-function dataset:size() return n_examples end -- 100 examples
-for i=1,dataset:size() do 
-  local input = lab.randn(inputs);     -- normally distributed example in 2d
-  local output = lab.randn(outputs);
-  dataset[i] = {input, output}
-end
-
-require "nn"
-mlp = nn.Sequential();  -- make a multi-layer perceptron
-mlp:add(nn.Linear(inputs, HUs))
-mlp:add(nn.Tanh())
-mlp:add(nn.Linear(HUs, outputs))
-
-
-criterion = nn.MSECriterion()  
-trainer = nn.StochasticGradient(mlp, criterion)
-trainer.learningRate = 0.01
-trainer.shuffleIndices = true
-trainer.maxIteration = 4
-trainer:train(dataset)
-