Benchmark Update

[nicholas-leonard]

Hi @glample,

This PR provides the following updates:

• Torch uses SeqLSTM . I give explicit credit to @jcjohnson for taking the time to implement such blazing fast code (see PR#207).
• There is only one input and target batch. This allows the benchmarking of huge networks as otherwise the xValues tensor was so big that it would blow up the memory.
• Input is now of shape seqlen x batchsize x hiddensize instead of batchsize x seqlen x hiddensize this gives a slight speedup to all implementations and is the natural way to present sequential data as it makes the step-wise tensors contiguous in memory.
• hiddensize = inputsize. This is the usual use-case for RNN/LSTMs.
• Benchmarks for hiddensize of 128, 512 and 1024, batchsize of 32 and 128 as mentioned in Benchmark hidden sizes that give better efficiency and bigger batches #9 .

[nicholas-leonard]

@ma2rten Is there a way to store the input and target tensor inside a GPU tensor before sending it to TensorFlow? I think that would make your code considerably faster.

[nicholas-leonard]

@glample I think this is ready to merge. I can incorporate any tweaks later.