Note that Jeff Donahue's implementation will be merged to Caffe (not this code).
Jeff's code is more modularized, while this code is optimized for LSTM.
This code computes gradient w.r.t. recurrent weights in a single matrix computation.

• Speed comparison (Titan X, 3-layer LSTM with 2048 units)

  • Batch size = 20, Length = 100

  Code	Forward(ms)	Backward(ms)	Total (ms)
  This code	248	291	539
  Jeff's code	264	462	726

  • Batch size = 4, Length = 100

  Code	Forward(ms)	Backward(ms)	Total (ms)
  This code	131	118	249
  Jeff's code	140	290	430

  • Batch size = 20, Length = 20

  Code	Forward(ms)	Backward(ms)	Total (ms)
  This code	49	59	108
  Jeff's code	52	92	144

  • Batch size = 4, Length = 20

  Code	Forward(ms)	Backward(ms)	Total (ms)
  This code	29	26	55
  Jeff's code	30	61	91

An example code is in /examples/lstm_sequence/.
In this code, LSTM network is trained to generate a predefined sequence without any inputs.
This experiment was introduced by Clockwork RNN.
Four different LSTM networks and shell scripts(.sh) for training are provided.
Each script generates a log file containing the predicted sequence and the true sequence.
You can use plot_result.m to visualize the result.
The result of four LSTM networks will be as follows:

• 1-layer LSTM with 15 hidden units for short sequence
• 1-layer LSTM with 50 hidden units for long sequence
• 3-layer deep LSTM with 7 hidden units for short sequence
• 3-layer deep LSTM with 23 hidden units for long sequence