LSTM APIs in Tensorflow

• BasicRNNCell – A vanilla RNN cell.
• GRUCell – A Gated Recurrent Unit cell.
• BasicLSTMCell – An LSTM cell based on Recurrent Neural Network Regularization. No peephole connection or cell clipping.
• LSTMCell – A more complex LSTM cell that allows for optional peephole connections and cell clipping.
• MultiRNNCell – A wrapper to combine multiple cells into a multi-layer cell.
• DropoutWrapper – A wrapper to add dropout to input and/or output connections of a cell.

and the contributed RNN cells and wrappers:

• CoupledInputForgetGateLSTMCell – An extended LSTMCell that has coupled input and forget gates based on LSTM: A Search Space Odyssey.
• TimeFreqLSTMCell – Time-Frequency LSTM cell based on Modeling Time-Frequency Patterns with LSTM vs. Convolutional Architectures for LVCSR Tasks
• GridLSTMCell – The cell from Grid Long Short-Term Memory.
• AttentionCellWrapper – Adds attention to an existing RNN cell, based on Long Short-Term Memory-Networks for Machine Reading.
• LSTMBlockCell – A faster version of the basic LSTM cell (Note: this one is in lstm_ops.py)

Multi LSTM == Stacked LSTM

Multi Dimension LSTM == grid LSTM

No peephole connection or cell clipping

peephole connection: 那些gate會不會參考cell state