Recurrent and Fully Connected Kernels Group

This group describes operations related to the RNN calculations and includes general fully connected kernel description. There are a lot of variations of RNN Cells architectures. MLI provides the most popular ones as a separate kernel. The RNN Dense kernel in combination with elementwise and transform kernels might be used to construct Recurrent cells which are not present in MLI.

Functions in this group typically use the mli_rnn_cell_cfg structure, defined as:

typedef struct {
   mli_rnn_direction direction;
   mli_rnn_results results;
   mli_rnn_out_activation act;
   mli_data_container scratch_data;
 } mli_rnn_cell_cfg;

mli_rnn_cell_cfg Structure Field Description

Field Name	Type	Enumeration Value	Description
direction	mli_rnn_direction (enumeration)	RNN_DIR_FORWARD	Process the input sequence in forward direction (from the first entity to the last one).
		RNN_DIR_BACKWARD	Process the input sequence in backward direction (from the last entity to the first one).
results	mli_rnn_results (enumeration)	RNN_OUT_LAST	Preserve only the last result after all RNN iterations in output tensor
		RNN_OUT_ALL	Preserve result of each RNN iteration in output tensor
act	mli_rnn_out_activation RNN output activation type (enumeration)	RNN_ACT_TANH	Hyperbolic tangent activation function
		RNN_ACT_SIGM	Logistic (sigmoid) activation function
		RNN_ACT_NONE	No activation
scratch_data	mli_data_container	–	Container with a scratch memory to keep cell’s intermediate results. Must contain a valid pointer in pi32 field (see mli_tensor Structure) to a memory of sufficient size for the kernel. The exact amount of memory is defined in the respective kernel descriptions in subsequent sections.

• Fully Connected Prototype and Function List
• RNN Dense Prototype and Function List
• Basic Long Short Term Memory (LSTM) Cell Prototype and Function List
• Gated Recurrent Unit (GRU) Cell Prototype and Function List