Gated Recurrent Unit Example

Description:

Demonstrates a gated recurrent unit (GRU) example with the use of fully-connected, Tanh/Sigmoid activation functions.

Model definition:

GRU is a type of recurrent neural network (RNN). It contains two sigmoid gates and one hidden state.

The computation can be summarized as:

z[t] = sigmoid( W_z ⋅ {h[t-1],x[t]} )
r[t] = sigmoid( W_r ⋅ {h[t-1],x[t]} ) 
n[t] = tanh( W_n ⋅ [r[t] × {h[t-1], x[t]} ) 
h[t] = (1 - z[t]) × h[t-1] + z[t] × n[t] 

Gate Recurrent Unit Diagram

Variables Description:

• update_gate_weights, reset_gate_weights, hidden_state_weights are weights corresponding to update gate (W_z), reset gate (W_r), and hidden state (W_n).
• update_gate_bias, reset_gate_bias, hidden_state_bias are layer bias arrays
• test_input1, test_input2, test_history are the inputs and initial history

The buffer is allocated as:

| reset | input | history | update | hidden_state |

In this way, the concatination is automatically done since (reset, input) and (input, history) are physically concatinated in memory.

The ordering of the weight matrix should be adjusted accordingly.

CMSIS DSP Software Library Functions Used:

• arm_fully_connected_mat_q7_vec_q15_opt()
• arm_nn_activations_direct_q15()
• arm_mult_q15()
• arm_offset_q15()
• arm_sub_q15()
• arm_copy_q15()

Refer arm_nnexamples_gru.cpp