/* ----------------------------------------------------------------------
* Copyright (C) 2010-2018 Arm Limited. All rights reserved.
*
*
* Project: CMSIS NN Library
* Title: arm_nnexamples_gru.cpp
*
* Description: Gated Recurrent Unit Example
*
* Target Processor: Cortex-M4/Cortex-M7
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* - Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* - Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in
* the documentation and/or other materials provided with the
* distribution.
* - Neither the name of Arm LIMITED nor the names of its contributors
* may be used to endorse or promote products derived from this
* software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
* FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
* COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
* ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
* -------------------------------------------------------------------- */
/**
* @ingroup groupExamples
*/
/**
* @defgroup GRUExample Gated Recurrent Unit Example
*
* \par Description:
* \par
* Demonstrates a gated recurrent unit (GRU) example with the use of fully-connected,
* Tanh/Sigmoid activation functions.
*
* \par Model definition:
* \par
* GRU is a type of recurrent neural network (RNN). It contains two sigmoid gates and one hidden
* state.
* \par
* The computation can be summarized as:
* <pre>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] </pre>
* \image html GRU.gif "Gate Recurrent Unit Diagram"
*
* \par Variables Description:
* \par
* \li \c update_gate_weights, \c reset_gate_weights, \c hidden_state_weights are weights corresponding to update gate (W_z), reset gate (W_r), and hidden state (W_n).
* \li \c update_gate_bias, \c reset_gate_bias, \c hidden_state_bias are layer bias arrays
* \li \c test_input1, \c test_input2, \c test_history are the inputs and initial history
*
* \par
* The buffer is allocated as:
* \par
* | reset | input | history | update | hidden_state |
* \par
* In this way, the concatination is automatically done since (reset, input) and (input, history)
* are physically concatinated in memory.
* \par
* The ordering of the weight matrix should be adjusted accordingly.
*
*
*
* \par CMSIS DSP Software Library Functions Used:
* \par
* - 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()
*
* <b> Refer </b>
* \link arm_nnexamples_gru.cpp \endlink
*
*/
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include "arm_nnexamples_gru_test_data.h"
#include "arm_math.h"
#include "arm_nnfunctions.h"
#ifdef _RTE_
#include "RTE_Components.h"
#ifdef RTE_Compiler_EventRecorder
#include "EventRecorder.h"
#endif
#endif
#define DIM_HISTORY 32
#define DIM_INPUT 32
#define DIM_VEC 64
#define USE_X4
#ifndef USE_X4
static q7_t update_gate_weights[DIM_VEC * DIM_HISTORY] = UPDATE_GATE_WEIGHT_X2;
static q7_t reset_gate_weights[DIM_VEC * DIM_HISTORY] = RESET_GATE_WEIGHT_X2;
static q7_t hidden_state_weights[DIM_VEC * DIM_HISTORY] = HIDDEN_STATE_WEIGHT_X2;
#else
static q7_t update_gate_weights[DIM_VEC * DIM_HISTORY] = UPDATE_GATE_WEIGHT_X4;
static q7_t reset_gate_weights[DIM_VEC * DIM_HISTORY] = RESET_GATE_WEIGHT_X4;
static q7_t hidden_state_weights[DIM_VEC * DIM_HISTORY] = HIDDEN_STATE_WEIGHT_X4;
#endif
static q7_t update_gate_bias[DIM_HISTORY] = UPDATE_GATE_BIAS;
static q7_t reset_gate_bias[DIM_HISTORY] = RESET_GATE_BIAS;
static q7_t hidden_state_bias[DIM_HISTORY] = HIDDEN_STATE_BIAS;
static q15_t test_input1[DIM_INPUT] = INPUT_DATA1;
static q15_t test_input2[DIM_INPUT] = INPUT_DATA2;
static q15_t test_history[DIM_HISTORY] = HISTORY_DATA;
q15_t scratch_buffer[DIM_HISTORY * 4 + DIM_INPUT];
void gru_example(q15_t * scratch_input, uint16_t input_size, uint16_t history_size,
q7_t * weights_update, q7_t * weights_reset, q7_t * weights_hidden_state,
q7_t * bias_update, q7_t * bias_reset, q7_t * bias_hidden_state)
{
q15_t *reset = scratch_input;
q15_t *input = scratch_input + history_size;
q15_t *history = scratch_input + history_size + input_size;
q15_t *update = scratch_input + 2 * history_size + input_size;
q15_t *hidden_state = scratch_input + 3 * history_size + input_size;
// reset gate calculation
// the range of the output can be adjusted with bias_shift and output_shift
#ifndef USE_X4
arm_fully_connected_mat_q7_vec_q15(input, weights_reset, input_size + history_size, history_size, 0, 15, bias_reset,
reset, NULL);
#else
arm_fully_connected_mat_q7_vec_q15_opt(input, weights_reset, input_size + history_size, history_size, 0, 15,
bias_reset, reset, NULL);
#endif
// sigmoid function, the size of the integer bit-width should be consistent with out_shift
arm_nn_activations_direct_q15(reset, history_size, 0, ARM_SIGMOID);
arm_mult_q15(history, reset, reset, history_size);
// update gate calculation
// the range of the output can be adjusted with bias_shift and output_shift
#ifndef USE_X4
arm_fully_connected_mat_q7_vec_q15(input, weights_update, input_size + history_size, history_size, 0, 15,
bias_update, update, NULL);
#else
arm_fully_connected_mat_q7_vec_q15_opt(input, weights_update, input_size + history_size, history_size, 0, 15,
bias_update, update, NULL);
#endif
// sigmoid function, the size of the integer bit-width should be consistent with out_shift
arm_nn_activations_direct_q15(update, history_size, 0, ARM_SIGMOID);
// hidden state calculation
#ifndef USE_X4
arm_fully_connected_mat_q7_vec_q15(reset, weights_hidden_state, input_size + history_size, history_size, 0, 15,
bias_hidden_state, hidden_state, NULL);
#else
arm_fully_connected_mat_q7_vec_q15_opt(reset, weights_hidden_state, input_size + history_size, history_size, 0, 15,
bias_hidden_state, hidden_state, NULL);
#endif
// tanh function, the size of the integer bit-width should be consistent with out_shift
arm_nn_activations_direct_q15(hidden_state, history_size, 0, ARM_TANH);
arm_mult_q15(update, hidden_state, hidden_state, history_size);
// we calculate z - 1 here
// so final addition becomes substraction
arm_offset_q15(update, 0x8000, update, history_size);
// multiply history
arm_mult_q15(history, update, update, history_size);
// calculate history_out
arm_sub_q15(hidden_state, update, history, history_size);
return;
}
int main()
{
#ifdef RTE_Compiler_EventRecorder
EventRecorderInitialize (EventRecordAll, 1); // initialize and start Event Recorder
#endif
printf("Start GRU execution\n");
int input_size = DIM_INPUT;
int history_size = DIM_HISTORY;
// copy over the input data
arm_copy_q15(test_input1, scratch_buffer + history_size, input_size);
arm_copy_q15(test_history, scratch_buffer + history_size + input_size, history_size);
gru_example(scratch_buffer, input_size, history_size,
update_gate_weights, reset_gate_weights, hidden_state_weights,
update_gate_bias, reset_gate_bias, hidden_state_bias);
printf("Complete first iteration on GRU\n");
arm_copy_q15(test_input2, scratch_buffer + history_size, input_size);
gru_example(scratch_buffer, input_size, history_size,
update_gate_weights, reset_gate_weights, hidden_state_weights,
update_gate_bias, reset_gate_bias, hidden_state_bias);
printf("Complete second iteration on GRU\n");
return 0;
}
