openmv/src/omv/ports/stm32/modules/py_micro_speech.c

292 lines
11 KiB
C

/*
* This file is part of the OpenMV project.
*
* Copyright (c) 2013-2021 Ibrahim Abdelkader <iabdalkader@openmv.io>
* Copyright (c) 2013-2021 Kwabena W. Agyeman <kwagyeman@openmv.io>
*
* This work is licensed under the MIT license, see the file LICENSE for details.
*
* Micro Speech Python module.
*/
#include <stdio.h>
#include "py/obj.h"
#include "py/objarray.h"
#include "py/runtime.h"
#include "py/mphal.h"
#include "systick.h"
#include "py/binary.h"
#include "py_assert.h"
#include "py_helper.h"
#include "fb_alloc.h"
#include "omv_boardconfig.h"
#include "libtf.h"
#include "py_tf.h"
#include "common.h"
#if MICROPY_PY_MICRO_SPEECH
#define kMaxAudioSampleSize (512)
#define kAudioSampleFrequency (16000)
// The following values are derived from values used during model training.
// If you change the way you preprocess the input, update all these constants.
#define kFeatureSliceSize (40)
#define kFeatureSliceCount (49)
#define kFeatureElementCount (kFeatureSliceSize * kFeatureSliceCount)
#define kFeatureSliceStrideMs (20)
#define kFeatureSliceDurationMs (30)
#define kCategoryCount (4)
#define kAverageWindowSamples (1020 / kFeatureSliceDurationMs)
#define RAISE_OS_EXCEPTION(msg) mp_raise_msg(&mp_type_OSError, MP_ERROR_TEXT(msg))
typedef struct _py_micro_speech_obj {
mp_obj_base_t base;
uint32_t n_slices;
bool new_slices;
int8_t spectrogram[kFeatureElementCount];
} py_micro_speech_obj_t;
static const mp_obj_type_t py_micro_speech_type;
static void py_micro_speech_print(const mp_print_t *print, mp_obj_t self_in, mp_print_kind_t kind)
{
py_micro_speech_obj_t *microspeech = MP_OBJ_TO_PTR(self_in);
printf("MicroSpeech obj n_slices: %lu new_slices :%d!\n",
microspeech->n_slices, microspeech->new_slices);
}
mp_obj_t py_micro_speech_make_new(const mp_obj_type_t *type, size_t n_args, size_t n_kw, const mp_obj_t *args)
{
// check arguments
//mp_arg_check_num(n_args, n_kw, 1, MP_OBJ_FUN_ARGS_MAX, true);
py_micro_speech_obj_t *o = m_new_obj(py_micro_speech_obj_t);
o->base.type = &py_micro_speech_type;
o->n_slices = 0;
o->new_slices = false;
memset(o->spectrogram, 0, kFeatureElementCount);
if (libtf_initialize_micro_features() != 0) {
RAISE_OS_EXCEPTION("Failed to initialize micro features!");
}
return MP_OBJ_FROM_PTR(o);
}
mp_obj_t py_micro_speech_audio_callback(mp_obj_t self_in, mp_obj_t buf_in)
{
py_micro_speech_obj_t *microspeech = MP_OBJ_TO_PTR(self_in);
mp_buffer_info_t pcmbuf;
mp_get_buffer_raise(buf_in, &pcmbuf, MP_BUFFER_READ);
if ((pcmbuf.len / 2) != kMaxAudioSampleSize) {
RAISE_OS_EXCEPTION("Audio data size too small!");
}
uint32_t slice_index = 0;
if (microspeech->n_slices < kFeatureSliceCount) {
slice_index = microspeech->n_slices++;
} else {
// Spectrogram is full, move old data up in the spectrogram
// and put the new slice at the end of the spectrogram.
// +-----------+ +-----------+
// | data@20ms | --> | data@40ms |
// +-----------+ -- +-----------+
// | data@40ms | -- --> | data@60ms |
// +-----------+ -- -- +-----------+
// | data@60ms | -- -- | data@80ms |
// +-----------+ -- +-----------+
// | data@80ms | -- |<new slice>|
// +-----------+ +-----------+
slice_index = (kFeatureSliceCount - 1);
memmove(microspeech->spectrogram,
microspeech->spectrogram + kFeatureSliceSize,
kFeatureElementCount - kFeatureSliceSize);
microspeech->new_slices = true;
}
size_t num_samples_read;
int8_t *new_slice = microspeech->spectrogram + (slice_index * kFeatureSliceSize);
if (libtf_generate_micro_features((int16_t*) pcmbuf.buf,
kMaxAudioSampleSize, kFeatureSliceSize, new_slice, &num_samples_read)) {
RAISE_OS_EXCEPTION("Feature generation failed!");
}
return mp_const_none;
}
STATIC MP_DEFINE_CONST_FUN_OBJ_2(py_micro_speech_audio_callback_obj, py_micro_speech_audio_callback);
STATIC void py_tf_input_callback(void *callback_data, void *model_input, libtf_parameters_t *params)
{
// Copy feature buffer to input tensor
for (int i = 0; i < kFeatureElementCount; i++) {
((int8_t *) model_input)[i] = ((int8_t *) callback_data)[i];
}
}
STATIC void py_tf_output_callback(void *callback_data, void *model_output, libtf_parameters_t *params)
{
uint8_t *scores = (uint8_t *) callback_data;
if (params->output_height != 1) {
mp_raise_msg(&mp_type_ValueError, MP_ERROR_TEXT("Expected model output height to be 1!"));
}
if (params->output_width != 1) {
mp_raise_msg(&mp_type_ValueError, MP_ERROR_TEXT("Expected model output width to be 1!"));
}
if (params->output_channels != 4) {
mp_raise_msg(&mp_type_ValueError, MP_ERROR_TEXT("Expected model output channels to be 4!"));
}
for (int i = 0, ii = params->output_channels; i < ii; i++) {
scores[i] = ((uint8_t *) model_output)[i] - params->output_zero_point;
debug_printf("%.2f ", (double) ((((uint8_t *) model_output)[i] - params->output_zero_point) * params->output_scale));
}
}
STATIC mp_obj_t py_micro_speech_listen(uint n_args, const mp_obj_t *args, mp_map_t *kw_args)
{
py_micro_speech_obj_t *microspeech = args[0];
py_tf_model_obj_t *arg_model = args[1];
float threshold = py_helper_keyword_float(n_args, args, 2, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_threshold), 0.9f);
uint32_t timeout = py_helper_keyword_int(n_args, args, 3, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_timeout), 1000);
size_t labels_filter_len = 0;
mp_obj_t *labels_filter = py_helper_keyword_iterable(n_args, args,
4, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_filter), &labels_filter_len);
fb_alloc_mark();
py_tf_alloc_putchar_buffer();
uint32_t tensor_arena_size;
uint8_t *tensor_arena = fb_alloc_all(&tensor_arena_size, FB_ALLOC_PREFER_SIZE);
libtf_parameters_t params;
if (libtf_get_parameters(arg_model->model_data, tensor_arena, tensor_arena_size, &params) != 0) {
mp_raise_msg(&mp_type_OSError, (mp_rom_error_text_t) py_tf_putchar_buffer);
}
fb_free(); // free fb_alloc_all()
tensor_arena = fb_alloc(params.tensor_arena_size, FB_ALLOC_PREFER_SPEED | FB_ALLOC_CACHE_ALIGN);
int8_t spectrogram[kFeatureElementCount];
uint32_t return_label = 0;
uint32_t results_count = 0;
uint8_t previous_scores[kAverageWindowSamples][kCategoryCount];
uint32_t average_scores[kCategoryCount];
memset(previous_scores, 0, kAverageWindowSamples * kCategoryCount);
memset(average_scores, 0, kCategoryCount * sizeof(*average_scores));
uint32_t start = HAL_GetTick();
while (timeout == 0 || (HAL_GetTick() - start) < timeout) {
__WFI();
if (microspeech->new_slices == false) {
continue;
}
// Copy spectrogram atomically
__disable_irq();
microspeech->new_slices = false;
memcpy(spectrogram, microspeech->spectrogram, kFeatureElementCount);
__enable_irq();
// Run model on updated spectrogram
if (libtf_invoke(arg_model->model_data,
tensor_arena,
&params,
py_tf_input_callback,
spectrogram,
py_tf_output_callback,
previous_scores[results_count]) != 0) {
mp_raise_msg(&mp_type_OSError, (mp_rom_error_text_t) py_tf_putchar_buffer);
}
// If we have enough samples calculate average scores.
if ((HAL_GetTick() - start) > (kAverageWindowSamples * kFeatureSliceDurationMs)) {
uint32_t highest_index = 0, highest_score = 0;
// Re/Calculate the average score for all labels in the window.
for (int i=0; i<kAverageWindowSamples; i++) {
for (int c=0; c<kCategoryCount; c++) {
if (i == 0) {
average_scores[c] = previous_scores[i][c];
} else {
average_scores[c] += previous_scores[i][c];
}
}
}
// Find the label index with the highest average score.
for (int i=0; i<kCategoryCount; i++) {
if (average_scores[i] > highest_score) {
highest_index = i;
highest_score = average_scores[i];
}
}
// If the highest average score is higher than the threshold return a command.
if (average_scores[highest_index] / (kAverageWindowSamples * 255.0f) > threshold) {
bool command_filtered = (labels_filter != NULL);
// If a list of labels is provided to filter commands, check if the
// detected command is in that list, otherwise continue the detection.
if (labels_filter != NULL) {
for (int i=0; i<labels_filter_len; i++) {
if (highest_index == mp_obj_get_int(labels_filter[i])) {
command_filtered = false;
break;
}
}
}
if (command_filtered == false) {
return_label = highest_index;
// Clear spectrogram
__disable_irq();
microspeech->n_slices = 0;
microspeech->new_slices = false;
__enable_irq();
break;
}
}
}
results_count = (results_count + 1) % kAverageWindowSamples;
}
fb_alloc_free_till_mark();
return mp_obj_new_int(return_label);
}
STATIC MP_DEFINE_CONST_FUN_OBJ_KW(py_micro_speech_listen_obj, 2, py_micro_speech_listen);
STATIC const mp_rom_map_elem_t py_micro_speech_locals_dict_table[] = {
// instance methods
{ MP_ROM_QSTR(MP_QSTR_audio_callback), MP_ROM_PTR(&py_micro_speech_audio_callback_obj) },
{ MP_ROM_QSTR(MP_QSTR_listen), MP_ROM_PTR(&py_micro_speech_listen_obj) },
// class constants
};
STATIC MP_DEFINE_CONST_DICT(py_micro_speech_locals_dict, py_micro_speech_locals_dict_table);
static const mp_obj_type_t py_micro_speech_type = {
{ &mp_type_type },
.name = MP_QSTR_MicroSpeech,
.print = py_micro_speech_print,
.make_new = py_micro_speech_make_new,
.locals_dict = (mp_obj_dict_t*)&py_micro_speech_locals_dict,
};
STATIC const mp_rom_map_elem_t module_globals_table[] = {
{ MP_ROM_QSTR(MP_QSTR___name__), MP_ROM_QSTR(MP_QSTR_micro_speech) },
{ MP_ROM_QSTR(MP_QSTR_MicroSpeech), MP_ROM_PTR(&py_micro_speech_type) },
};
STATIC MP_DEFINE_CONST_DICT(module_globals, module_globals_table);
const mp_obj_module_t micro_speech_module = {
.base = { &mp_type_module },
.globals = (mp_obj_dict_t*)&module_globals,
};
MP_REGISTER_MODULE(MP_QSTR_micro_speech, micro_speech_module);
#endif // MICROPY_PY_MICRO_SPEECH