#include "interpolate.h"
#include "level.h"
#include "tap.h"
+#include "filter.h"
#define BUFSIZE 4096
#define C64_FREQUENCY 985248
#define SYNC_PULSE_LENGTH 378.0
#define SYNC_TEST_TOLERANCE 1.10
+// SPSA options
#define NUM_FILTER_COEFF 32
+#define NUM_ITER 5000
+#define A NUM_ITER/10 // approx
+#define INITIAL_A 0.005 // A bit of trial and error...
+#define INITIAL_C 0.02 // This too.
+#define GAMMA 0.166
+#define ALPHA 1.0
static float hysteresis_limit = 3000.0 / 32768.0;
static bool do_calibrate = true;
static bool output_cycles_plot = false;
-static bool use_filter = false;
static bool do_crop = false;
static float crop_start = 0.0f, crop_end = HUGE_VAL;
+
+static bool use_fir_filter = false;
static float filter_coeff[NUM_FILTER_COEFF] = { 1.0f }; // The rest is filled with 0.
+static bool use_rc_filter = false;
+static float rc_filter_freq;
static bool output_filtered = false;
+
static bool quiet = false;
static bool do_auto_level = false;
static bool output_leveled = false;
+static std::vector<float> train_snap_points;
+static bool do_train = false;
+
+// The minimum estimated sound level (for do_auto_level) at any given point.
+// If you decrease this, you'll be able to amplify really silent signals
+// by more, but you'll also increase the level of silent (ie. noise-only) segments,
+// possibly caused misdetected pulses in these segments.
+static float min_level = 0.05f;
// between [x,x+1]
double find_zerocrossing(const std::vector<float> &pcm, int x)
static struct option long_options[] = {
{"auto-level", 0, 0, 'a' },
+ {"output-leveled", 0, 0, 'A' },
{"no-calibrate", 0, 0, 's' },
{"plot-cycles", 0, 0, 'p' },
{"hysteresis-limit", required_argument, 0, 'l' },
{"filter", required_argument, 0, 'f' },
+ {"rc-filter", required_argument, 0, 'r' },
{"output-filtered", 0, 0, 'F' },
{"crop", required_argument, 0, 'c' },
{"quiet", 0, 0, 'q' },
fprintf(stderr, "\n");
fprintf(stderr, " -a, --auto-level automatically adjust amplitude levels throughout the file\n");
fprintf(stderr, " -A, --output-leveled output leveled waveform to leveled.raw\n");
+ fprintf(stderr, " -m, --min-level minimum estimated sound level (0..32768) for --auto-level\n");
fprintf(stderr, " -s, --no-calibrate do not try to calibrate on sync pulse length\n");
fprintf(stderr, " -p, --plot-cycles output debugging info to cycles.plot\n");
fprintf(stderr, " -l, --hysteresis-limit VAL change amplitude threshold for ignoring pulses (0..32768)\n");
fprintf(stderr, " -f, --filter C1:C2:C3:... specify FIR filter (up to %d coefficients)\n", NUM_FILTER_COEFF);
+ fprintf(stderr, " -r, --rc-filter FREQ send signal through a highpass RC filter with given frequency (in Hertz)\n");
fprintf(stderr, " -F, --output-filtered output filtered waveform to filtered.raw\n");
fprintf(stderr, " -c, --crop START[:END] use only the given part of the file\n");
+ fprintf(stderr, " -t, --train LEN1:LEN2:... train a filter for detecting any of the given number of cycles\n");
+ fprintf(stderr, " (implies --no-calibrate and --quiet unless overridden)\n");
fprintf(stderr, " -q, --quiet suppress some informational messages\n");
fprintf(stderr, " -h, --help display this help, then exit\n");
exit(1);
{
for ( ;; ) {
int option_index = 0;
- int c = getopt_long(argc, argv, "aAspl:f:Fc:qh", long_options, &option_index);
+ int c = getopt_long(argc, argv, "aAm:spl:f:r:Fc:t:qh", long_options, &option_index);
if (c == -1)
break;
output_leveled = true;
break;
+ case 'm':
+ min_level = atof(optarg) / 32768.0;
+ break;
+
case 's':
do_calibrate = false;
break;
break;
case 'f': {
- const char *coeffstr = strtok(optarg, ":");
+ const char *coeffstr = strtok(optarg, ": ");
int coeff_index = 0;
while (coeff_index < NUM_FILTER_COEFF && coeffstr != NULL) {
filter_coeff[coeff_index++] = atof(coeffstr);
- coeffstr = strtok(NULL, ":");
+ coeffstr = strtok(NULL, ": ");
}
- use_filter = true;
+ use_fir_filter = true;
break;
}
+ case 'r':
+ use_rc_filter = true;
+ rc_filter_freq = atof(optarg);
+ break;
+
case 'F':
output_filtered = true;
break;
break;
}
+ case 't': {
+ const char *cyclestr = strtok(optarg, ":");
+ while (cyclestr != NULL) {
+ train_snap_points.push_back(atof(cyclestr));
+ cyclestr = strtok(NULL, ":");
+ }
+ do_train = true;
+
+ // Set reasonable defaults (can be overridden later on the command line).
+ do_calibrate = false;
+ quiet = true;
+ break;
+ }
+
case 'q':
quiet = true;
break;
}
// TODO: Support AVX here.
-std::vector<float> do_filter(const std::vector<float>& pcm, const float* filter)
+std::vector<float> do_fir_filter(const std::vector<float>& pcm, const float* filter)
{
std::vector<float> filtered_pcm;
filtered_pcm.reserve(pcm.size());
return filtered_pcm;
}
+std::vector<float> do_rc_filter(const std::vector<float>& pcm, float freq, int sample_rate)
+{
+ std::vector<float> filtered_pcm;
+ filtered_pcm.resize(pcm.size());
+ Filter filter = Filter::hpf(M_PI * freq / sample_rate);
+ for (unsigned i = 0; i < pcm.size(); ++i) {
+ filtered_pcm[i] = filter.update(pcm[i]);
+ }
+
+ if (output_filtered) {
+ FILE *fp = fopen("filtered.raw", "wb");
+ fwrite(filtered_pcm.data(), filtered_pcm.size() * sizeof(filtered_pcm[0]), 1, fp);
+ fclose(fp);
+ }
+
+ return filtered_pcm;
+}
+
std::vector<pulse> detect_pulses(const std::vector<float> &pcm, int sample_rate)
{
std::vector<pulse> pulses;
fclose(fp);
}
+std::pair<int, double> find_closest_point(double x, const std::vector<float> &points)
+{
+ int best_point = 0;
+ double best_dist = (x - points[0]) * (x - points[0]);
+ for (unsigned j = 1; j < train_snap_points.size(); ++j) {
+ double dist = (x - points[j]) * (x - points[j]);
+ if (dist < best_dist) {
+ best_point = j;
+ best_dist = dist;
+ }
+ }
+ return std::make_pair(best_point, best_dist);
+}
+
+float eval_badness(const std::vector<pulse>& pulses, double calibration_factor)
+{
+ double sum_badness = 0.0;
+ for (unsigned i = 0; i < pulses.size(); ++i) {
+ double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
+ if (cycles > 2000.0) cycles = 2000.0; // Don't make pauses arbitrarily bad.
+ std::pair<int, double> selected_point_and_sq_dist = find_closest_point(cycles, train_snap_points);
+ sum_badness += selected_point_and_sq_dist.second;
+ }
+ return sqrt(sum_badness / (pulses.size() - 1));
+}
+
+void find_kmeans(const std::vector<pulse> &pulses, double calibration_factor, const std::vector<float> &initial_centers)
+{
+ std::vector<float> last_centers = initial_centers;
+ std::vector<float> sums;
+ std::vector<float> num;
+ sums.resize(initial_centers.size());
+ num.resize(initial_centers.size());
+ for ( ;; ) {
+ for (unsigned i = 0; i < initial_centers.size(); ++i) {
+ sums[i] = 0.0f;
+ num[i] = 0;
+ }
+ for (unsigned i = 0; i < pulses.size(); ++i) {
+ double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
+ // Ignore heavy outliers, which are almost always long pauses.
+ if (cycles > 2000.0) {
+ continue;
+ }
+ std::pair<int, double> selected_point_and_sq_dist = find_closest_point(cycles, last_centers);
+ int p = selected_point_and_sq_dist.first;
+ sums[p] += cycles;
+ ++num[p];
+ }
+ bool any_moved = false;
+ for (unsigned i = 0; i < initial_centers.size(); ++i) {
+ if (num[i] == 0) {
+ printf("K-means broke down, can't output new reference training points\n");
+ return;
+ }
+ float new_center = sums[i] / num[i];
+ if (fabs(new_center - last_centers[i]) > 1e-3) {
+ any_moved = true;
+ }
+ last_centers[i] = new_center;
+ }
+ if (!any_moved) {
+ break;
+ }
+ }
+ printf("New reference training points:");
+ for (unsigned i = 0; i < last_centers.size(); ++i) {
+ printf(" %.3f", last_centers[i]);
+ }
+ printf("\n");
+}
+
+void spsa_train(const std::vector<float> &pcm, int sample_rate)
+{
+ float filter[NUM_FILTER_COEFF] = { 1.0f }; // The rest is filled with 0.
+
+ float start_c = INITIAL_C;
+ double best_badness = HUGE_VAL;
+
+ for (int n = 1; n < NUM_ITER; ++n) {
+ float a = INITIAL_A * pow(n + A, -ALPHA);
+ float c = start_c * pow(n, -GAMMA);
+
+ // find a random perturbation
+ float p[NUM_FILTER_COEFF];
+ float filter1[NUM_FILTER_COEFF], filter2[NUM_FILTER_COEFF];
+ for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
+ p[i] = (rand() % 2) ? 1.0 : -1.0;
+ filter1[i] = std::max(std::min(filter[i] - c * p[i], 1.0f), -1.0f);
+ filter2[i] = std::max(std::min(filter[i] + c * p[i], 1.0f), -1.0f);
+ }
+
+ std::vector<pulse> pulses1 = detect_pulses(do_fir_filter(pcm, filter1), sample_rate);
+ std::vector<pulse> pulses2 = detect_pulses(do_fir_filter(pcm, filter2), sample_rate);
+ float badness1 = eval_badness(pulses1, 1.0);
+ float badness2 = eval_badness(pulses2, 1.0);
+
+ // Find the gradient estimator
+ float g[NUM_FILTER_COEFF];
+ for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
+ g[i] = (badness2 - badness1) / (2.0 * c * p[i]);
+ filter[i] -= a * g[i];
+ filter[i] = std::max(std::min(filter[i], 1.0f), -1.0f);
+ }
+ if (badness2 < badness1) {
+ std::swap(badness1, badness2);
+ std::swap(filter1, filter2);
+ std::swap(pulses1, pulses2);
+ }
+ if (badness1 < best_badness) {
+ printf("\nNew best filter (badness=%f):", badness1);
+ for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
+ printf(" %.5f", filter1[i]);
+ }
+ best_badness = badness1;
+ printf("\n");
+
+ find_kmeans(pulses1, 1.0, train_snap_points);
+
+ if (output_cycles_plot) {
+ output_cycle_plot(pulses1, 1.0);
+ }
+ }
+ printf("%d ", n);
+ fflush(stdout);
+ }
+}
+
int main(int argc, char **argv)
{
parse_options(argc, argv);
pcm = crop(pcm, crop_start, crop_end, sample_rate);
}
- if (use_filter) {
- pcm = do_filter(pcm, filter_coeff);
+ if (use_fir_filter) {
+ pcm = do_fir_filter(pcm, filter_coeff);
+ }
+
+ if (use_rc_filter) {
+ pcm = do_rc_filter(pcm, rc_filter_freq, sample_rate);
}
if (do_auto_level) {
- pcm = level_samples(pcm, sample_rate);
+ pcm = level_samples(pcm, min_level, sample_rate);
if (output_leveled) {
FILE *fp = fopen("leveled.raw", "wb");
fwrite(pcm.data(), pcm.size() * sizeof(pcm[0]), 1, fp);
}
#endif
+ if (do_train) {
+ spsa_train(pcm, sample_rate);
+ exit(0);
+ }
+
std::vector<pulse> pulses = detect_pulses(pcm, sample_rate);
double calibration_factor = 1.0;