// Sets the report string to a combined human and machine-readable report
// string of the error rates.
// Returns false if there is no data, leaving report unchanged, unless
// even_if_empty is true.
bool ErrorCounter::ReportString(bool even_if_empty, const Counts& counts,
STRING* report) {
// Compute the error rates.
double rates[CT_SIZE];
if (!ComputeRates(counts, rates) && !even_if_empty)
return false;
// Using %.4g%%, the length of the output string should exactly match the
// length of the format string, but in case of overflow, allow for +eddd
// on each number.
const int kMaxExtraLength = 5; // Length of +eddd.
// Keep this format string and the snprintf in sync with the CountTypes enum.
const char* format_str = "Unichar=%.4g%%[1], %.4g%%[2], %.4g%%[n], %.4g%%[T] "
"Mult=%.4g%%, Jn=%.4g%%, Brk=%.4g%%, Rej=%.4g%%, "
"FontAttr=%.4g%%, Multi=%.4g%%, "
"Answers=%.3g, Rank=%.3g, "
"OKjunk=%.4g%%, Badjunk=%.4g%%";
int max_str_len = strlen(format_str) + kMaxExtraLength * (CT_SIZE - 1) + 1;
char* formatted_str = new char[max_str_len];
snprintf(formatted_str, max_str_len, format_str,
rates[CT_UNICHAR_TOP1_ERR] * 100.0,
rates[CT_UNICHAR_TOP2_ERR] * 100.0,
rates[CT_UNICHAR_TOPN_ERR] * 100.0,
rates[CT_UNICHAR_TOPTOP_ERR] * 100.0,
rates[CT_OK_MULTI_UNICHAR] * 100.0,
rates[CT_OK_JOINED] * 100.0,
rates[CT_OK_BROKEN] * 100.0,
rates[CT_REJECT] * 100.0,
rates[CT_FONT_ATTR_ERR] * 100.0,
rates[CT_OK_MULTI_FONT] * 100.0,
rates[CT_NUM_RESULTS],
rates[CT_RANK],
100.0 * rates[CT_REJECTED_JUNK],
100.0 * rates[CT_ACCEPTED_JUNK]);
*report = formatted_str;
delete [] formatted_str;
// Now append each field of counts with a tab in front so the result can
// be loaded into a spreadsheet.
for (int ct = 0; ct < CT_SIZE; ++ct)
report->add_str_int("\t", counts.n[ct]);
return true;
}
// Creates a report of the error rate. The report_level controls the detail
// that is reported to stderr via tprintf:
// 0 -> no output.
// >=1 -> bottom-line error rate.
// >=3 -> font-level error rate.
// boosting_mode determines the return value. It selects which (un-weighted)
// error rate to return.
// The fontinfo_table from MasterTrainer provides the names of fonts.
// The it determines the current subset of the training samples.
// If not NULL, the top-choice unichar error rate is saved in unichar_error.
// If not NULL, the report string is saved in fonts_report.
// (Ignoring report_level).
double ErrorCounter::ReportErrors(int report_level, CountTypes boosting_mode,
const UnicityTable<FontInfo>& fontinfo_table,
const SampleIterator& it,
double* unichar_error,
STRING* fonts_report) {
// Compute totals over all the fonts and report individual font results
// when required.
Counts totals;
int fontsize = font_counts_.size();
for (int f = 0; f < fontsize; ++f) {
// Accumulate counts over fonts.
totals += font_counts_[f];
STRING font_report;
if (ReportString(font_counts_[f], &font_report)) {
if (fonts_report != NULL) {
*fonts_report += fontinfo_table.get(f).name;
*fonts_report += ": ";
*fonts_report += font_report;
*fonts_report += "\n";
}
if (report_level > 2) {
// Report individual font error rates.
tprintf("%s: %s\n", fontinfo_table.get(f).name, font_report.string());
}
}
}
if (report_level > 0) {
// Report the totals.
STRING total_report;
if (ReportString(totals, &total_report)) {
tprintf("TOTAL Scaled Err=%.4g%%, %s\n",
scaled_error_ * 100.0, total_report.string());
}
// Report the worst substitution error only for now.
if (totals.n[CT_UNICHAR_TOP1_ERR] > 0) {
const UNICHARSET& unicharset = it.shape_table()->unicharset();
int charsetsize = unicharset.size();
int shapesize = it.CompactCharsetSize();
int worst_uni_id = 0;
int worst_shape_id = 0;
int worst_err = 0;
for (int u = 0; u < charsetsize; ++u) {
for (int s = 0; s < shapesize; ++s) {
if (unichar_counts_(u, s) > worst_err) {
worst_err = unichar_counts_(u, s);
worst_uni_id = u;
worst_shape_id = s;
}
}
}
if (worst_err > 0) {
tprintf("Worst error = %d:%s -> %s with %d/%d=%.2f%% errors\n",
worst_uni_id, unicharset.id_to_unichar(worst_uni_id),
it.shape_table()->DebugStr(worst_shape_id).string(),
worst_err, totals.n[CT_UNICHAR_TOP1_ERR],
100.0 * worst_err / totals.n[CT_UNICHAR_TOP1_ERR]);
}
}
}
double rates[CT_SIZE];
if (!ComputeRates(totals, rates))
return 0.0;
// Set output values if asked for.
if (unichar_error != NULL)
*unichar_error = rates[CT_UNICHAR_TOP1_ERR];
return rates[boosting_mode];
}
// Creates a report of the error rate. The report_level controls the detail
// that is reported to stderr via tprintf:
// 0 -> no output.
// >=1 -> bottom-line error rate.
// >=3 -> font-level error rate.
// boosting_mode determines the return value. It selects which (un-weighted)
// error rate to return.
// The fontinfo_table from MasterTrainer provides the names of fonts.
// The it determines the current subset of the training samples.
// If not NULL, the top-choice unichar error rate is saved in unichar_error.
// If not NULL, the report string is saved in fonts_report.
// (Ignoring report_level).
double ErrorCounter::ReportErrors(int report_level, CountTypes boosting_mode,
const FontInfoTable& fontinfo_table,
const SampleIterator& it,
double* unichar_error,
STRING* fonts_report) {
// Compute totals over all the fonts and report individual font results
// when required.
Counts totals;
int fontsize = font_counts_.size();
for (int f = 0; f < fontsize; ++f) {
// Accumulate counts over fonts.
totals += font_counts_[f];
STRING font_report;
if (ReportString(false, font_counts_[f], &font_report)) {
if (fonts_report != NULL) {
*fonts_report += fontinfo_table.get(f).name;
*fonts_report += ": ";
*fonts_report += font_report;
*fonts_report += "\n";
}
if (report_level > 2) {
// Report individual font error rates.
tprintf("%s: %s\n", fontinfo_table.get(f).name, font_report.string());
}
}
}
// Report the totals.
STRING total_report;
bool any_results = ReportString(true, totals, &total_report);
if (fonts_report != NULL && fonts_report->length() == 0) {
// Make sure we return something even if there were no samples.
*fonts_report = "NoSamplesFound: ";
*fonts_report += total_report;
*fonts_report += "\n";
}
if (report_level > 0) {
// Report the totals.
STRING total_report;
if (any_results) {
tprintf("TOTAL Scaled Err=%.4g%%, %s\n",
scaled_error_ * 100.0, total_report.string());
}
// Report the worst substitution error only for now.
if (totals.n[CT_UNICHAR_TOP1_ERR] > 0) {
int charsetsize = unicharset_.size();
int worst_uni_id = 0;
int worst_result_id = 0;
int worst_err = 0;
for (int u = 0; u < charsetsize; ++u) {
for (int v = 0; v < charsetsize; ++v) {
if (unichar_counts_(u, v) > worst_err) {
worst_err = unichar_counts_(u, v);
worst_uni_id = u;
worst_result_id = v;
}
}
}
if (worst_err > 0) {
tprintf("Worst error = %d:%s -> %s with %d/%d=%.2f%% errors\n",
worst_uni_id, unicharset_.id_to_unichar(worst_uni_id),
unicharset_.id_to_unichar(worst_result_id),
worst_err, totals.n[CT_UNICHAR_TOP1_ERR],
100.0 * worst_err / totals.n[CT_UNICHAR_TOP1_ERR]);
}
}
tprintf("Multi-unichar shape use:\n");
for (int u = 0; u < multi_unichar_counts_.size(); ++u) {
if (multi_unichar_counts_[u] > 0) {
tprintf("%d multiple answers for unichar: %s\n",
multi_unichar_counts_[u],
unicharset_.id_to_unichar(u));
}
}
tprintf("OK Score histogram:\n");
ok_score_hist_.print();
tprintf("ERROR Score histogram:\n");
bad_score_hist_.print();
}
double rates[CT_SIZE];
if (!ComputeRates(totals, rates))
return 0.0;
// Set output values if asked for.
if (unichar_error != NULL)
*unichar_error = rates[CT_UNICHAR_TOP1_ERR];
return rates[boosting_mode];
}
