// Helper returns true if the null_char is the winner at t, and it beats the
// null_threshold, or the next choice is space, in which case we will use the
// null anyway.
static bool NullIsBest(const NetworkIO& output, float null_thr,
int null_char, int t) {
if (output.f(t)[null_char] >= null_thr) return true;
if (output.BestLabel(t, null_char, null_char, NULL) != UNICHAR_SPACE)
return false;
return output.f(t)[null_char] > output.f(t)[UNICHAR_SPACE];
}
// Prints debug output detailing activations and 2nd choice over a range
// of positions.
void LSTMRecognizer::DebugActivationRange(const NetworkIO& outputs,
const char* label, int best_choice,
int x_start, int x_end) {
tprintf("%s=%d On [%d, %d), scores=", label, best_choice, x_start, x_end);
double max_score = 0.0;
double mean_score = 0.0;
int width = x_end - x_start;
for (int x = x_start; x < x_end; ++x) {
const float* line = outputs.f(x);
double score = line[best_choice] * 100.0;
if (score > max_score) max_score = score;
mean_score += score / width;
int best_c = 0;
double best_score = 0.0;
for (int c = 0; c < outputs.NumFeatures(); ++c) {
if (c != best_choice && line[c] > best_score) {
best_c = c;
best_score = line[c];
}
}
tprintf(" %.3g(%s=%d=%.3g)", score, DecodeSingleLabel(best_c), best_c,
best_score * 100.0);
}
tprintf(", Mean=%g, max=%g\n", mean_score, max_score);
}
// Helper computes min and mean best results in the output.
void LSTMRecognizer::OutputStats(const NetworkIO& outputs, float* min_output,
float* mean_output, float* sd) {
const int kOutputScale = MAX_INT8;
STATS stats(0, kOutputScale + 1);
for (int t = 0; t < outputs.Width(); ++t) {
int best_label = outputs.BestLabel(t, NULL);
if (best_label != null_char_ || t == 0) {
float best_output = outputs.f(t)[best_label];
stats.add(static_cast<int>(kOutputScale * best_output), 1);
}
}
*min_output = static_cast<float>(stats.min_bucket()) / kOutputScale;
*mean_output = stats.mean() / kOutputScale;
*sd = stats.sd() / kOutputScale;
}
