diff options
Diffstat (limited to 'native/jni/src/proximity_info_state_utils.cpp')
| -rw-r--r-- | native/jni/src/proximity_info_state_utils.cpp | 368 |
1 files changed, 368 insertions, 0 deletions
diff --git a/native/jni/src/proximity_info_state_utils.cpp b/native/jni/src/proximity_info_state_utils.cpp index 65a258309..e5567f7a7 100644 --- a/native/jni/src/proximity_info_state_utils.cpp +++ b/native/jni/src/proximity_info_state_utils.cpp @@ -14,6 +14,7 @@ * limitations under the License. */ +#include <sstream> // for debug prints #include <vector> #include "defines.h" @@ -481,4 +482,371 @@ namespace latinime { // TODO: Detect double letter more smartly return 0.01f + static_cast<float>(beelineDistance) / static_cast<float>(time) / averageSpeed; } + +/* static */ float ProximityInfoStateUtils::getPointAngle( + const std::vector<int> *const sampledInputXs, + const std::vector<int> *const sampledInputYs, const int index) { + if (!sampledInputXs || !sampledInputYs) { + return 0.0f; + } + const int sampledInputSize = sampledInputXs->size(); + if (index <= 0 || index >= sampledInputSize - 1) { + return 0.0f; + } + const float previousDirection = getDirection(sampledInputXs, sampledInputYs, index - 1, index); + const float nextDirection = getDirection(sampledInputXs, sampledInputYs, index, index + 1); + const float directionDiff = getAngleDiff(previousDirection, nextDirection); + return directionDiff; +} + +/* static */ float ProximityInfoStateUtils::getPointsAngle( + const std::vector<int> *const sampledInputXs, + const std::vector<int> *const sampledInputYs, + const int index0, const int index1, const int index2) { + if (!sampledInputXs || !sampledInputYs) { + return 0.0f; + } + const int sampledInputSize = sampledInputXs->size(); + if (index0 < 0 || index0 > sampledInputSize - 1) { + return 0.0f; + } + if (index1 < 0 || index1 > sampledInputSize - 1) { + return 0.0f; + } + if (index2 < 0 || index2 > sampledInputSize - 1) { + return 0.0f; + } + const float previousDirection = getDirection(sampledInputXs, sampledInputYs, index0, index1); + const float nextDirection = getDirection(sampledInputXs, sampledInputYs, index1, index2); + return getAngleDiff(previousDirection, nextDirection); +} + +// TODO: Remove the "scale" parameter +// This function basically converts from a length to an edit distance. Accordingly, it's obviously +// wrong to compare with mMaxPointToKeyLength. +/* static */ float ProximityInfoStateUtils::getPointToKeyByIdLength(const float maxPointToKeyLength, + const std::vector<float> *const distanceCache_G, const int keyCount, + const int inputIndex, const int keyId, const float scale) { + if (keyId != NOT_AN_INDEX) { + const int index = inputIndex * keyCount + keyId; + return min((*distanceCache_G)[index] * scale, maxPointToKeyLength); + } + // If the char is not a key on the keyboard then return the max length. + return static_cast<float>(MAX_POINT_TO_KEY_LENGTH); +} + +/* static */ float ProximityInfoStateUtils::getPointToKeyByIdLength(const float maxPointToKeyLength, + const std::vector<float> *const distanceCache_G, const int keyCount, + const int inputIndex, const int keyId) { + return getPointToKeyByIdLength(maxPointToKeyLength, distanceCache_G, keyCount, inputIndex, + keyId, 1.0f); +} + +// Updates probabilities of aligning to some keys and skipping. +// Word suggestion should be based on this probabilities. +/* static */ void ProximityInfoStateUtils::updateAlignPointProbabilities( + const float maxPointToKeyLength, const int mostCommonKeyWidth, const int keyCount, + const int start, const int sampledInputSize, const std::vector<int> *const sampledInputXs, + const std::vector<int> *const sampledInputYs, + const std::vector<float> *const sampledSpeedRates, + const std::vector<int> *const sampledLengthCache, + const std::vector<float> *const distanceCache_G, + std::vector<NearKeycodesSet> *nearKeysVector, + std::vector<hash_map_compat<int, float> > *charProbabilities) { + static const float MIN_PROBABILITY = 0.000001f; + static const float MAX_SKIP_PROBABILITY = 0.95f; + static const float SKIP_FIRST_POINT_PROBABILITY = 0.01f; + static const float SKIP_LAST_POINT_PROBABILITY = 0.1f; + static const float MIN_SPEED_RATE_FOR_SKIP_PROBABILITY = 0.15f; + static const float SPEED_WEIGHT_FOR_SKIP_PROBABILITY = 0.9f; + static const float SLOW_STRAIGHT_WEIGHT_FOR_SKIP_PROBABILITY = 0.6f; + static const float NEAREST_DISTANCE_WEIGHT = 0.5f; + static const float NEAREST_DISTANCE_BIAS = 0.5f; + static const float NEAREST_DISTANCE_WEIGHT_FOR_LAST = 0.6f; + static const float NEAREST_DISTANCE_BIAS_FOR_LAST = 0.4f; + + static const float ANGLE_WEIGHT = 0.90f; + static const float DEEP_CORNER_ANGLE_THRESHOLD = M_PI_F * 60.0f / 180.0f; + static const float SKIP_DEEP_CORNER_PROBABILITY = 0.1f; + static const float CORNER_ANGLE_THRESHOLD = M_PI_F * 30.0f / 180.0f; + static const float STRAIGHT_ANGLE_THRESHOLD = M_PI_F * 15.0f / 180.0f; + static const float SKIP_CORNER_PROBABILITY = 0.4f; + static const float SPEED_MARGIN = 0.1f; + static const float CENTER_VALUE_OF_NORMALIZED_DISTRIBUTION = 0.0f; + + charProbabilities->resize(sampledInputSize); + // Calculates probabilities of using a point as a correlated point with the character + // for each point. + for (int i = start; i < sampledInputSize; ++i) { + (*charProbabilities)[i].clear(); + // First, calculates skip probability. Starts form MIN_SKIP_PROBABILITY. + // Note that all values that are multiplied to this probability should be in [0.0, 1.0]; + float skipProbability = MAX_SKIP_PROBABILITY; + + const float currentAngle = getPointAngle(sampledInputXs, sampledInputYs, i); + const float speedRate = (*sampledSpeedRates)[i]; + + float nearestKeyDistance = static_cast<float>(MAX_POINT_TO_KEY_LENGTH); + for (int j = 0; j < keyCount; ++j) { + if ((*nearKeysVector)[i].test(j)) { + const float distance = getPointToKeyByIdLength( + maxPointToKeyLength, distanceCache_G, keyCount, i, j); + if (distance < nearestKeyDistance) { + nearestKeyDistance = distance; + } + } + } + + if (i == 0) { + skipProbability *= min(1.0f, nearestKeyDistance * NEAREST_DISTANCE_WEIGHT + + NEAREST_DISTANCE_BIAS); + // Promote the first point + skipProbability *= SKIP_FIRST_POINT_PROBABILITY; + } else if (i == sampledInputSize - 1) { + skipProbability *= min(1.0f, nearestKeyDistance * NEAREST_DISTANCE_WEIGHT_FOR_LAST + + NEAREST_DISTANCE_BIAS_FOR_LAST); + // Promote the last point + skipProbability *= SKIP_LAST_POINT_PROBABILITY; + } else { + // If the current speed is relatively slower than adjacent keys, we promote this point. + if ((*sampledSpeedRates)[i - 1] - SPEED_MARGIN > speedRate + && speedRate < (*sampledSpeedRates)[i + 1] - SPEED_MARGIN) { + if (currentAngle < CORNER_ANGLE_THRESHOLD) { + skipProbability *= min(1.0f, speedRate + * SLOW_STRAIGHT_WEIGHT_FOR_SKIP_PROBABILITY); + } else { + // If the angle is small enough, we promote this point more. (e.g. pit vs put) + skipProbability *= min(1.0f, speedRate * SPEED_WEIGHT_FOR_SKIP_PROBABILITY + + MIN_SPEED_RATE_FOR_SKIP_PROBABILITY); + } + } + + skipProbability *= min(1.0f, speedRate * nearestKeyDistance * + NEAREST_DISTANCE_WEIGHT + NEAREST_DISTANCE_BIAS); + + // Adjusts skip probability by a rate depending on angle. + // ANGLE_RATE of skipProbability is adjusted by current angle. + skipProbability *= (M_PI_F - currentAngle) / M_PI_F * ANGLE_WEIGHT + + (1.0f - ANGLE_WEIGHT); + if (currentAngle > DEEP_CORNER_ANGLE_THRESHOLD) { + skipProbability *= SKIP_DEEP_CORNER_PROBABILITY; + } + // We assume the angle of this point is the angle for point[i], point[i - 2] + // and point[i - 3]. The reason why we don't use the angle for point[i], point[i - 1] + // and point[i - 2] is this angle can be more affected by the noise. + const float prevAngle = getPointsAngle(sampledInputXs, sampledInputYs, i, i - 2, i - 3); + if (i >= 3 && prevAngle < STRAIGHT_ANGLE_THRESHOLD + && currentAngle > CORNER_ANGLE_THRESHOLD) { + skipProbability *= SKIP_CORNER_PROBABILITY; + } + } + + // probabilities must be in [0.0, MAX_SKIP_PROBABILITY]; + ASSERT(skipProbability >= 0.0f); + ASSERT(skipProbability <= MAX_SKIP_PROBABILITY); + (*charProbabilities)[i][NOT_AN_INDEX] = skipProbability; + + // Second, calculates key probabilities by dividing the rest probability + // (1.0f - skipProbability). + const float inputCharProbability = 1.0f - skipProbability; + + // TODO: The variance is critical for accuracy; thus, adjusting these parameter by machine + // learning or something would be efficient. + static const float SPEEDxANGLE_WEIGHT_FOR_STANDARD_DIVIATION = 0.3f; + static const float MAX_SPEEDxANGLE_RATE_FOR_STANDERD_DIVIATION = 0.25f; + static const float SPEEDxNEAREST_WEIGHT_FOR_STANDARD_DIVIATION = 0.5f; + static const float MAX_SPEEDxNEAREST_RATE_FOR_STANDERD_DIVIATION = 0.15f; + static const float MIN_STANDERD_DIVIATION = 0.37f; + + const float speedxAngleRate = min(speedRate * currentAngle / M_PI_F + * SPEEDxANGLE_WEIGHT_FOR_STANDARD_DIVIATION, + MAX_SPEEDxANGLE_RATE_FOR_STANDERD_DIVIATION); + const float speedxNearestKeyDistanceRate = min(speedRate * nearestKeyDistance + * SPEEDxNEAREST_WEIGHT_FOR_STANDARD_DIVIATION, + MAX_SPEEDxNEAREST_RATE_FOR_STANDERD_DIVIATION); + const float sigma = speedxAngleRate + speedxNearestKeyDistanceRate + MIN_STANDERD_DIVIATION; + + ProximityInfoUtils::NormalDistribution + distribution(CENTER_VALUE_OF_NORMALIZED_DISTRIBUTION, sigma); + static const float PREV_DISTANCE_WEIGHT = 0.5f; + static const float NEXT_DISTANCE_WEIGHT = 0.6f; + // Summing up probability densities of all near keys. + float sumOfProbabilityDensities = 0.0f; + for (int j = 0; j < keyCount; ++j) { + if ((*nearKeysVector)[i].test(j)) { + float distance = sqrtf(getPointToKeyByIdLength( + maxPointToKeyLength, distanceCache_G, keyCount, i, j)); + if (i == 0 && i != sampledInputSize - 1) { + // For the first point, weighted average of distances from first point and the + // next point to the key is used as a point to key distance. + const float nextDistance = sqrtf(getPointToKeyByIdLength( + maxPointToKeyLength, distanceCache_G, keyCount, i + 1, j)); + if (nextDistance < distance) { + // The distance of the first point tends to bigger than continuing + // points because the first touch by the user can be sloppy. + // So we promote the first point if the distance of that point is larger + // than the distance of the next point. + distance = (distance + nextDistance * NEXT_DISTANCE_WEIGHT) + / (1.0f + NEXT_DISTANCE_WEIGHT); + } + } else if (i != 0 && i == sampledInputSize - 1) { + // For the first point, weighted average of distances from last point and + // the previous point to the key is used as a point to key distance. + const float previousDistance = sqrtf(getPointToKeyByIdLength( + maxPointToKeyLength, distanceCache_G, keyCount, i - 1, j)); + if (previousDistance < distance) { + // The distance of the last point tends to bigger than continuing points + // because the last touch by the user can be sloppy. So we promote the + // last point if the distance of that point is larger than the distance of + // the previous point. + distance = (distance + previousDistance * PREV_DISTANCE_WEIGHT) + / (1.0f + PREV_DISTANCE_WEIGHT); + } + } + // TODO: Promote the first point when the extended line from the next input is near + // from a key. Also, promote the last point as well. + sumOfProbabilityDensities += distribution.getProbabilityDensity(distance); + } + } + + // Split the probability of an input point to keys that are close to the input point. + for (int j = 0; j < keyCount; ++j) { + if ((*nearKeysVector)[i].test(j)) { + float distance = sqrtf(getPointToKeyByIdLength( + maxPointToKeyLength, distanceCache_G, keyCount, i, j)); + if (i == 0 && i != sampledInputSize - 1) { + // For the first point, weighted average of distances from the first point and + // the next point to the key is used as a point to key distance. + const float prevDistance = sqrtf(getPointToKeyByIdLength( + maxPointToKeyLength, distanceCache_G, keyCount, i + 1, j)); + if (prevDistance < distance) { + distance = (distance + prevDistance * NEXT_DISTANCE_WEIGHT) + / (1.0f + NEXT_DISTANCE_WEIGHT); + } + } else if (i != 0 && i == sampledInputSize - 1) { + // For the first point, weighted average of distances from last point and + // the previous point to the key is used as a point to key distance. + const float prevDistance = sqrtf(getPointToKeyByIdLength( + maxPointToKeyLength, distanceCache_G, keyCount, i - 1, j)); + if (prevDistance < distance) { + distance = (distance + prevDistance * PREV_DISTANCE_WEIGHT) + / (1.0f + PREV_DISTANCE_WEIGHT); + } + } + const float probabilityDensity = distribution.getProbabilityDensity(distance); + const float probability = inputCharProbability * probabilityDensity + / sumOfProbabilityDensities; + (*charProbabilities)[i][j] = probability; + } + } + } + + + if (DEBUG_POINTS_PROBABILITY) { + for (int i = 0; i < sampledInputSize; ++i) { + std::stringstream sstream; + sstream << i << ", "; + sstream << "(" << (*sampledInputXs)[i] << ", " << (*sampledInputYs)[i] << "), "; + sstream << "Speed: "<< (*sampledSpeedRates)[i] << ", "; + sstream << "Angle: "<< getPointAngle(sampledInputXs, sampledInputYs, i) << ", \n"; + + for (hash_map_compat<int, float>::iterator it = (*charProbabilities)[i].begin(); + it != (*charProbabilities)[i].end(); ++it) { + if (it->first == NOT_AN_INDEX) { + sstream << it->first + << "(skip):" + << it->second + << "\n"; + } else { + sstream << it->first + << "(" + //<< static_cast<char>(mProximityInfo->getCodePointOf(it->first)) + << "):" + << it->second + << "\n"; + } + } + AKLOGI("%s", sstream.str().c_str()); + } + } + + // Decrease key probabilities of points which don't have the highest probability of that key + // among nearby points. Probabilities of the first point and the last point are not suppressed. + for (int i = max(start, 1); i < sampledInputSize; ++i) { + for (int j = i + 1; j < sampledInputSize; ++j) { + if (!suppressCharProbabilities( + mostCommonKeyWidth, sampledInputSize, sampledLengthCache, i, j, + charProbabilities)) { + break; + } + } + for (int j = i - 1; j >= max(start, 0); --j) { + if (!suppressCharProbabilities( + mostCommonKeyWidth, sampledInputSize, sampledLengthCache, i, j, + charProbabilities)) { + break; + } + } + } + + // Converting from raw probabilities to log probabilities to calculate spatial distance. + for (int i = start; i < sampledInputSize; ++i) { + for (int j = 0; j < keyCount; ++j) { + hash_map_compat<int, float>::iterator it = (*charProbabilities)[i].find(j); + if (it == (*charProbabilities)[i].end()){ + (*nearKeysVector)[i].reset(j); + } else if(it->second < MIN_PROBABILITY) { + // Erases from near keys vector because it has very low probability. + (*nearKeysVector)[i].reset(j); + (*charProbabilities)[i].erase(j); + } else { + it->second = -logf(it->second); + } + } + (*charProbabilities)[i][NOT_AN_INDEX] = -logf((*charProbabilities)[i][NOT_AN_INDEX]); + } +} + +// Decreases char probabilities of index0 by checking probabilities of a near point (index1) and +// increases char probabilities of index1 by checking probabilities of index0. +/* static */ bool ProximityInfoStateUtils::suppressCharProbabilities(const int mostCommonKeyWidth, + const int sampledInputSize, const std::vector<int> *const lengthCache, + const int index0, const int index1, + std::vector<hash_map_compat<int, float> > *charProbabilities) { + ASSERT(0 <= index0 && index0 < sampledInputSize); + ASSERT(0 <= index1 && index1 < sampledInputSize); + + static const float SUPPRESSION_LENGTH_WEIGHT = 1.5f; + static const float MIN_SUPPRESSION_RATE = 0.1f; + static const float SUPPRESSION_WEIGHT = 0.5f; + static const float SUPPRESSION_WEIGHT_FOR_PROBABILITY_GAIN = 0.1f; + static const float SKIP_PROBABALITY_WEIGHT_FOR_PROBABILITY_GAIN = 0.3f; + + const float keyWidthFloat = static_cast<float>(mostCommonKeyWidth); + const float diff = fabsf(static_cast<float>((*lengthCache)[index0] - (*lengthCache)[index1])); + if (diff > keyWidthFloat * SUPPRESSION_LENGTH_WEIGHT) { + return false; + } + const float suppressionRate = MIN_SUPPRESSION_RATE + + diff / keyWidthFloat / SUPPRESSION_LENGTH_WEIGHT * SUPPRESSION_WEIGHT; + for (hash_map_compat<int, float>::iterator it = (*charProbabilities)[index0].begin(); + it != (*charProbabilities)[index0].end(); ++it) { + hash_map_compat<int, float>::iterator it2 = (*charProbabilities)[index1].find(it->first); + if (it2 != (*charProbabilities)[index1].end() && it->second < it2->second) { + const float newProbability = it->second * suppressionRate; + const float suppression = it->second - newProbability; + it->second = newProbability; + // mCharProbabilities[index0][NOT_AN_INDEX] is the probability of skipping this point. + (*charProbabilities)[index0][NOT_AN_INDEX] += suppression; + + // Add the probability of the same key nearby index1 + const float probabilityGain = min(suppression * SUPPRESSION_WEIGHT_FOR_PROBABILITY_GAIN, + (*charProbabilities)[index1][NOT_AN_INDEX] + * SKIP_PROBABALITY_WEIGHT_FOR_PROBABILITY_GAIN); + it2->second += probabilityGain; + (*charProbabilities)[index1][NOT_AN_INDEX] -= probabilityGain; + } + } + return true; +} } // namespace latinime |