Mercurial Hosting > traffic-intelligence
view scripts/learn-motion-patterns.py @ 1052:1748c02f9ac3
Modifying storage.py [configloaders]
author | Wendlasida |
---|---|
date | Thu, 05 Jul 2018 13:48:12 -0400 |
parents | 5621e4ad2428 |
children | b735895c8815 |
line wrap: on
line source
#! /usr/bin/env python3 import sys, argparse import numpy as np import matplotlib.pyplot as plt from trafficintelligence import ml, utils, storage, moving parser = argparse.ArgumentParser(description='''The program clusters trajectories, each cluster being represented by a trajectory. It can either work on the same dataset (database) or different ones, but only does learning or assignment at a time to avoid issues (the minimum cluster size argument is not used for now as it may change prototypes when assigning other trajectories)''') #, epilog = '' #parser.add_argument('--cfg', dest = 'configFilename', help = 'name of the configuration file') parser.add_argument('-d', dest = 'databaseFilename', help = 'name of the Sqlite database file', required = True) parser.add_argument('-o', dest = 'outputPrototypeDatabaseFilename', help = 'name of the Sqlite database file to save prototypes') parser.add_argument('-i', dest = 'inputPrototypeDatabaseFilename', help = 'name of the Sqlite database file for prototypes to start the algorithm with') parser.add_argument('-t', dest = 'trajectoryType', help = 'type of trajectories to learn from', choices = ['objectfeature', 'feature', 'object'], default = 'objectfeatures') parser.add_argument('--max-nobjectfeatures', dest = 'maxNObjectFeatures', help = 'maximum number of features per object to load', type = int, default = 1) parser.add_argument('-n', dest = 'nTrajectories', help = 'number of the object or feature trajectories to load', type = int, default = None) parser.add_argument('-e', dest = 'epsilon', help = 'distance for the similarity of trajectory points', type = float, required = True) parser.add_argument('--metric', dest = 'metric', help = 'metric for the similarity of trajectory points', default = 'cityblock') # default is manhattan distance parser.add_argument('-s', dest = 'minSimilarity', help = 'minimum similarity to put a trajectory in a cluster', type = float, required = True) parser.add_argument('-c', dest = 'minClusterSize', help = 'minimum cluster size', type = int, default = 0) parser.add_argument('--learn', dest = 'learn', help = 'learn', action = 'store_true') parser.add_argument('--optimize', dest = 'optimizeCentroid', help = 'recompute centroid at each assignment', action = 'store_true') parser.add_argument('--random', dest = 'randomInitialization', help = 'random initialization of clustering algorithm', action = 'store_true') parser.add_argument('--subsample', dest = 'positionSubsamplingRate', help = 'rate of position subsampling (1 every n positions)', type = int) parser.add_argument('--display', dest = 'display', help = 'display trajectories', action = 'store_true') parser.add_argument('--save-similarities', dest = 'saveSimilarities', help = 'save computed similarities (in addition to prototypes)', action = 'store_true') parser.add_argument('--save-assignments', dest = 'saveAssignments', help = 'saves the assignments of the objects to the prototypes', action = 'store_true') parser.add_argument('--assign', dest = 'assign', help = 'assigns the objects to the prototypes and saves the assignments', action = 'store_true') args = parser.parse_args() # use cases # 1. learn proto from one file, save in same or another # 2. load proto, load objects (from same or other db), update proto matchings, save proto # TODO 3. on same dataset, learn and assign trajectories (could be done with min cluster size) # TODO? 4. when assigning, allow min cluster size only to avoid assigning to small clusters (but prototypes are not removed even if in small clusters, can be done after assignment with nmatchings) # TODO add possibility to cluster with velocities # TODO add possibility to load all trajectories and use minclustersize # load trajectories to cluster or assign if args.trajectoryType == 'objectfeature': trajectoryType = 'feature' objectFeatureNumbers = storage.loadObjectFeatureFrameNumbers(args.databaseFilename, objectNumbers = args.nTrajectories) featureNumbers = [] for numbers in objectFeatureNumbers.values(): featureNumbers += numbers[:min(len(numbers), args.maxNObjectFeatures)] objects = storage.loadTrajectoriesFromSqlite(args.databaseFilename, 'feature', objectNumbers = featureNumbers, timeStep = args.positionSubsamplingRate) else: trajectoryType = args.trajectoryType objects = storage.loadTrajectoriesFromSqlite(args.databaseFilename, trajectoryType, objectNumbers = args.nTrajectories, timeStep = args.positionSubsamplingRate) trajectories = [o.getPositions().asArray().T for o in objects] # load initial prototypes, if any if args.inputPrototypeDatabaseFilename is not None: initialPrototypes = storage.loadPrototypesFromSqlite(args.inputPrototypeDatabaseFilename, True) trajectories = [p.getMovingObject().getPositions().asArray().T for p in initialPrototypes]+trajectories if len(initialPrototypes) > 0: initialPrototypeIndices = list(range(len(initialPrototypes))) else: initialPrototypeIndices = None else: initialPrototypes = [] initialPrototypeIndices = None lcss = utils.LCSS(metric = args.metric, epsilon = args.epsilon) nTrajectories = len(trajectories) similarities = -np.ones((nTrajectories, nTrajectories)) similarityFunc = lambda x,y : lcss.computeNormalized(x, y) # the next line can be called again without reinitializing similarities if args.learn: prototypeIndices = ml.prototypeCluster(trajectories, similarities, args.minSimilarity, similarityFunc, args.optimizeCentroid, args.randomInitialization, initialPrototypeIndices) else: prototypeIndices = initialPrototypeIndices if args.assign: # TODO don't touch initial prototypes if not from same db as trajectories #if not args.learn and args.minClusterSize >= 1: # allow only # print('Warning: you did not learn the prototypes and you are using minimum cluster size of {}, which may lead to removing prototypes and assigning them to others'.format(args.minClusterSize)) # if args.minClusterSize >= 1: # if initialPrototypeIndices is None: # prototypeIndices, labels = ml.assignToPrototypeClusters(trajectories, prototypeIndices, similarities, args.minSimilarity, similarityFunc, args.minClusterSize) # else: # print('Not assigning with non-zero minimum cluster size and initial prototypes (would remove initial prototypes based on other trajectories') # else: # prototypeIndices, labels = ml.assignToPrototypeClusters(trajectories, prototypeIndices, similarities, args.minSimilarity, similarityFunc) assignedPrototypeIndices, labels = ml.assignToPrototypeClusters(trajectories, prototypeIndices, similarities, args.minSimilarity, similarityFunc) if args.learn and not args.assign: prototypes = [] for i in prototypeIndices: if i<len(initialPrototypes): prototypes.append(initialPrototypes[i]) else: prototypes.append(moving.Prototype(args.databaseFilename, objects[i-len(initialPrototypes)].getNum(), trajectoryType)) if args.outputPrototypeDatabaseFilename is None: outputPrototypeDatabaseFilename = args.databaseFilename else: outputPrototypeDatabaseFilename = args.outputPrototypeDatabaseFilename if args.inputPrototypeDatabaseFilename == args.outputPrototypeDatabaseFilename: storage.deleteFromSqlite(args.outputPrototypeDatabaseFilename, 'prototype') storage.savePrototypesToSqlite(outputPrototypeDatabaseFilename, prototypes) if args.display: plt.figure() for p in prototypes: p.getMovingObject().plot() plt.axis('equal') plt.show() if not args.learn and args.assign: # no modification to prototypes, can work with initialPrototypes clusterSizes = ml.computeClusterSizes(labels, prototypeIndices, -1) for i in prototypeIndices: nMatchings = clusterSizes[i]-1 if initialPrototypes[i].nMatchings is None: initialPrototypes[i].nMatchings = nMatchings else: initialPrototypes[i].nMatchings += nMatchings if args.outputPrototypeDatabaseFilename is None: outputPrototypeDatabaseFilename = args.databaseFilename else: outputPrototypeDatabaseFilename = args.outputPrototypeDatabaseFilename storage.setPrototypeMatchingsInSqlite(outputPrototypeDatabaseFilename, initialPrototypes) if args.saveAssignments: if args.trajectoryType == 'objectfeature': # consider that the object is assigned through its longest features objectNumbers = [] objectLabels = [] for objNum, objFeatureNumbers in objectFeatureNumbers.items(): objLabels = [] for i, o in enumerate(objects): if o.getNum() in objFeatureNumbers: objLabels.append(labels[i+len(initialPrototypes)]) objectLabels.append(utils.mostCommon(objLabels)) objectNumbers.append(objNum) storage.savePrototypeAssignmentsToSqlite(args.databaseFilename, objectNumbers, 'object', objectLabels, initialPrototypes) else: storage.savePrototypeAssignmentsToSqlite(args.databaseFilename, [obj.getNum() for obj in objects], trajectoryType, labels[len(initialPrototypes):], initialPrototypes) if args.display: plt.figure() for i,o in enumerate(objects): if labels[i+len(initialPrototypes)] < 0: o.plot('kx-') else: o.plot(utils.colors[labels[i+len(initialPrototypes)]]) for i,p in enumerate(initialPrototypes): p.getMovingObject().plot(utils.colors[i]+'o') plt.axis('equal') plt.show() if (args.learn or args.assign) and args.saveSimilarities: np.savetxt(utils.removeExtension(args.databaseFilename)+'-prototype-similarities.txt.gz', similarities, '%.4f')