def mergeDatasetAndRecords(self, train, val):
numTrain = len(self.dataset) * (1 - config.getf('percentOfValidation'))
numVal = len(self.dataset) * config.getf('percentOfValidation')
random.shuffle(self.dataset)
for i in range(len(self.dataset)):
imgPath = config.get('imageDir') + self.dataset[i]['img'] + '.jpg'
# record format: Action, reward, discountedMaxQ, x1, y1, x2, y2,
record = [self.dataset[i]['A'], self.dataset[i]['R'], 0.0
] + self.dataset[i]['box']
record += [self.dataset[i]['Sp']] + self.dataset[i]['Xp'] + [
self.dataset[i]['Sc']
] + self.dataset[i]['Xc']
record += [self.dataset[i]['Ap']]
if i < numTrain:
try:
train[imgPath].append(record)
except:
train[imgPath] = [record]
else:
try:
val[imgPath].append(record)
except:
val[imgPath] = [record]
return train, val
def __init__(self, workingDir):
self.directory = workingDir
self.writeSolverFile()
self.solver = caffe.SGDSolver(self.directory + 'solver.prototxt')
self.iter = 0
self.itersPerEpisode = config.geti('trainingIterationsPerBatch')
self.lr = config.getf('learningRate')
self.stepSize = config.geti('stepSize')
self.gamma = config.getf('gamma')
print 'CAFFE SOLVER INITALIZED'
def __init__(self, workingDir):
self.directory = workingDir
self.writeSolverFile()
self.solver = caffe.SGDSolver(self.directory + 'solver.prototxt')
self.iter = 0
self.itersPerEpisode = config.geti('trainingIterationsPerBatch')
self.lr = config.getf('learningRate')
self.stepSize = config.geti('stepSize')
self.gamma = config.getf('gamma')
print 'CAFFE SOLVER INITALIZED'
def train(self):
interactions = config.geti('trainInteractions')
minEpsilon = config.getf('minTrainingEpsilon')
epochSize = len(self.environment.db.images) / 2
epsilon = 1.0
self.controller.setEpsilonGreedy(epsilon)
print 'Epoch 0: Exploration'
self.runEpoch(interactions, len(self.environment.db.images))
self.learner = QLearning()
self.agent.learner = self.learner
epoch = 1
egEpochs = config.geti('epsilonGreedyEpochs')
while epoch <= egEpochs:
epsilon = epsilon - (1.0 - minEpsilon) / float(egEpochs)
if epsilon < minEpsilon: epsilon = minEpsilon
self.controller.setEpsilonGreedy(epsilon)
print 'Epoch', epoch, '(epsilon-greedy:{:5.3f})'.format(epsilon)
self.runEpoch(interactions, epochSize)
epoch += 1
epoch = 1
maxEpochs = config.geti('exploitLearningEpochs')
while epoch <= maxEpochs:
print 'Epoch', epoch + egEpochs, '(exploitation mode: epsilon={:5.3f})'.format(
epsilon)
self.runEpoch(interactions, epochSize)
epoch += 1
def train(self):
interactions = config.geti('trainInteractions')
minEpsilon = config.getf('minTrainingEpsilon')
epochSize = len(self.environment.db.images)/2
epsilon = 1.0
self.controller.setEpsilonGreedy(epsilon)
print 'Epoch 0: Exploration'
self.runEpoch(interactions, len(self.environment.db.images))
self.learner = QLearning()
self.agent.learner = self.learner
epoch = 1
egEpochs = config.geti('epsilonGreedyEpochs')
while epoch <= egEpochs:
epsilon = epsilon - (1.0-minEpsilon)/float(egEpochs)
if epsilon < minEpsilon: epsilon = minEpsilon
self.controller.setEpsilonGreedy(epsilon)
print 'Epoch',epoch ,'(epsilon-greedy:{:5.3f})'.format(epsilon)
self.runEpoch(interactions, epochSize)
epoch += 1
epoch = 1
maxEpochs = config.geti('exploitLearningEpochs')
while epoch <= maxEpochs:
print 'Epoch',epoch+egEpochs,'(exploitation mode: epsilon={:5.3f})'.format(epsilon)
self.runEpoch(interactions, epochSize)
epoch += 1
def __init__(self, imageList, mode, groundTruthFile=None):
self.mode = mode
self.cnn = cn.ConvNet()
self.testRecord = None
self.idx = -1
self.imageList = [x.strip() for x in open(imageList)]
self.groundTruth = cu.loadBoxIndexFile(groundTruthFile)
#self.imageList = self.rankImages()
#self.imageList = self.imageList[0:10]
allImgs = set([x.strip() for x in open(config.get('allImagesList'))])
self.negativeSamples = list(allImgs.difference(set(self.groundTruth.keys())))
self.negativeEpisode = False
if self.mode == 'train':
self.negativeProbability = config.getf('negativeEpisodeProb')
random.shuffle(self.imageList)
self.loadNextEpisode()
def __init__(self, imageList, mode, groundTruthFile=None):
self.mode = mode
self.cnn = cn.ConvNet()
self.testRecord = None
self.idx = -1
self.imageList = [x.strip() for x in open(imageList)]
self.groundTruth = cu.loadBoxIndexFile(groundTruthFile)
#self.imageList = self.rankImages()
#self.imageList = self.imageList[0:10]
allImgs = set([x.strip() for x in open(config.get('allImagesList'))])
self.negativeSamples = list(
allImgs.difference(set(self.groundTruth.keys())))
self.negativeEpisode = False
if self.mode == 'train':
self.negativeProbability = config.getf('negativeEpisodeProb')
random.shuffle(self.imageList)
self.loadNextEpisode()
def train(self):
networkFile = config.get('networkDir') + config.get(
'snapshotPrefix') + '_iter_' + config.get(
'trainingIterationsPerBatch') + '.caffemodel'
interactions = config.geti('trainInteractions')
minEpsilon = config.getf('minTrainingEpsilon')
epochSize = len(self.environment.imageList) / 1
epsilon = 1.0
self.controller.setEpsilonGreedy(epsilon,
self.environment.sampleAction)
epoch = 1
exEpochs = config.geti('explorationEpochs')
while epoch <= exEpochs:
s = cu.tic()
print 'Epoch', epoch, ': Exploration (epsilon=1.0)'
self.runEpoch(interactions, len(self.environment.imageList))
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ', s)
epoch += 1
self.learner = QLearning()
self.agent.learner = self.learner
egEpochs = config.geti('epsilonGreedyEpochs')
while epoch <= egEpochs + exEpochs:
s = cu.tic()
epsilon = epsilon - (1.0 - minEpsilon) / float(egEpochs)
if epsilon < minEpsilon: epsilon = minEpsilon
self.controller.setEpsilonGreedy(epsilon,
self.environment.sampleAction)
print 'Epoch', epoch, '(epsilon-greedy:{:5.3f})'.format(epsilon)
self.runEpoch(interactions, epochSize)
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ', s)
epoch += 1
maxEpochs = config.geti('exploitLearningEpochs') + exEpochs + egEpochs
while epoch <= maxEpochs:
s = cu.tic()
print 'Epoch', epoch, '(exploitation mode: epsilon={:5.3f})'.format(
epsilon)
self.runEpoch(interactions, epochSize)
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ', s)
shutil.copy(networkFile, networkFile + '.' + str(epoch))
epoch += 1
def balanceTrainingExamples(self, candidates):
pos, neg = {},{}
psc, ngc = 0,0
for k in candidates.keys():
for box in candidates[k]:
if box[5] > config.getf('minPositiveOverlap'): # and box[6] >= 1.0:
try: pos[k].append(box)
except: pos[k] = [box]
psc += 1
else:
try: neg[k].append(box)
except: neg[k] = [box]
ngc += 1
self.pos = ImageBoxIndex(pos,True)
self.neg = ImageBoxIndex(neg,True)
self.probNeg = max(float(psc)/float(ngc),0.05)
print 'Positives:',psc,'Negatives:',ngc
def train(self):
networkFile = config.get('networkDir') + config.get('snapshotPrefix') + '_iter_' + config.get('trainingIterationsPerBatch') + '.caffemodel'
interactions = config.geti('trainInteractions')
minEpsilon = config.getf('minTrainingEpsilon')
epochSize = len(self.environment.imageList)/1
epsilon = 1.0
self.controller.setEpsilonGreedy(epsilon, self.environment.sampleAction)
epoch = 1
exEpochs = config.geti('explorationEpochs')
while epoch <= exEpochs:
s = cu.tic()
print 'Epoch',epoch,': Exploration (epsilon=1.0)'
self.runEpoch(interactions, len(self.environment.imageList))
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ',s)
epoch += 1
self.learner = QLearning()
self.agent.learner = self.learner
egEpochs = config.geti('epsilonGreedyEpochs')
while epoch <= egEpochs + exEpochs:
s = cu.tic()
epsilon = epsilon - (1.0-minEpsilon)/float(egEpochs)
if epsilon < minEpsilon: epsilon = minEpsilon
self.controller.setEpsilonGreedy(epsilon, self.environment.sampleAction)
print 'Epoch',epoch ,'(epsilon-greedy:{:5.3f})'.format(epsilon)
self.runEpoch(interactions, epochSize)
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ',s)
epoch += 1
maxEpochs = config.geti('exploitLearningEpochs') + exEpochs + egEpochs
while epoch <= maxEpochs:
s = cu.tic()
print 'Epoch',epoch,'(exploitation mode: epsilon={:5.3f})'.format(epsilon)
self.runEpoch(interactions, epochSize)
self.task.flushStats()
self.doValidation(epoch)
s = cu.toc('Epoch done in ',s)
shutil.copy(networkFile, networkFile + '.' + str(epoch))
epoch += 1
def balanceTrainingExamples(self, candidates):
pos, neg = {}, {}
psc, ngc = 0, 0
for k in candidates.keys():
for box in candidates[k]:
if box[5] > config.getf(
'minPositiveOverlap'): # and box[6] >= 1.0:
try:
pos[k].append(box)
except:
pos[k] = [box]
psc += 1
else:
try:
neg[k].append(box)
except:
neg[k] = [box]
ngc += 1
self.pos = ImageBoxIndex(pos, True)
self.neg = ImageBoxIndex(neg, True)
self.probNeg = max(float(psc) / float(ngc), 0.05)
print 'Positives:', psc, 'Negatives:', ngc
def test(self):
interactions = config.geti('testInteractions')
self.controller.setEpsilonGreedy(config.getf('testEpsilon'))
self.runEpoch(interactions, len(self.environment.db.images))
def __init__(self, alpha=0.5):
ValueBasedLearner.__init__(self)
self.alpha = alpha
self.gamma = config.getf('gammaDiscountReward')
self.netManager = CaffeMultiLayerPerceptronManagement(config.get('networkDir'))
import RLConfig as config
import numpy as np
import scipy.io
import MemoryUsage
import RLConfig as config
import BoxSearchState as bss
import random
STATE_FEATURES = config.geti('stateFeatures')/config.geti('temporalWindow')
NUM_ACTIONS = config.geti('outputActions')
TEMPORAL_WINDOW = config.geti('temporalWindow')
HISTORY_FACTOR = config.geti('historyFactor')
NEGATIVE_PROBABILITY = config.getf('negativeEpisodeProb')
class BoxSearchAgent():
image = None
observation = None
action = None
reward = None
timer = 0
def __init__(self, qnet, learner=None):
self.controller = qnet
self.learner = learner
self.avgReward = 0
self.replayMemory = None
def test(self):
interactions = config.geti('testInteractions')
self.controller.setEpsilonGreedy(config.getf('testEpsilon'))
self.runEpoch(interactions, len(self.environment.imageList))
__author__ = "Juan C. Caicedo, [email protected]" from pybrain.rl.environments import Task import BoxSearchState as bss import utils as cu import libDetection as det import numpy as np import RLConfig as config MIN_ACCEPTABLE_IOU = config.getf('minAcceptableIoU') def center(box): return [ (box[2] + box[0])/2.0 , (box[3] + box[1])/2.0 ] def euclideanDist(c1, c2): return (c1[0] - c2[0])**2 + (c1[1] - c2[1])**2 class BoxSearchTask(Task): def __init__(self, environment=None, groundTruthFile=None): Task.__init__(self, environment) if groundTruthFile is not None: self.groundTruth = cu.loadBoxIndexFile(groundTruthFile) self.image = '' self.epochRecall = [] self.epochMaxIoU = [] self.epochLandmarks = [] def getReward(self):
import RLConfig as config
import numpy as np
import scipy.io
import MemoryUsage
import RLConfig as config
import BoxSearchState as bss
import random
STATE_FEATURES = config.geti('stateFeatures') / config.geti('temporalWindow')
NUM_ACTIONS = config.geti('outputActions')
TEMPORAL_WINDOW = config.geti('temporalWindow')
HISTORY_FACTOR = config.geti('historyFactor')
NEGATIVE_PROBABILITY = config.getf('negativeEpisodeProb')
class BoxSearchAgent():
image = None
observation = None
action = None
reward = None
timer = 0
def __init__(self, qnet, learner=None):
self.controller = qnet
self.learner = learner
self.avgReward = 0
self.replayMemory = None
__author__ = "Juan C. Caicedo, [email protected]" from pybrain.rl.environments import Task import BoxSearchState as bss import utils as cu import libDetection as det import numpy as np import RLConfig as config MIN_ACCEPTABLE_IOU = config.getf('minAcceptableIoU') def center(box): return [(box[2] + box[0]) / 2.0, (box[3] + box[1]) / 2.0] def euclideanDist(c1, c2): return (c1[0] - c2[0])**2 + (c1[1] - c2[1])**2 class BoxSearchTask(Task): def __init__(self, environment=None, groundTruthFile=None): Task.__init__(self, environment) if groundTruthFile is not None: self.groundTruth = cu.loadBoxIndexFile(groundTruthFile) self.image = '' self.epochRecall = [] self.epochMaxIoU = [] self.epochLandmarks = []
def __init__(self, alpha=0.5):
ValueBasedLearner.__init__(self)
self.alpha = alpha
self.gamma = config.getf('gammaDiscountReward')
self.netManager = CaffeMultiLayerPerceptronManagement(
config.get('networkDir'))