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):
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()
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 __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.priorMemory = PriorMemory(config.get('allObjectsBoxes'), self.groundTruth, self.cnn)
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()
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 __init__(self, alpha=0.5):
ValueBasedLearner.__init__(self)
self.alpha = alpha
self.gamma = config.getf('gammaDiscountReward')
self.netManager = CaffeMultiLayerPerceptronManagement(config.get('networkDir'))
__author__ = "Juan C. Caicedo, [email protected]" import os import random import numpy as np import scipy.io import caffe import learn.rl.RLConfig as config import BoxSearchState as bs from pybrain.rl.learners.valuebased.valuebased import ValueBasedLearner DETECTION_REWARD = config.getf('detectionReward') ACTION_HISTORY_SIZE = bs.NUM_ACTIONS*config.geti('actionHistoryLength') ACTION_HISTORY_LENTH = config.geti('actionHistoryLength') NETWORK_INPUTS = config.geti('stateFeatures')/config.geti('temporalWindow') REPLAY_MEMORY_SIZE = config.geti('trainingIterationsPerBatch')*config.geti('trainingBatchSize') def generateRandomActionHistory(): actions = np.zeros((ACTION_HISTORY_SIZE)) history = [i*bs.NUM_ACTIONS + np.random.randint(0,bs.PLACE_LANDMARK) for i in range(ACTION_HISTORY_LENTH)] actions[history] = 1 return actions class QLearning(ValueBasedLearner): offPolicy = True batchMode = True dataset = [] trainingSamples = 0
__author__ = "Juan C. Caicedo, [email protected]" from pybrain.rl.environments import Task import BoxSearchState as bss import utils.utils as cu import utils.libDetection as det import numpy as np import learn.rl.RLConfig as config MIN_ACCEPTABLE_IOU = config.getf('minAcceptableIoU') DETECTION_REWARD = config.getf('detectionReward') 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 = []
X_COORD_UP = 0
Y_COORD_UP = 1
SCALE_UP = 2
ASPECT_RATIO_UP = 3
X_COORD_DOWN = 4
Y_COORD_DOWN = 5
SCALE_DOWN = 6
ASPECT_RATIO_DOWN = 7
PLACE_LANDMARK = 8
SKIP_REGION = 9
# BOX LIMITS
MIN_ASPECT_RATIO = 0.15
MAX_ASPECT_RATIO = 6.00
MIN_BOX_SIDE = 10
STEP_FACTOR = config.getf('boxResizeStep')
DELTA_SIZE = config.getf('boxResizeStep')
# OTHER DEFINITIONS
NUM_ACTIONS = config.geti('outputActions')
RESET_BOX_FACTOR = 2
QUADRANT_SIZE = 0.7
def fingerprint(b):
return '_'.join( map(str, map(int, b)) )
class BoxSearchState():
def __init__(self, imageName, boxReset='Full', groundTruth=None):
self.imageName = imageName
self.visibleImage = Image.open(config.get('imageDir') + '/' + self.imageName + '.jpg')
def __init__(self, alpha=0.5):
ValueBasedLearner.__init__(self)
self.alpha = alpha
self.gamma = config.getf('gammaDiscountReward')
self.netManager = CaffeMultiLayerPerceptronManagement(
config.get('networkDir'))
__author__ = "Juan C. Caicedo, [email protected]" from pybrain.rl.environments import Task import BoxSearchState as bss import utils.utils as cu import utils.libDetection as det import numpy as np import learn.rl.RLConfig as config MIN_ACCEPTABLE_IOU = config.getf('minAcceptableIoU') DETECTION_REWARD = config.getf('detectionReward') 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 = []
__author__ = "Juan C. Caicedo, [email protected]" import os import random import numpy as np import scipy.io import caffe import learn.rl.RLConfig as config import BoxSearchState as bs from pybrain.rl.learners.valuebased.valuebased import ValueBasedLearner DETECTION_REWARD = config.getf('detectionReward') ACTION_HISTORY_SIZE = bs.NUM_ACTIONS * config.geti('actionHistoryLength') ACTION_HISTORY_LENTH = config.geti('actionHistoryLength') NETWORK_INPUTS = config.geti('stateFeatures') / config.geti('temporalWindow') REPLAY_MEMORY_SIZE = config.geti('trainingIterationsPerBatch') * config.geti( 'trainingBatchSize') def generateRandomActionHistory(): actions = np.zeros((ACTION_HISTORY_SIZE)) history = [ i * bs.NUM_ACTIONS + np.random.randint(0, bs.PLACE_LANDMARK) for i in range(ACTION_HISTORY_LENTH) ] actions[history] = 1 return actions class QLearning(ValueBasedLearner):
def test(self):
interactions = config.geti('testInteractions')
self.controller.setEpsilonGreedy(config.getf('testEpsilon'))
self.runEpoch(interactions, len(self.environment.imageList))
import learn.rl.RLConfig as config
import numpy as np
import scipy.io
import utils.MemoryUsage
import BoxSearchState as bss
import PriorMemory as prm
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
self.priorMemory = None
__author__ = "Juan C. Caicedo, [email protected]" import os import utils.utils as cu import numpy as np import caffe from caffe import wrapperv0 import learn.rl.RLConfig as config LAYER = config.get('convnetLayer') MARK_WIDTH = config.getf('markWidth') class ConvNet(): def __init__(self): self.net = None self.image = '' self.id = 0 self.loadNetwork() def loadNetwork(self): self.imgDim = config.geti('imageDim') self.cropSize = config.geti('cropSize') self.contextPad = config.geti('contextPad') #self.stateContextFactor = config.geti('stateContextFactor') modelFile = config.get('convnetDir') + config.get('convNetDef') networkFile = config.get('convnetDir') + config.get('trainedConvNet') self.net = wrapperv0.ImageNetClassifier(modelFile, networkFile, IMAGE_DIM=self.imgDim, CROPPED_DIM=self.cropSize, MEAN_IMAGE=config.get('meanImage')) self.net.caffenet.set_mode_gpu()
def test(self):
interactions = config.geti('testInteractions')
self.controller.setEpsilonGreedy(config.getf('testEpsilon'))
self.runEpoch(interactions, len(self.environment.imageList))
import learn.rl.RLConfig as config
import numpy as np
import scipy.io
import utils.MemoryUsage
import BoxSearchState as bss
import PriorMemory as prm
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]" import os import utils.utils as cu import numpy as np import caffe from caffe import wrapperv0 import learn.rl.RLConfig as config LAYER = config.get('convnetLayer') MARK_WIDTH = config.getf('markWidth') class ConvNet(): def __init__(self): self.net = None self.image = '' self.id = 0 self.loadNetwork() def loadNetwork(self): self.imgDim = config.geti('imageDim') self.cropSize = config.geti('cropSize') self.contextPad = config.geti('contextPad') #self.stateContextFactor = config.geti('stateContextFactor') modelFile = config.get('convnetDir') + config.get('convNetDef') networkFile = config.get('convnetDir') + config.get('trainedConvNet') self.net = wrapperv0.ImageNetClassifier( modelFile,