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 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()
self.net.caffenet.set_phase_test()
self.imageMean = self.net._IMAGENET_MEAN.swapaxes(1, 2).swapaxes(0, 1).astype('float32')
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()
self.net.caffenet.set_phase_test()
self.imageMean = self.net._IMAGENET_MEAN.swapaxes(1, 2).swapaxes(
0, 1).astype('float32')
def doNetworkTraining(self, samples, labels):
self.solver.net.set_input_arrays(samples, labels)
self.solver.solve()
self.iter += config.geti('trainingIterationsPerBatch')
if self.iter % self.stepSize == 0:
newLR = self.lr * (self.gamma**int(self.iter / self.stepSize))
print 'Changing LR to:', newLR
self.solver.change_lr(newLR)
def doNetworkTraining(self, samples, labels):
self.solver.net.set_input_arrays(samples, labels)
self.solver.solve()
self.iter += config.geti('trainingIterationsPerBatch')
if self.iter % self.stepSize == 0:
newLR = self.lr * ( self.gamma** int(self.iter/self.stepSize) )
print 'Changing LR to:',newLR
self.solver.change_lr(newLR)
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 run(self):
if self.mode == 'train':
self.agent.persistMemory = True
self.agent.startReplayMemory(len(self.environment.imageList), config.geti('trainInteractions'))
#self.agent.assignPriorMemory(self.environment.priorMemory)
self.train()
elif self.mode == 'test':
self.agent.persistMemory = False
self.test()
def run(self):
if self.mode == 'train':
self.agent.persistMemory = True
self.agent.startReplayMemory(len(self.environment.imageList),
config.geti('trainInteractions'))
#self.agent.assignPriorMemory(self.environment.priorMemory)
self.train()
elif self.mode == 'test':
self.agent.persistMemory = False
self.test()
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, imageName, boxReset='Full', groundTruth=None):
self.imageName = imageName
self.visibleImage = Image.open(config.get('imageDir') + '/' + self.imageName + '.jpg')
self.box = [0,0,0,0]
self.resets = 1
self.reset(boxReset)
self.landmarkIndex = {}
self.actionChosen = 2
self.actionValue = 0
self.groundTruth = groundTruth
if self.groundTruth is not None:
self.taskSimulator = bst.BoxSearchTask()
self.taskSimulator.groundTruth = self.groundTruth
self.taskSimulator.loadGroundTruth(self.imageName)
self.stepsWithoutLandmark = 0
self.actionHistory = [0 for i in range(NUM_ACTIONS*config.geti('actionHistoryLength'))]
def doValidation(self, epoch):
if epoch % config.geti('validationEpochs') != 0:
return
auxRL = BoxSearchRunner('test')
auxRL.run()
indexType = config.get('evaluationIndexType')
category = config.get('category')
if indexType == 'pascal':
categories, catIndex = bse.get20Categories()
elif indexType == 'relations':
categories, catIndex = bse.getCategories()
elif indexType == 'finetunedRelations':
categories, catIndex = bse.getRelationCategories()
catI = categories.index(category)
scoredDetections = bse.loadScores(config.get('testMemory'), catI)
groundTruthFile = config.get('testGroundTruth')
ps,rs = bse.evaluateCategory(scoredDetections, 'scores', groundTruthFile)
pl,rl = bse.evaluateCategory(scoredDetections, 'landmarks', groundTruthFile)
line = lambda x,y,z: x + '\t{:5.3f}\t{:5.3f}\n'.format(y,z)
print line('Validation Scores:',ps,rs)
print line('Validation Landmarks:',pl,rl)
def doValidation(self, epoch):
if epoch % config.geti('validationEpochs') != 0:
return
auxRL = BoxSearchRunner('test')
auxRL.run()
indexType = config.get('evaluationIndexType')
category = config.get('category')
if indexType == 'pascal':
categories, catIndex = bse.get20Categories()
elif indexType == 'relations':
categories, catIndex = bse.getCategories()
elif indexType == 'finetunedRelations':
categories, catIndex = bse.getRelationCategories()
catI = categories.index(category)
scoredDetections = bse.loadScores(config.get('testMemory'), catI)
groundTruthFile = config.get('testGroundTruth')
ps, rs = bse.evaluateCategory(scoredDetections, 'scores',
groundTruthFile)
pl, rl = bse.evaluateCategory(scoredDetections, 'landmarks',
groundTruthFile)
line = lambda x, y, z: x + '\t{:5.3f}\t{:5.3f}\n'.format(y, z)
print line('Validation Scores:', ps, rs)
print line('Validation Landmarks:', pl, rl)
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]" 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
def defaultSampler():
return np.random.random([1, config.geti('outputActions')])
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')
self.box = [0,0,0,0]
self.resets = 1
self.reset(boxReset)
self.landmarkIndex = {}
import json
import utils.utils as cu
import utils.libDetection as det
import learn.rl.RLConfig as config
def sigmoid(x, a=1.0, b=0.0):
return 1.0 / (1.0 + np.exp(-a * x + b))
def tanh(x, a=5, b=0.5, c=2.0):
return c * np.tanh(a * x + b)
TEST_TIME_OUT = config.geti('testTimeOut')
ACTION_HISTORY_SIZE = bs.NUM_ACTIONS * config.geti('actionHistoryLength')
class BoxSearchEnvironment(Environment, Named):
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(
__author__ = "Juan C. Caicedo, [email protected]" 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
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]" 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
__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):