def main():
"""
Inputs:
None
Outputs:
Runs Testing code
"""
# TODO: Make LogDir
# TODO: Display time to end and cleanup other print statements with color
# TODO: Make logging file a parameter
# Parse Command Line arguments
Parser = argparse.ArgumentParser()
Parser.add_argument('--ModelPath', dest='ModelPath', default='/media/nitin/Research/EVDodge/CheckpointsEVDBHomographyDodgeNetLR5e-4Epoch200/199model.ckpt',\
help='Path to load latest model from, Default:ModelPath')
Parser.add_argument('--ReadPath', dest='ReadPath', default='/media/nitin/Research/EVDodge/DatasetChethanEvents/DeblurredHomography',\
help='Path to load images from, Default:ReadPath')
Parser.add_argument('--WritePath', dest='WritePath', default='/media/nitin/Research/EVDodge/DatasetChethanEvents/DeblurredHomographyDodgeNet',\
help='Path to write images to, Default:WritePath')
Parser.add_argument(
'--GPUDevice',
type=int,
default=0,
help='What GPU do you want to use? -1 for CPU, Default:0')
Args = Parser.parse_args()
ModelPath = Args.ModelPath
ReadPath = Args.ReadPath
WritePath = Args.WritePath
GPUDevice = Args.GPUDevice
# Set GPUDevice
tu.SetGPU(GPUDevice)
# Setup all needed parameters including file reading
TrainNames, ImageSize, PatchSize, NumTrainSamples, NumImgsStack = SetupAll(
ReadPath)
# Define PlaceHolder variables for Input and Predicted output
PatchPH = tf.placeholder(tf.float32,
shape=(1, PatchSize[0], PatchSize[1],
2 * PatchSize[2]),
name='Input')
TestOperation(PatchPH, PatchSize, ModelPath, ReadPath, WritePath,
TrainNames, NumTrainSamples)
def main():
"""
Inputs:
None
Outputs:
Runs the Training and testing code based on the Flag
"""
# TODO: Make LogDir
# TODO: Make logging file a parameter
# TODO: Time to complete print
# Parse Command Line arguments
Parser = argparse.ArgumentParser()
Parser.add_argument('--BasePath', default='/home/nitin/EVDodge/DatasetsForPaper/DownfacingEventsProcessed', help='Base path of images, Default:/home/nitin/EVDodge/DatasetsForPaper/DownfacingEventsProcessed')
Parser.add_argument('--NumEpochs', type=int, default=200, help='Number of Epochs to Train for, Default:200')
Parser.add_argument('--DivTrain', type=int, default=1, help='Factor to reduce Train data by per epoch, Default:1')
Parser.add_argument('--MiniBatchSize', type=int, default=256, help='Size of the MiniBatch to use, Default:256')
Parser.add_argument('--LoadCheckPoint', type=int, default=0, help='Load Model from latest Checkpoint from CheckPointsPath?, Default:0')
Parser.add_argument('--LossFuncName', default='PhotoL1', help='Choice of Loss functions, choose from PhotoL1, PhotoChab, PhotoRobust when using TrainingType as US. Default:PhotoL1')
Parser.add_argument('--NetworkType', default='Small', help='Choice of Network type, choose from Small, Large, Default:Small')
Parser.add_argument('--CheckPointPath', default='../CheckpointsHomography/', help='Path to save checkpoints, Default:../CheckpointsHomography/')
Parser.add_argument('--LogsPath', default='/media/nitin/Research/EVDodge/Logs/', help='Path to save Logs, Default:/media/nitin/Research/EVDodge/Logs/')
Parser.add_argument('--GPUDevice', type=int, default=0, help='What GPU do you want to use? -1 for CPU, Default:0')
Parser.add_argument('--LR', type=float, default=1e-4, help='Learning Rate, Default: 1e-4')
Parser.add_argument('--TrainingType', default='US', help='Training Type, S: Supervised, US: Unsupervised, Default: US')
Args = Parser.parse_args()
NumEpochs = Args.NumEpochs
BasePath = Args.BasePath
DivTrain = float(Args.DivTrain)
MiniBatchSize = Args.MiniBatchSize
LoadCheckPoint = Args.LoadCheckPoint
LossFuncName = Args.LossFuncName
NetworkType = Args.NetworkType
CheckPointPath = Args.CheckPointPath
LogsPath = Args.LogsPath
GPUDevice = Args.GPUDevice
LearningRate = Args.LR
TrainingType = Args.TrainingType
# Set GPUDevice
tu.SetGPU(GPUDevice)
# Setup all needed parameters including file reading
TrainNames, ValNames, TestNames, OptimizerParams,\
SaveCheckPoint, ImageSize, Rho, NumTrainSamples, NumValSamples, NumTestSamples,\
NumTestRunsPerEpoch = SetupAll(BasePath, LearningRate)
# If CheckPointPath doesn't exist make the path
if(not (os.path.isdir(CheckPointPath))):
os.makedirs(CheckPointPath)
# Find Latest Checkpoint File
if LoadCheckPoint==1:
LatestFile = FindLatestModel(CheckPointPath)
else:
LatestFile = None
# Pretty print stats
PrettyPrint(NumEpochs, DivTrain, MiniBatchSize, NumTrainSamples, NumTestSamples, LatestFile)
# Define PlaceHolder variables for Input and Predicted output
ImgPH = tf.placeholder(tf.float32, shape=(MiniBatchSize, ImageSize[0], ImageSize[1], 6), name='Input')
# PH for losses
I1PH = tf.placeholder(tf.float32, shape=(MiniBatchSize, 260, 346, 3), name='I1')
I2PH = tf.placeholder(tf.float32, shape=(MiniBatchSize, 260, 346, 3), name='I2')
MaskPH = tf.placeholder(tf.float32, shape=(MiniBatchSize, 260, 346, 3), name='Mask')
AllPtsPH = tf.placeholder(tf.float32, shape=(MiniBatchSize, 4, 2), name='AllPts')
LabelPH = tf.placeholder(tf.float32, shape=(MiniBatchSize, 8, 1), name='Label')
TrainOperation(ImgPH, I1PH, I2PH, MaskPH, AllPtsPH, LabelPH, TrainNames, TestNames, NumTrainSamples, ImageSize, Rho,
NumEpochs, MiniBatchSize, OptimizerParams, SaveCheckPoint, CheckPointPath, NumTestRunsPerEpoch,
DivTrain, LatestFile, LossFuncName, NetworkType, BasePath, LogsPath, TrainingType)
def TrainOperation(ImgPH, I1PH, I2PH, MaskPH, AllPtsPH, LabelPH, TrainNames, TestNames, NumTrainSamples, ImageSize, Rho,
NumEpochs, MiniBatchSize, OptimizerParams, SaveCheckPoint, CheckPointPath, NumTestRunsPerEpoch,
DivTrain, LatestFile, LossFuncName, NetworkType, BasePath, LogsPath, TrainingType):
"""
Inputs:
ImgPH is the Input Image placeholder
HomingVecPH is the ground truth homing vector placeholder
DirNames - Full path to all image files without extension
Train/Val - Idxs of all the images to be used for training/validation (held-out testing in this case)
Train/ValLabels - Labels corresponding to Train/Val
NumTrain/ValSamples - length(Train/Val)
ImageSize - Size of the image
NumEpochs - Number of passes through the Train data
MiniBatchSize is the size of the MiniBatch
OptimizerParams - List of all OptimizerParams: depends on Optimizer
SaveCheckPoint - Save checkpoint every SaveCheckPoint iteration in every epoch, checkpoint saved automatically after every epoch
CheckPointPath - Path to save checkpoints/model
NumTestRunsPerEpoch - Number of passes of Val data with MiniBatchSize
DivTrain - Divide the data by this number for Epoch calculation, use if you have a lot of data
LatestFile - Latest checkpointfile to continue training
Outputs:
Saves Trained network in CheckPointPath
"""
# Predict output with forward pass
if(NetworkType == 'Small'):
prHVal = EVHomographyNetUnsupSmall(ImgPH, ImageSize, MiniBatchSize)
with tf.name_scope('Loss'):
loss, WarpI1, WarpMask, WarpI1Patch, I2Patch, _ = LossFunc(I1PH, I2PH, MaskPH, AllPtsPH, LabelPH, prHVal, MiniBatchSize, LossFuncName, TrainingType)
elif(NetworkType == 'Large'):
prHVal = EVHomographyNetUnsup(ImgPH, ImageSize, MiniBatchSize)
with tf.name_scope('Loss'):
loss, WarpI1, WarpMask, WarpI1Patch, I2Patch, _ = LossFunc(I1PH, I2PH, MaskPH, AllPtsPH, LabelPH, prHVal, MiniBatchSize, LossFuncName, TrainingType)
elif(NetworkType == 'SmallRobust'):
prHVal, prAlpha = EVHomographyNetUnsupSmallRobust(ImgPH, ImageSize, MiniBatchSize)
with tf.name_scope('Loss'):
loss, WarpI1, WarpMask, WarpI1Patch, I2Patch, Alpha = LossFunc(I1PH, I2PH, MaskPH, AllPtsPH, LabelPH, prHVal, MiniBatchSize, LossFuncName, TrainingType, a=prAlpha, ImageSize=ImageSize)
with tf.name_scope('Adam'):
Optimizer = tf.train.AdamOptimizer(learning_rate=OptimizerParams[0], beta1=OptimizerParams[1],
beta2=OptimizerParams[2], epsilon=OptimizerParams[3])
Gradients = Optimizer.compute_gradients(loss)
OptimizerUpdate = Optimizer.apply_gradients(Gradients)
#Optimizer = tf.train.GradientDescentOptimizer(learning_rate=1e-8).minimize(loss)
#Optimizer = tf.train.MomentumOptimizer(learning_rate=1e-3, momentum=0.9, use_nesterov=True).minimize(loss)
# Tensorboard
# Create a summary to monitor loss tensor
tf.summary.scalar('LossEveryIter', loss)
tf.summary.image('WarpedImg', WarpI1[:,:,:,0:3])
tf.summary.image('I1', I1PH[:,:,:,0:3])
tf.summary.image('I2', I2PH[:,:,:,0:3])
tf.summary.image('Mask', MaskPH[:,:,:,0:3])
tf.summary.image('WarpMask', WarpMask[:,:,:,0:3])
# tf.summary.image('WarpI1Patch', WarpI1Patch[:,:,:,0:3])
# tf.summary.image('I2Patch', I2Patch[:,:,:,0:3])
# tf.summary.image('I2Patch - WarpI1Patch', tf.abs(I2Patch[:,:,:,0:3] - WarpI1Patch[:,:,:,0:3]))
tf.summary.histogram('prHVal', prHVal, collections=None, family=None)
if(NetworkType == 'SmallRobust'):
tf.summary.image('Alpha1', Alpha[:,:,:,0:1])
tf.summary.image('Alpha2', Alpha[:,:,:,1:2])
tf.summary.image('Alpha3', Alpha[:,:,:,2:3])
tf.summary.histogram('Alpha1', Alpha[:,:,:,0:1], collections=None, family=None)
tf.summary.histogram('Alpha2', Alpha[:,:,:,1:2], collections=None, family=None)
tf.summary.histogram('Alpha3', Alpha[:,:,:,2:3], collections=None, family=None)
# Merge all summaries into a single operation
MergedSummaryOP = tf.summary.merge_all()
AllEpochLoss = [0.0]
EachIterLoss = [0.0]
# Setup Saver
Saver = tf.train.Saver()
with tf.Session() as sess:
if LatestFile is not None:
Saver.restore(sess, CheckPointPath + LatestFile + '.ckpt')
# Extract only numbers from the name
StartEpoch = int(''.join(c for c in LatestFile.split('a')[0] if c.isdigit()))
print('Loaded latest checkpoint with the name ' + LatestFile + '....')
else:
sess.run(tf.global_variables_initializer())
StartEpoch = 0
print('New model initialized....')
# Print Number of parameters in the network
tu.FindNumParams(1)
# Tensorboard
Writer = tf.summary.FileWriter(LogsPath, graph=tf.get_default_graph())
TotalTimeElapsed = 0.0
TimerOverall = tic()
for Epochs in tqdm(range(StartEpoch, NumEpochs)):
EpochLoss = 0.0
Timer1 = tic()
NumIterationsPerEpoch = int(NumTrainSamples/MiniBatchSize/DivTrain)
for PerEpochCounter in tqdm(range(NumIterationsPerEpoch)):
print('Epoch ' + str(Epochs) + ' PerEpochCounter ' + str(PerEpochCounter))
Timer2 = tic()
IBatch, LabelBatch, I1Batch, I2Batch, I1PatchBatch, I2PatchBatch, \
AllPtsBatch, PerturbPtsBatch, MaskBatch = GenerateBatch(TrainNames, ImageSize, MiniBatchSize, Rho, BasePath)
FeedDict = {ImgPH: IBatch, I1PH: I1Batch, AllPtsPH: AllPtsBatch, I2PH: I2Batch, MaskPH: MaskBatch, LabelPH: LabelBatch}
_, LossThisBatch, Summary = sess.run([OptimizerUpdate, loss, MergedSummaryOP], feed_dict=FeedDict)
# Tensorboard
Writer.add_summary(Summary, Epochs*NumIterationsPerEpoch + PerEpochCounter)
# If you don't flush the tensorboard doesn't update until a lot of iterations!
Writer.flush()
# Calculate and print Train accuracy (also called EpochLoss) every epoch
EpochLoss += LossThisBatch
# Save All losses
EachIterLoss.append(LossThisBatch)
TimeLastMiniBatch = toc(Timer2)
# Print LossThisBatch
print('LossThisBatch is '+ str(LossThisBatch))
# Save checkpoint every some SaveCheckPoint's iterations
if PerEpochCounter % SaveCheckPoint == 0:
# Save the Model learnt in this epoch
SaveName = CheckPointPath + str(Epochs) + 'a' + str(PerEpochCounter) + 'model.ckpt'
Saver.save(sess, save_path=SaveName)
print(SaveName + ' Model Saved...')
# Print timing information
EstimatedTimeToCompletionThisEpoch = float(TimeLastMiniBatch)*float(NumIterationsPerEpoch-PerEpochCounter-1.0)
EstimatedTimeToCompletionTotal = float(TimeLastMiniBatch)*float(NumIterationsPerEpoch-PerEpochCounter-1.0) +\
float(TimeLastMiniBatch)*float(NumIterationsPerEpoch-1.0)*float(NumEpochs-Epochs)
TotalTimeElapsed = toc(TimerOverall)
print('Percentage complete in total epochs ' + str(float(Epochs+1)/float(NumEpochs-StartEpoch+1)*100.0))
print('Percentage complete in this Train epoch ' + str(float(PerEpochCounter)/float(NumIterationsPerEpoch)*100.0))
print('Last MiniBatch took '+ str(TimeLastMiniBatch) + ' secs, time taken till now ' + str(TotalTimeElapsed) + \
' estimated time to completion of this epoch is ' + str(EstimatedTimeToCompletionThisEpoch))
print('Estimated Total time remaining is ' + str(EstimatedTimeToCompletionTotal))
TimeLastEpoch = toc(Timer1)
EstimatedTimeToCompletion = float(TotalTimeElapsed)/float(Epochs+1.0)*float(NumEpochs-Epochs-1.0)
# Save Each Epoch loss
AllEpochLoss.append(EpochLoss)
# Save model every epoch
SaveName = CheckPointPath + str(Epochs) + 'model.ckpt'
Saver.save(sess, save_path=SaveName)
print(SaveName + ' Model Saved...')
def main():
"""
Inputs:
None
Outputs:
Runs the Training and testing code based on the Flag
"""
# TODO: Make LogDir
# TODO: Make logging file a parameter
# TODO: Time to complete print
# Parse Command Line arguments
Parser = argparse.ArgumentParser()
Parser.add_argument('--BasePath', default='/media/nitin/Research/EVDodge/downfacing_processed', help='Base path of images, Default:/media/nitin/Research/EVDodge/downfacing_processed')
Parser.add_argument('--NumEpochs', type=int, default=200, help='Number of Epochs to Train for, Default:200')
Parser.add_argument('--DivTrain', type=int, default=1, help='Factor to reduce Train data by per epoch, Default:1')
Parser.add_argument('--MiniBatchSize', type=int, default=256, help='Size of the MiniBatch to use, Default:256')
Parser.add_argument('--LoadCheckPoint', type=int, default=0, help='Load Model from latest Checkpoint from CheckPointsPath?, Default:0')
Parser.add_argument('--CheckPointPath', default='../CheckpointsHomography/', help='Path to save checkpoints, Default:../CheckpointsHomography/')
Parser.add_argument('--LogsPath', default='/media/nitin/Research/EVDodge/Logs/', help='Path to save Logs, Default:/media/nitin/Research/EVDodge/Logs/')
Parser.add_argument('--LossFuncName', default='M', help='Choice of Loss functions, choose from M for Mean, V for Variance, Default:M')
Parser.add_argument('--GPUDevice', type=int, default=0, help='What GPU do you want to use? -1 for CPU, Default:0')
Parser.add_argument('--LR', type=float, default=1e-4, help='Learning Rate, Default: 1e-4')
Parser.add_argument('--SymType', default='L1', help='Similarity mapping, choose from L1 and Chab, Default:L1')
Args = Parser.parse_args()
NumEpochs = Args.NumEpochs
BasePath = Args.BasePath
DivTrain = float(Args.DivTrain)
MiniBatchSize = Args.MiniBatchSize
LoadCheckPoint = Args.LoadCheckPoint
CheckPointPath = Args.CheckPointPath
LogsPath = Args.LogsPath
LossFuncName = Args.LossFuncName
GPUDevice = Args.GPUDevice
LearningRate = Args.LR
SymType = Args.SymType
# Set GPUDevice
tu.SetGPU(GPUDevice)
# Setup all needed parameters including file reading
TrainNames, ValNames, TestNames, OptimizerParams,\
SaveCheckPoint, ImageSize, PatchSize, NumTrainSamples, NumValSamples, NumTestSamples,\
NumTestRunsPerEpoch = SetupAll(BasePath, LearningRate)
# Find Latest Checkpoint File
if LoadCheckPoint==1:
LatestFile = FindLatestModel(CheckPointPath)
else:
LatestFile = None
# Pretty print stats
PrettyPrint(NumEpochs, DivTrain, MiniBatchSize, NumTrainSamples, NumTestSamples, LatestFile)
# Define PlaceHolder variables for Input and Predicted output
PatchPH = tf.placeholder(tf.float32, shape=(MiniBatchSize, PatchSize[0], PatchSize[1], PatchSize[2]), name='Input')
# PH for losses
IPH = tf.placeholder(tf.float32, shape=(MiniBatchSize, ImageSize[0], ImageSize[1], ImageSize[2]), name='IPH')
MaskPH = tf.placeholder(tf.float32, shape=(MiniBatchSize, ImageSize[0], ImageSize[1], ImageSize[2]), name='Mask')
TrainOperation(PatchPH, IPH, MaskPH, TrainNames, TestNames, NumTrainSamples, ImageSize, PatchSize,
NumEpochs, MiniBatchSize, OptimizerParams, SaveCheckPoint, CheckPointPath, NumTestRunsPerEpoch,
DivTrain, LatestFile, LossFuncName, BasePath, LogsPath, SymType)
def main():
"""
Inputs:
None
Outputs:
Runs Testing code
"""
# TODO: Make LogDir
# TODO: Display time to end and cleanup other print statements with color
# TODO: Make logging file a parameter
# Parse Command Line arguments
Parser = argparse.ArgumentParser()
Parser.add_argument('--ModelPath', dest='ModelPath', default='/media/nitin/Research/EVDodge/CheckpointsDeblurHomographyLR1e-4Epochs400/399model.ckpt',\
help='Path to load latest model from, Default:ModelPath')
Parser.add_argument('--ReadPath', dest='ReadPath', default='/media/nitin/Research/EVDodge/DatasetChethanEvents/Deblurred',\
help='Path to load images from, Default:ReadPath')
Parser.add_argument('--WritePath', dest='WritePath', default='/media/nitin/Research/EVDodge/DatasetChethanEvents/DeblurredHomography',\
help='Path to load images from, Default:WritePath')
Parser.add_argument(
'--GPUDevice',
type=int,
default=0,
help='What GPU do you want to use? -1 for CPU, Default:0')
Parser.add_argument(
'--CropType',
dest='CropType',
default='C',
help=
'What kind of crop do you want to perform? R: Random, C: Center, Default: C'
)
Args = Parser.parse_args()
ModelPath = Args.ModelPath
ReadPath = Args.ReadPath
WritePath = Args.WritePath
GPUDevice = Args.GPUDevice
CropType = Args.CropType
# Set GPUNum
tu.SetGPU(GPUDevice)
# Setup all needed parameters including file reading
TrainNames, ImageSize, PatchSize, NumTrainSamples = SetupAll(ReadPath)
# Define PlaceHolder variables for Input and Predicted output
PatchPH = tf.placeholder(tf.float32,
shape=(1, PatchSize[0], PatchSize[1],
PatchSize[2] * 2),
name='Input')
I1PH = tf.placeholder(tf.float32,
shape=(1, PatchSize[0], PatchSize[1], PatchSize[2]),
name='I1')
I2PH = tf.placeholder(tf.float32,
shape=(1, PatchSize[0], PatchSize[1], PatchSize[2]),
name='I2')
if (not os.path.exists(WritePath)):
cprint("WARNING: %s doesnt exist, Creating it." % WritePath, 'yellow')
os.mkdir(WritePath)
TestOperation(PatchPH, I1PH, I2PH, PatchSize, ModelPath, ReadPath,
WritePath, TrainNames, NumTrainSamples, CropType)
def TestOperation(PatchPH, I1PH, I2PH, PatchSize, ModelPath, ReadPath,
WritePath, TrainNames, NumTrainSamples, CropType):
"""
Inputs:
ImgPH is the Input Image placeholder
HomingVecPH is the ground truth homing vector placeholder
NumTrainSamples - length(Train)
ModelPath - Path to load trained model from
DataPath - Paths of all images where testing will be run on
Outputs:
Predictions written to ./TxtFiles/PredOut.txt
"""
# Generate indexes for center crop of train size
CenterX = PatchSize[1] / 2
CenterY = PatchSize[0] / 2
RandX = np.ceil(CenterX - PatchSize[1] / 2)
RandY = np.ceil(CenterY - PatchSize[0] / 2)
p1 = (RandX, RandY)
p2 = (RandX, RandY + PatchSize[0])
p3 = (RandX + PatchSize[1], RandY + PatchSize[0])
p4 = (RandX + PatchSize[1], RandY)
AllPts = [p1, p2, p3, p4]
# Predict output with forward pass, MiniBatchSize for Test is 1
prHVal = EVHomographyNetUnsupSmall(PatchPH, PatchSize, 1)
prHVal = tf.reshape(prHVal, (-1, 8, 1))
HMat = stn.solve_DLT(1, AllPts, prHVal)
# Warp I1 to I2
out_size = [128, 128]
WarpI1 = stn.transform(out_size, HMat, 1, I1PH)
# Setup Saver
Saver = tf.train.Saver()
with tf.Session() as sess:
Saver.restore(sess, ModelPath)
tu.FindNumParams(1)
PredOuts = open(WritePath + os.sep + 'PredOuts.txt', 'w')
for dirs in tqdm(next(os.walk(ReadPath))[1]):
CurrReadPath = ReadPath + os.sep + dirs
NumFilesInCurrDir = len(glob.glob(CurrReadPath + os.sep + '*.png'))
I = cv2.imread(CurrReadPath + os.sep + "event_%d" % 1 + '.png')
CurrWritePath = WritePath + os.sep + dirs
# Create Write Folder if doesn't exist
if (not os.path.exists(CurrWritePath)):
os.makedirs(CurrWritePath)
for ImgNum in tqdm(range(0, NumFilesInCurrDir)):
# for StackNum in range(0, NumImgsStack):
INow = cv2.imread(CurrReadPath + os.sep + "event_%d" %
(ImgNum + 1) + '.png')
Rho = [25] # [10, 10]
IBatch, I1Batch, I2Batch, AllPtsBatch, PerturbPtsBatch, H4PtColBatch, MaskBatch = GenerateBatch(
INow, Rho, PatchSize, CropType, Vis=False)
# TODO: Better way is to feed data into a MiniBatch and Extract it again
# INow = np.hsplit(INow, 2)[0] # Imgs have a stack of 2 in this case, hence extract one
FeedDict = {PatchPH: IBatch, I1PH: I1Batch, I2PH: I2Batch}
prHPredVal = sess.run(prHVal, FeedDict)
prHTrue = np.float32(np.reshape(H4PtColBatch[0],
(-1, 4, 2)))[0]
ErrorNow = np.sum(
np.sqrt((prHPredVal[:, 0] - prHTrue[:, 0])**2 +
(prHPredVal[:, 1] - prHTrue[:, 1])**2)) / 4
# print(ErrorNow)
# print(prHPredVal)
# print(prHTrue)
# a = input('a')
# Timer1 = tic()
# WarpI1Ret = sess.run(WarpI1, FeedDict)
# cv2.imshow('a', WarpI1Ret[0])
# cv2.imshow('b', I1Batch[0])
# cv2.imshow('c', I2Batch[0])
# cv2.imshow('d', np.abs(WarpI1Ret[0]- I2Batch[0]))
# cv2.waitKey(0)
# print(toc(Timer1))
# WarpI1Ret = WarpI1Ret[0]
# Remap to [0,255] range
# WarpI1Ret = np.uint8(remap(WarpI1Ret, 0.0, 255.0, np.amin(WarpI1Ret), np.amax(WarpI1Ret)))
# Crop out junk pixels as they are appended in top left corner due to padding
# WarpI1Ret = WarpI1Ret[-PatchSize[0]:, -PatchSize[1]:, :]
# IStacked = np.hstack((WarpI1Ret, I2Batch[0]))
# Write Image to file
# cv2.imwrite(CurrWritePath + os.sep + 'events' + os.sep + "event_%d"%(ImgNum+1) + '.png', IStacked)
PredOuts.write(dirs + os.sep + "event_%d" % ImgNum + '.png' +
'\t' + str(ErrorNow) + '\n')
PredOuts.close()
def main():
"""
Inputs:
None
Outputs:
Runs the Training and testing code based on the Flag
"""
# TODO: Make LogDir
# TODO: Make logging file a parameter
# TODO: Time to complete print
# Parse Command Line arguments
Parser = argparse.ArgumentParser()
Parser.add_argument(
'--BasePath',
default=
'/media/nitin/Research/EVDodge/DatasetChethanEvents/DeblurredHomography',
help=
'Base path of images, Default:/media/nitin/Research/EVDodge/processed')
Parser.add_argument('--NumEpochs',
type=int,
default=200,
help='Number of Epochs to Train for, Default:200')
Parser.add_argument(
'--DivTrain',
type=int,
default=1,
help='Factor to reduce Train data by per epoch, Default:1')
Parser.add_argument('--MiniBatchSize',
type=int,
default=256,
help='Size of the MiniBatch to use, Default:256')
Parser.add_argument(
'--LoadCheckPoint',
type=int,
default=0,
help=
'Load Model from latest Checkpoint from CheckPointsPath?, Default:0')
Parser.add_argument(
'--GPUDevice',
type=int,
default=0,
help='What GPU do you want to use? -1 for CPU, Default:0')
Parser.add_argument(
'--CheckPointPath',
default='../CheckpointsSeg/',
help='Path to save checkpoints, Default:../CheckpointsSeg/')
Parser.add_argument(
'--LogsPath',
default='/media/nitin/Research/EVDodge/LogsSeg/',
help='Path to save Logs, Default:/media/nitin/Research/EVDodge/LogsSeg/'
)
Parser.add_argument('--LR',
type=float,
default=1e-3,
help='Learning Rate, Default: 1e-3')
Parser.add_argument(
'--MaxFrameDiff',
type=int,
default=1,
help='Maximum Frame difference to feed into network, Default: 1')
Args = Parser.parse_args()
NumEpochs = Args.NumEpochs
BasePath = Args.BasePath
DivTrain = float(Args.DivTrain)
MiniBatchSize = Args.MiniBatchSize
LoadCheckPoint = Args.LoadCheckPoint
GPUDevice = Args.GPUDevice
CheckPointPath = Args.CheckPointPath
LogsPath = Args.LogsPath
LearningRate = Args.LR
MaxFrameDiff = Args.MaxFrameDiff
# Set GPUDevice
tu.SetGPU(GPUDevice)
# If CheckPointPath doesn't exist make the path
if (not (os.path.isdir(CheckPointPath))):
os.makedirs(CheckPointPath)
# If LogsPath doesn't exist make the path
if (not (os.path.isdir(LogsPath))):
os.makedirs(LogsPath)
# Setup all needed parameters including file reading
TrainNames, TrainLabels, OptimizerParams, SaveCheckPoint, Factor, ImageSize, NumTrainSamples = SetupAll(
BasePath, LearningRate)
# Find Latest Checkpoint File
if LoadCheckPoint == 1:
LatestFile = FindLatestModel(CheckPointPath)
else:
LatestFile = None
# Pretty print stats
PrettyPrint(NumEpochs, DivTrain, MiniBatchSize, NumTrainSamples,
LatestFile)
# Define PlaceHolder variables for Input and Predicted output
ImgPH = tf.placeholder(tf.float32,
shape=(MiniBatchSize, ImageSize[0], ImageSize[1],
6),
name='Input')
I1PH = tf.placeholder(tf.float32,
shape=(MiniBatchSize, ImageSize[0], ImageSize[1], 3),
name='I1')
I2PH = tf.placeholder(tf.float32,
shape=(MiniBatchSize, ImageSize[0], ImageSize[1], 3),
name='I2')
LabelPH = tf.placeholder(tf.float32,
shape=(MiniBatchSize, ImageSize[0], ImageSize[1],
2),
name='Label') # 2 classes
TrainOperation(ImgPH, I1PH, I2PH, LabelPH, TrainNames, TrainLabels,
NumTrainSamples, Factor, ImageSize, NumEpochs,
MiniBatchSize, OptimizerParams, SaveCheckPoint,
CheckPointPath, DivTrain, LatestFile, MaxFrameDiff,
LogsPath, BasePath)