# Epipolar data - Translation
epi_trans = np.empty((0, 3))
for cam in usedCams:
epi_trans_cam = getH5pyData(config,
epiFilesDict,
usedSeqs,
'epi_trans_xyz',
3,
camera=cam)
epi_trans = np.append(epi_trans, epi_trans_cam, axis=0)
epi_trans_parms = calcNormParms(epi_trans, trainIdxs)
# Save to File
normParmsFile = config.getOutputFiles(normParmsFilesDict, recalcDataNorm)
if normParmsFile is None:
sys.exit()
with h5py.File(normParmsFile, 'w') as f:
f.create_dataset('rot_xyz', data=rot_parms)
f.create_dataset('trans_xyz', data=trans_parms)
f.create_dataset('trans_rtp', data=polar_parms)
f.create_dataset('noisy_rot_xyz', data=imu_parms)
f.create_dataset('epi_rot_xyz', data=epi_rot_parms)
f.create_dataset('epi_trans_xyz', data=epi_trans_parms)
print('^ Normalization Parameters Calculated')
print()
print()
print('v Standardizing Images')
for cam in usedCams:
for seqStr in kittiSeqs:
seq = int(seqStr)
print('%sReading Camera %s, Sequence %02d' % (' ' * 2, cam, seq))
calFile = config.getInputFiles(calFilesDict, seq)
if (calFile is None):
continue
origImages = config.getInputFiles(origImageFilesDict, seq, cam)
if (origImages is None):
continue
standardImages = config.getOutputFiles(standardImageFilesDict,
recalcStandardImages, seq, cam)
if (standardImages is None):
continue
standardImageFolder = config.getFolderRef(standardImages)
# get cal matrix for sequence
cameraCalMat = getCalMat(calFile, cam)
map_x, map_y = getMaps(destImageShape, standardDict['overlap'],
cameraCalMat)
numImages = len(origImages)
for idx, origImageName in enumerate(origImages):
if showProgress:
percentComplete = int(idx / numImages * 100)
for cam in usedCams:
for seqStr in kittiSeqs:
seq = int(seqStr)
print('%sReading Camera %s, Sequence %02d' % (' ' * 2, cam, seq))
# Get image names
origImageNames = config.getInputFiles(origImageFilesDict, seq, cam)
# if (origImageNames is None):
# continue
# get calibration files
calFile = config.getInputFiles(calFilesDict, seq)
# if (calFile is None):
# continue
epiFile = config.getOutputFiles(epiFilesDict, recalcEpiData, seq, cam)
# if epiFile is None:
# continue
# get calibration matrix
cameraCalMat = getCalMat(calFile, cam)
epi_rot_xyz = np.empty((0, 3))
epi_trans_xyz = np.empty((0, 3))
numImagePairs = len(origImageNames) - 1
# get first image keypoints and descriptors
srcKp, srcDesc = getKpsAndDescs(SIFT_detector, origImageNames[0])
for idx, origImageName in enumerate(origImageNames[1:]):
norm_epi_rot = applyNorm(epi_rot, epi_rot_parms)
# Epipolar data - Translation
epi_trans = np.empty((0, 3))
for cam in usedCams:
epi_trans_cam = getH5pyData(config,
epiFilesDict,
usedSeqs,
'epi_trans_xyz',
3,
camera=cam)
epi_trans = np.append(epi_trans, epi_trans_cam, axis=0)
norm_epi_trans = applyNorm(epi_trans, epi_trans_parms)
# Save to File
normDataFile = config.getOutputFiles(normDataFilesDict, reapplyDataNorm)
if normDataFile is not None:
print('Saving Normalized Truth and IMU Data')
with h5py.File(normDataFile, 'w') as f:
f.create_dataset('rot_xyz', data=norm_truth_rot)
f.create_dataset('trans_xyz', data=norm_truth_xyz)
f.create_dataset('trans_rtp', data=norm_truth_polar)
f.create_dataset('noisy_rot_xyz', data=norm_imu_rot)
normEpiFile = config.getOutputFiles(normEpiFilesDict, reapplyDataNorm)
if normEpiFile is not None:
print('Saving Normalized Epipolar Data')
with h5py.File(normEpiFile, 'w') as f:
f.create_dataset('epi_rot_xyz', data=norm_epi_rot)
f.create_dataset('epi_trans_xyz', data=norm_epi_trans)
arwError**2 * 1 / 3600 * 1 / sampleRate) * np.pi / 180 # radians
# sqrt( deg^2/hr * hr/sec * sec ) * radians/deg
# VALUE IS APPROXIMATELY 0.002635 deg
print()
print('v Creating IMU Data')
for cam in usedCams:
for seqStr in kittiSeqs:
seq = int(seqStr)
print('%sReading Camera %s, Sequence %02d' % (' ' * 2, cam, seq))
truthFile = config.getInputFiles(truthFilesDict, seq)
if (truthFile is None):
continue
imuFile = config.getOutputFiles(imuFilesDict, recalcIMUData, seq)
if (imuFile is None):
continue
# Read in rotation truth poses
with h5py.File(truthFile, 'r') as f:
rot_xyz = np.array(f['rot_xyz']) # radians
# Add noise to rotation piece
noise_xyz = np.random.randn(rot_xyz.shape[0], rot_xyz.shape[1])
noisy_rot_xyz = rot_xyz + angular_std * noise_xyz
# Save noisy rotation angles to h5py file
with h5py.File(imuFile, 'w') as f:
os.chmod(imuFile, 0o666)
f.create_dataset('noisy_rot_xyz', data=noisy_rot_xyz)
splitEpiLayers = config.constraintParms['splitEpiLayers'] postConstraintsLayerSize = config.constraintParms['postConstraintsLayerSize'] defaultLossFunc = config.modelParms['defaultLossFunction'] lossRotScale = config.modelParms['lossRotScale'] optimizerType = config.modelParms['optimizerType'] model_yaml_filename = config.modelParms['modelYamlFilename'] # Callbacks checkpoint_filename = config.checkpointParms['filename'] history_filename = config.trainingParms['histFilename'] # Get Files modelYamlsDict = config.trainPaths['model'] modelYamls = config.getOutputFiles(modelYamlsDict, True) saveModelPath = config.getFolderRef(modelYamls) model_filepath = os.path.join(saveModelPath, model_yaml_filename) checkpointFilesDict = config.trainPaths['checkpoint'] checkpointFiles = config.getOutputFiles(checkpointFilesDict, True) saveCheckpointPath = config.getFolderRef(checkpointFiles) checkpoint_filepath = os.path.join(saveCheckpointPath, checkpoint_filename) historyFilesDict = config.trainPaths['history'] historyFiles = config.getOutputFiles(historyFilesDict, True) saveHistoryPath = config.getFolderRef(historyFiles) history_filepath = os.path.join(saveHistoryPath, history_filename) # Get Callbacks List callbacksList = getCallbacksList(config, history_filepath, checkpoint_filepath)
print()
print('v Downsampling Images and Normalizing Pixels')
for cam in usedCams:
for seqStr in kittiSeqs:
seq = int(seqStr)
print('%sReading Camera %s, Sequence %02d' % (' ' * 2, cam, seq))
standardImages = config.getInputFiles(standardImageFilesDict, seq, cam)
if (standardImages is None):
continue
normImages = config.getOutputFiles(normImageFilesDict, recalcNormImages, seq, cam)
if (normImages is None):
continue
normImageFolder = config.getFolderRef(normImages)
numImages = len(standardImages)
for idx, standardImageName in enumerate(standardImages):
if showProgress:
percentComplete = int(idx/numImages*100)
if divmod(idx,300)[1]==0:
print('Percent Complete: %d%%' % percentComplete)
srcImage = cv.cvtColor(cv.imread(standardImageName), cv.COLOR_BGR2GRAY)
dstImage = cv.resize(srcImage, dsize=destImageShape[::-1], interpolation=cv.INTER_AREA)
# Files
poseFilesDict = config.kittiOriginal['truth']
truthFilesDict = config.kittiPrepared['truth']
print()
print('v Creating Truth Data')
for cam in usedCams:
for seqStr in kittiSeqs:
seq = int(seqStr)
print('%sReading Camera %s, Sequence %02d' % (' ' * 2, cam, seq))
poseFile = config.getInputFiles(poseFilesDict, seq)
if (poseFile is None):
continue
truthFile = config.getOutputFiles(truthFilesDict, recalcTruthData, seq)
if (truthFile is None):
continue
# Get true R and t from pose files
# Row i represents the ith pose of the left camera coordinate
# system (z pointing forwards) via a 3x4 transformation matrix.
# The matricies take a point in the ith coordinate system and project
# it to the 0th coordinate system. The translational part corresponds
# to the pose of the left camera in the ith frame with respect to the
# 0th frame.
# T_0_i = [ R_0_i , t_0_0toi ]
# [ 0_(1x3), 1 ]
def loadPredictions(tests, eType):
errors_list = []
preds_list = []
truth_list = []
filetruthrots_list = []
for i, test in enumerate(tests):
# if True:
# test = tests[0]
configFile = 'exp_configs/%s.yaml' % test
config = ThesisConfig(configFile)
# Parameters
name = config.expName
numOutputs = config.modelParms['numOutputs']
# Get Files
evalFilesDict = config.resultPaths['evaluations']
figFilesDict = config.resultPaths['figures']
truthFilesDict = config.kittiPrepared['truth']
figFolder = config.getFolderRef(
config.getOutputFiles(figFilesDict, True))
# Get Predictions Save File
evalFolder = ''
for pathSection in evalFilesDict['dir']:
evalFolder = os.path.join(evalFolder, pathSection)
predictionsFile = os.path.join(evalFolder, eType + '_predictions.hdf5')
if os.path.exists(predictionsFile):
# Load Predictions
y_pred_real, y_true_real, evalType, min_val_epoch, min_val_loss = loadPredFile(
predictionsFile)
print('Min Validation Loss: %s, Epoch %s' %
(min_val_loss, min_val_epoch))
# get test idxs
# numTested = y_true_real.shape[0]
splitFilesDict = config.kittiPrepared['split']
splitFile = config.getInputFiles(splitFilesDict)
with h5py.File(splitFile, 'r') as f:
if evalType == 'test':
turnIdxs = np.array(f['testTurnIdxs'])
nonTurnIdxs = np.array(f['testNonTurnIdxs'])
elif evalType == 'val':
turnIdxs = np.array(f['valTurnIdxs'])
nonTurnIdxs = np.array(f['valNonTurnIdxs'])
idxs = np.sort(np.concatenate(
(turnIdxs, nonTurnIdxs))) #[:numTested]
truthData = np.empty((0, 3))
for seq in config.usedSeqs:
truthFile = config.getInputFiles(truthFilesDict, seq)
with h5py.File(truthFile, 'r') as f:
rot_xyz = np.array(f['rot_xyz'])
truthData = np.append(truthData, rot_xyz, axis=0)
file_truth_rots = truthData[idxs, :]
# Calculate average loss in each direction
errors = y_true_real - y_pred_real
errors_list.append(errors)
preds_list.append(y_pred_real)
truth_list.append(y_true_real)
filetruthrots_list.append(file_truth_rots)
else:
print('predictions file %s does not exist' % predictionsFile)
return (errors_list, preds_list, truth_list, filetruthrots_list)
for configFile in configFiles:
config = ThesisConfig(configFile)
# Parameters
name = config.expName
numOutputs = config.modelParms['numOutputs']
# Get Files
evalFilesDict = config.resultPaths['evaluations']
figFilesDict = config.resultPaths['figures']
figFolder = config.getFolderRef(config.getOutputFiles(figFilesDict, True))
# Get Predictions Save File
evalFolder = ''
for pathSection in evalFilesDict['dir']:
evalFolder = os.path.join(evalFolder, pathSection)
predictionsFile = os.path.join(evalFolder, evalType+'_predictions.hdf5')
if os.path.exists(predictionsFile):
# Load Predictions
with h5py.File(predictionsFile, 'r') as f:
from src.helpers.cfg import ThesisConfig
from src.helpers.helper_functions import getOptimizer
from src.helpers.training_helpers import getCallbacksList, getGenerator, getTrainAndValGenerators
from src.helpers.custom_loss import scaledMSE_RT
import tensorflow as tf
from keras import backend as K
import keras.losses
# Load Configuration
configFile = 'exp_configs/scale_test_4.yaml'
config = ThesisConfig(configFile)
checkpoint_filename = config.checkpointParms['filename']
checkpointFilesDict = config.trainPaths['checkpoint']
checkpointFiles = config.getOutputFiles(checkpointFilesDict, True)
saveCheckpointPath = config.getFolderRef(checkpointFiles)
checkpoint_filepath = os.path.join(saveCheckpointPath, checkpoint_filename)
numOutputs = config.modelParms['numOutputs']
lossRotScale = config.modelParms['lossRotScale']
defaultLossFunc = config.modelParms['defaultLossFunction']
if numOutputs > 3:
lossFunc = scaledMSE_RT(lossRotScale)
keras.losses.lossFunction = lossFunc
else:
lossFunc = defaultLossFunc
modelPath = checkpoint_filepath.format(epoch=1)
model = load_model(modelPath)
# Parameters
kittiSeqs = config.kittiSeqs
usedCams = config.usedCams
usedSeqs = config.usedSeqs
splitFracs = config.splitFracs
turnThresh_rad = config.thesisKittiParms['turnThreshold'] * np.pi / 180
# Files
truthFilesDict = config.kittiPrepared['truth']
splitFilesDict = config.kittiPrepared['split']
print()
print('v Splitting Dataset')
splitFile = config.getOutputFiles(splitFilesDict)
if splitFile is None:
print('^ Dataset Splitting Complete')
print()
sys.exit()
# Check if file exists
# Get Sequences Counts and Turning Indexes
truthFilesList = []
for seq in usedSeqs:
truthFilesList.append(config.getInputFiles(truthFilesDict, seq))
idxCvt = IndexConverter(truthFilesList, turnThresh_rad)
# SPLIT TURNING INDEXES
# 'exp_configs/CNN_test_15.yaml',
# 'exp_configs/CNN_test_16.yaml',
# 'exp_configs/CNN_test_17.yaml']
configFileList = ['exp_configs/scale_test_3.yaml']
for configFile in configFileList:
config = ThesisConfig(configFile)
history_filename = config.trainingParms['histFilename']
historyFilesDict = config.trainPaths['history']
historyFiles = config.getInputFiles(historyFilesDict)
saveHistoryPath = config.getFolderRef(historyFiles)
history_filepath = os.path.join(saveHistoryPath, history_filename)
saveFigFilesDict = config.resultPaths['figures']
figureFile = config.getOutputFiles(saveFigFilesDict, True)
saveFigFolder = config.getFolderRef(figureFile)
saveFigFilename = config.experiment['experiment'][
'name'] + config.resultPaths['figures']['type']
saveFigFile = os.path.join(saveFigFolder, saveFigFilename)
if os.path.exists(history_filepath):
print()
with h5py.File(history_filepath, 'r') as f:
epochs = np.array(f['epochs'], dtype=np.int)
numEpochs = len(epochs)
if 'loss' in f:
train_loss = np.array(f['loss'])
min_train_loss = np.min(train_loss)
min_train_loss_epoch = (epochs[np.argmin(train_loss)])
print('Min Training Loss: %s, Epoch %s' %