def initFromConfig(self, cfg, sectionKey):
super(ConvLayerParams, self).initFromConfig(cfg, sectionKey)
self._inputDim = readCfgIntNoneListParam(cfg, sectionKey, 'inputDim',
self._inputDim)
self._nFilters = readCfgIntParam(cfg, sectionKey, 'nFilters',
self._nFilters)
self._filterDim = readCfgIntNoneListParam(cfg, sectionKey, 'filterDim',
self._filterDim)
#self._filter_shape = filter_shape
#self._image_shape = image_shape
#self._outputDim = readCfgIntNoneListParam(cfg,sectionKey,'outputDim',self._outputDim)
self._wInitMode = readCfgIntParam(cfg, sectionKey, 'wInitMode',
self._wInitMode)
self._wInitOrthogonal = readCfgBooleanParam(cfg, sectionKey,
'wInitOrthogonal',
self._wInitOrthogonal)
self.update()
def initFromConfig(self, cfg, sectionKey):
super(PoolLayerParams, self).initFromConfig(cfg, sectionKey)
self._inputDim = readCfgIntNoneListParam(cfg, sectionKey, 'inputDim',
self._inputDim)
self._poolType = readCfgIntParam(cfg, sectionKey, 'poolType',
self._poolType)
self._poolsize = readCfgIntNoneListParam(cfg, sectionKey, 'poolsize',
self._poolsize)
self.update()
def initFromConfig(self, cfg, sectionKey):
super(ConvPoolLayerParams, self).initFromConfig(cfg, sectionKey)
self._inputDim = readCfgIntNoneListParam(cfg, sectionKey, 'inputDim',
self._inputDim)
self._nFilters = readCfgIntParam(cfg, sectionKey, 'nFilters',
self._nFilters)
self._filterDim = readCfgIntNoneListParam(cfg, sectionKey, 'filterDim',
self._filterDim)
self._wInitMode = readCfgIntParam(cfg, sectionKey, 'wInitMode',
self._wInitMode)
self._wInitOrthogonal = readCfgBooleanParam(cfg, sectionKey,
'wInitOrthogonal',
self._wInitOrthogonal)
self._poolType = readCfgIntParam(cfg, sectionKey, 'poolType',
self._poolType)
self._poolsize = readCfgIntNoneListParam(cfg, sectionKey, 'poolsize',
self._poolsize)
self.update()
def initFromConfig(self,cfg,sectionKey):
super(CatLayerParams,self).initFromConfig(cfg,sectionKey)
self._inputDim = readCfgIntNoneListParam(cfg,sectionKey,'inputDim',self._inputDim)
self._axis = readCfgIntParam(cfg,sectionKey,'axis',self._axis)
self.update()
def initFromConfig(self,cfg,sectionKey):
super(HiddenLayerParams,self).initFromConfig(cfg,sectionKey)
self._nOutputs = readCfgIntParam(cfg,sectionKey,'outputDim',self._nOutputs)
self.update()
def linemod_test_main(fileName=None, cfg=None):
theano.config.exception_verbosity = 'high'
if fileName is None:
g = glob.glob('../data/results/tmp/*_descrNet.pkl')
#g = glob.glob('../data/results/cuda03_tmp/*_descrNet.pkl')
fileName = sorted(g)[-1]
if os.path.isdir(fileName):
g = glob.glob('{}/*_descrNet.pkl'.format(fileName))
fileName = sorted(g)[-1]
outFNBase = fileName[:-len("descrNet.pkl")] + 'test'
print("outFNBase {}".format(outFNBase))
showPlots = readCfgIntParam(cfg, 'misc', 'showPlots', default=1) > 0
######## load net
trainRes = loadTrainResultsPickle(fileName)
descrNet = trainRes[0]
cfg = trainRes[2]
descrNetTrainerParams = trainRes[3]
#
imgsPklFileName = readCfgParam(cfg, 'input', 'imgsPklFileName', default="")
trainSetPklFileName = readCfgParam(cfg,
'train',
'trainSetPklFileName',
default="")
valSetPklFileName = readCfgParam(cfg,
'train',
'valSetPklFileName',
default="")
#-------------
inputDim = descrNet.cfgParams._inputDim
#if len(inputDim) > 1 and len(inputDim[0]) > 1:
if isinstance(inputDim[0], collections.Iterable):
inputMode = 3 # IMG/DPT
inputSize = inputDim[0][2]
# check
heights = [idim[2] for idim in inputDim]
widths = [idim[3] for idim in inputDim]
assert all([w == inputSize for w in widths]) and all(
[h == inputSize
for h in heights]), "can only handle quadratic inputs"
else:
nChan = inputDim[1]
if nChan == 1:
inputMode = 0 # DPT
elif nChan == 3:
inputMode = 1 # IMG
elif nChan == 4:
inputMode = 2 # IMG + DPT
else:
raise ValueError(
"Unknown input mode: #channels = {}".format(nChan))
inputSize = inputDim[2]
print("inputMode = {}".format(inputMode))
print("inputSize = {}".format(inputSize))
print("num weight is descrNet: {}".format(descrNet.getNumWeights()))
print(" total: {}".format(
numpy.sum([numpy.prod(ws) for ws in descrNet.getNumWeights()])))
######## prepare test data
#lmDataBasepath = '/home/wohlhart/work/data/linemod/'
lmDataBasepath = readCfgParam(cfg,
'paths',
'lmDataBase',
default='/home/wohlhart/work/data/linemod/')
seqNamesStr = readCfgParam(cfg, 'input', 'objs', default="")
seqNames = map(str.strip, seqNamesStr.split(","))
zRotInvStr = readCfgParam(cfg, 'input', 'zRotInv', default=[])
zRotInv = list(eval(zRotInvStr))
print "trainset: {}".format(trainSetPklFileName)
print "images: {}".format(imgsPklFileName)
###############################################################################################
# create test set from test seq that was stored by main_linemod_realandsynth
# which did the split of the linemod data into train and test
print("Loading images from pkl ... ")
# LOAD from pkl
pklload_start_time = time.clock()
with open(imgsPklFileName, 'rb') as f:
_, _, testSeqs = cPickle.load(f)
f.close()
pklload_end_time = time.clock()
print('Loading took %.1fs' % ((pklload_end_time - pklload_start_time)))
# patch0 = testSeqs['cat'].data[52]
# print("fn {}".format(patch0.frame.filename))
# print("cn {}".format(patch0.frame.className))
# cv2.imshow("cat52",patch0.dpt/2.+1.)
# cv2.waitKey()
#print("###### bs {}".format(descrNet.cfgParams.batch_size))
testdata_set = BatchedImgSeqDataset()
testdata_set.initFromImgSeqs(testSeqs,
inputMode=inputMode,
batchSize=descrNet.cfgParams.batch_size)
# test it
img0 = numpy.concatenate([x[0]
for x in testdata_set.x], axis=0) if isinstance(
testdata_set.x, list) else testdata_set.x[0]
img0 = numpy.swapaxes(numpy.swapaxes(img0, 0, 1), 1, 2)
print(img0.shape)
if showPlots:
cv2.imshow("img0", img0 + 0.5)
#cv2.waitKey()
# DEBUG: show cam poses
#visBatchedDatasetCamPoses(testdata_set)
#plt.show()
#raise ValueError("stop here")
#############################################################
# Templates dataset
#
tmpl_set = loadTmplSeqsBatchedDataset(
lmDataBasepath,
targetSize=inputSize,
batchSize=descrNet.cfgParams.batch_size,
inputMode=inputMode,
seqNames=seqNames,
zRotInv=zRotInv)
assert isinstance(tmpl_set.x,
list) or (tmpl_set.x.shape[1] == nChan
) # if it's a list, it's more complicated :(
#############################################################
# general informations about tmpl and test datasets
#
uLab = numpy.setdiff1d(numpy.union1d(testdata_set.y, tmpl_set.y),
numpy.array(-1),
assume_unique=True)
print("uLab {}".format(uLab))
numClasses = len(uLab)
testDataValidIdx, = (testdata_set.y >= 0).nonzero()
tmplDataValidIdx, = (tmpl_set.y >= 0).nonzero()
testLabels = testdata_set.y[testDataValidIdx]
tmplLabels = tmpl_set.y[tmplDataValidIdx]
nTestSamp = len(testDataValidIdx)
nTmpls = len(tmplLabels)
numTmplPerClass = nTmpls / numClasses
print("numTmplPerClass {}".format(numTmplPerClass))
print("tmplDataValidIdx {}".format(len(tmplDataValidIdx)))
tmplRots = tmpl_set.sampleInfo['rots'][tmplDataValidIdx]
print("tmplRots.shape {}".format(tmplRots.shape))
tmplRots = tmplRots / numpy.tile(
numpy.linalg.norm(tmplRots, axis=1).reshape(nTmpls, 1),
(1, 3)) # normalize
testRots = testdata_set.sampleInfo['rots'][testDataValidIdx]
testRots = testRots / numpy.tile(
numpy.linalg.norm(testRots, axis=1).reshape(nTestSamp, 1),
(1, 3)) # normalize
print("tmplRots.shape {}, testRots.shape {}".format(
tmplRots.shape, testRots.shape))
#################################################################################
# create a dataManager
dataManagerConfig = MacroBatchManagerConfig()
dataManagerConfig.batch_size = testdata_set.batchSize
macroBatchSize = readCfgIntParam(cfg,
'trainset',
'macroBatchSize',
default=-1)
maxBytes = eval(
readCfgParam(cfg, 'trainset', 'maxBytes', default=50 * 1024 * 1024))
nTestBatches = testdata_set.numSamples / testdata_set.batchSize
nTmplBatches = tmpl_set.numSamples / tmpl_set.batchSize
batchSizeBytes = testdata_set.dataSizeNBytes / nTestBatches
macroBatchSize = findNiceMacroBatchSize(macroBatchSize, maxBytes,
batchSizeBytes,
[nTestBatches, nTmplBatches])
dataManager = MacroBatchManager(dataManagerConfig,
macroBatchSize,
acceptPartialMacroBatches=False)
dataManager.setupVariables(testdata_set, doBorrow=True)
#################################################################################
# visualize testdata and traindata
visTestDataDescrs = True
testOnTrainData = False
if testOnTrainData or visTestDataDescrs:
# LOAD DATA
print("Load train_set: {}".format(trainSetPklFileName))
pklload_start_time = time.clock()
#f = open(trainSetPklFileName, 'rb')
#train_set,val_set = cPickle.load(f)
#f.close()
train_set = BatchedDataset.loadFromPkl(trainSetPklFileName)
#val_set = BatchedDataset.loadFromPkl(valSetPklFileName)
pklload_end_time = time.clock()
print('Loading train_set took %.1fs' %
((pklload_end_time - pklload_start_time)))
if visTestDataDescrs:
calcAndVis3DDescriptorsWithRot(descrNet,
testdata_set,
dataManager,
fileName=outFNBase +
'_testdata_descrs.png')
plt.title('test data')
if testOnTrainData:
#descrNetTrainer.test_on_train()
calcAndVis3DDescriptorsWithRot(descrNet,
train_set,
dataManager,
fileName=outFNBase +
'_traindata_descrs.png')
plt.title('train data')
# test if the classes are mapped correctly by showing one sample of each class from train and test
# for cl in range(4):
# idx = numpy.flatnonzero(testdata_set.y == cl)
# img = testdata_set.x[idx[0]].squeeze()
# cv2.imshow("test{}".format(cl),img)
# idx = numpy.flatnonzero(train_set.y == cl)
# img = train_set.x[idx[0]].squeeze()
# cv2.imshow("train{}".format(cl),img)
#rng = numpy.random.RandomState(23455)
#descrNetTrainer = DescrNetTrainer(descrNet,descrNetTrainerParams,rng)
#descrNetTrainer.setDataAndCompileFunctions(train_set,val_set,compileDebugFcts=False)#True)
#descrNetTrainer.setData(train_set,val_set,macroBatchSize=1)
#descrNetTrainer.test_on_train()
##
# vis cam poses for all data
visCamPoses = False
if visCamPoses and showPlots:
visBatchedDatasetCamPoses(testdata_set)
visBatchedDatasetCamPoses(train_set)
visBatchedTrainDatasetCamPoses(train_set)
plt.show(block=False)
#raise ValueError("stop here")
##############################
# TEST results
descr_comp_start_time = time.clock()
res = descrNet.computeDescriptors(testdata_set,
dataManager,
batch_size=testdata_set.batchSize)
descr_comp_end_time = time.clock()
# throw away results for invalid samples (that were added to fill up minibatches)
testDescrs = res[testDataValidIdx]
print('Computing descriptors for %d samples took %.1fs' %
(testdata_set.numSamples,
(descr_comp_end_time - descr_comp_start_time)))
# calc descrs for templates
#res = descrNet.computeDescriptors(tmpl_set.x,batch_size=tmpl_set.batchSize)
dataManager.prepareForNewData()
res = descrNet.computeDescriptors(tmpl_set,
dataManager,
batch_size=tmpl_set.batchSize)
tmplDescrs = res[tmplDataValidIdx]
print("tmplDescrs.shape {}".format(tmplDescrs.shape))
# distances of descriptors
dst = scipy.spatial.distance.cdist(testDescrs, tmplDescrs)
print("dst.shape {}".format(dst.shape))
# calculate similarity of poses of train sample and templates of the sample class
sims = []
for lab in uLab:
sim = numpy.dot(testRots[testLabels == lab],
tmplRots[tmplLabels == lab].T)
if zRotInv[lab] == 2:
sim = numpy.maximum(
sim,
numpy.dot(
testRots[testLabels == lab] * numpy.array([-1, -1, 1]),
tmplRots[tmplLabels == lab].T))
sims.append(sim + 2.)
sim = scipy.linalg.block_diag(*sims) - 2.
print("sim.shape {}".format(sim.shape))
print("sim mn/mx {},{}".format(numpy.min(sim), numpy.max(sim)))
# show sim matrix
#cv2.imshow("sims",(sim-numpy.min(sim))/(numpy.max(sim) - numpy.min(sim)))
# TODO, eval metrics:
# * pure accuracy: closest tmpl is the correct one (target)
# * cmc curve (target tmpl is among the first k closest)
# * curve: number of samples for which the closest tmpl has a rotation error below x
# make sim block diag
#------
(acc, classAcc) = getAccRel(sim, dst)
print("total test acc: {}%".format(acc * 100.))
print("total test class acc: {}%".format(classAcc * 100.))
# perSampSepScore
perSampSepScore = getClassSepScore(sim, dst)
minSepScore = numpy.min(perSampSepScore)
maxSepScore = numpy.max(perSampSepScore)
meanSepScore = numpy.mean(perSampSepScore)
medSepScore = numpy.median(perSampSepScore)
print("perSampSepScore mn/mx/mean/med {},{},{},{}".format(
minSepScore, maxSepScore, meanSepScore, medSepScore))
tikzFN = outFNBase + '_class_sep.tikz'
pdfFN = outFNBase + '_class_sep.pdf'
pngFN = outFNBase + '_class_sep.png'
pklFN = outFNBase + '_class_sep.pkl'
plotPerSampSepScore(perSampSepScore,
fileName=[pngFN, pdfFN],
tikzFileName=tikzFN,
pklFileName=pklFN)
if showPlots:
plt.show(block=False)
tikzFN = outFNBase + '_acc_vs_angleerror.tikz'
pdfFN = outFNBase + '_acc_vs_angleerror.pdf'
pngFN = outFNBase + '_acc_vs_angleerror.png'
pklFN = outFNBase + '_acc_vs_angleerror.pkl'
print("saving {}".format(pngFN))
plotAngleErrors(sim,
dst,
fileName=[pdfFN, pngFN],
tikzFileName=tikzFN,
pklFileName=pklFN,
showPlot=showPlots)
# # vis min dst and max sim
# minDst = numpy.min(dst,axis=1)
# cv2.imshow("minDst",(dst == minDst).astype(numpy.float))
# maxSim = numpy.min(sim,axis=1)
# cv2.imshow("maxSim",(sim == maxSim).astype(numpy.float))
visClosestTmpls(testdata_set,
tmpl_set,
dst,
fileName=outFNBase + '_samples.png',
showPlot=showPlots)
visWrongestClosestTmpls(testdata_set,
tmpl_set,
dst,
sim,
fileName=outFNBase + '_wrongestsamples.png',
showPlot=showPlots)
##
# scatter all sims vs dsts per class
tikzFN = outFNBase + '_sim_vs_dst.tikz'
#pdfFN = outFNBase+'_sim_vs_dst.pdf'
pngFN = outFNBase + '_sim_vs_dst.png'
#visSimsVsDsts(sim=sim,dst=dst,uLab=uLab,rowLabels=testLabels,colLabels=tmplLabels,classLabels=seqNames,fileName=[pdfFN,pngFN],tikzFileName=tikzFN)#,mxDst=0.7)
visSimsVsDsts(sim=sim,
dst=dst,
uLab=uLab,
rowLabels=testLabels,
colLabels=tmplLabels,
classLabels=seqNames,
fileName=pngFN) #,mxDst=0.7)
if showPlots:
plt.show(block=False)
tikzFN = outFNBase + '_sim_vs_dst_clipped.tikz'
pngFN = outFNBase + '_sim_vs_dst_clipped.png'
#visSimsVsDsts(sim=sim,dst=dst,uLab=uLab,rowLabels=testLabels,colLabels=tmplLabels,classLabels=seqNames,fileName=[pdfFN,pngFN],tikzFileName=tikzFN)#,mxDst=0.7)
visSimsVsDsts(sim=sim,
dst=dst,
uLab=uLab,
rowLabels=testLabels,
colLabels=tmplLabels,
classLabels=seqNames,
clip=25,
fileName=pngFN) #,mxDst=0.7)
if showPlots:
plt.show(block=False)
# save config to file
targetCfgFile = outFNBase + '_cfg.ini'
cfg.set('net', 'weightsshapes', "{}".format(descrNet.getNumWeights()))
cfg.set(
'net', 'totalnumweights', "{}".format(
numpy.sum([numpy.prod(ws) for ws in descrNet.getNumWeights()])))
if not cfg.has_section('results'):
cfg.add_section('results')
cfg.set('results', 'acc', "{}".format(acc))
cfg.set('results', 'classAcc', "{}".format(classAcc))
cfg.set('results', 'minSepScore', "{}".format(minSepScore))
cfg.set('results', 'maxSepScore', "{}".format(maxSepScore))
cfg.set('results', 'meanSepScore', "{}".format(meanSepScore))
cfg.set('results', 'medSepScore', "{}".format(medSepScore))
with open(targetCfgFile, 'wb') as configfile:
cfg.write(configfile)
print("Saveing cfg along with other stuff to:")
print(" '{}'".format(targetCfgFile))
######################################
# wait for quit
if showPlots:
ch = 0
plt.show(block=False)
print("waiting for 'q'")
while ((ch & 0xFF) != ord('q')) and (ch >= 0):
ch = cv2.waitKey()
print(" ... got '{}' ({})".format(chr(ch & 0xFF), ch))
cv2.destroyAllWindows()
if (showPlots):
print("showing plots")
else:
print("not showing plots")
def linemod_vis_test_descr_main(fileName=None,cfg=None):
theano.config.exception_verbosity = 'high'
if fileName is None:
g = glob.glob('../data/results/tmp/*_descrNet.pkl')
fileName = sorted(g)[-1]
if os.path.isdir(fileName):
g = glob.glob('{}/*_descrNet.pkl'.format(fileName))
fileName = sorted(g)[-1]
outFNBase = fileName[:-len("descrNet.pkl")] + 'test'
print("outFNBase {}".format(outFNBase))
showPlots = readCfgIntParam(cfg,'misc','showPlots',default=1) > 0
######## load net
res = loadTrainResultsPickle(fileName)
descrNet = res[0]
cfg = res[2]
descrNetTrainerParams = res[3]
imgsPklFileName = readCfgParam(cfg, 'input', 'imgsPklFileName', default="")
trainSetPklFileName = readCfgParam(cfg, 'train', 'trainSetPklFileName', default="")
#-------------
nChan = descrNet.cfgParams.layers[0].image_shape[1]
if nChan == 1:
inputMode = 0
elif nChan == 3:
inputMode = 1
elif nChan == 4:
inputMode = 2
else:
raise ValueError("Unknown input mode: #channels = {}".format(nChan))
inputSize = descrNet.cfgParams.layers[0].image_shape[2]
print("inputMode = {}".format(inputMode))
print("inputSize = {}".format(inputSize))
######## prepare test data
lmDataBasepath = readCfgParam(cfg,'paths', 'lmDataBase',default='/home/wohlhart/work/data/linemod/')
seqNamesStr = readCfgParam(cfg, 'input','objs', default="")
seqNames = map(str.strip,seqNamesStr.split(","))
zRotInvStr = readCfgParam(cfg, 'input','zRotInv', default=[])
zRotInv = list(eval(zRotInvStr))
###############################################################################################
# create test set from test seq that was stored by main_linemod_realandsynth
# which did the split of the linemod data into train and test
print("Loading images from pkl ... ")
# LOAD from pkl
pklload_start_time = time.clock()
with open(imgsPklFileName, 'rb') as f:
_,_,testSeqs = cPickle.load(f)
pklload_end_time = time.clock()
print ('Loading took %.1fs' % ((pklload_end_time - pklload_start_time)) )
testdata_set = BatchedImgSeqDataset()
testdata_set.initFromImgSeqs(testSeqs, inputMode=inputMode, batchSize=descrNet.cfgParams.batch_size)
#############################################################
# general informations about tmpl and test datasets
#
testDataValidIdx, = (testdata_set.y >= 0).nonzero()
testLabels = testdata_set.y[testDataValidIdx]
nTestSamp = len(testDataValidIdx)
testRots = testdata_set.sampleInfo['rots'][testDataValidIdx]
testRots = testRots / numpy.tile(numpy.linalg.norm(testRots,axis=1).reshape(nTestSamp,1),(1,3)) # normalize
#print("tmplRots.shape {}, testRots.shape {}".format(tmplRots.shape,testRots.shape))
#################################################################################
# create a dataManager
dataManagerConfig = MacroBatchManagerConfig()
dataManagerConfig.batch_size = testdata_set.batchSize
macroBatchSize = readCfgIntParam(cfg,'trainset','macroBatchSize',default=-1)
maxBytes = eval(readCfgParam(cfg,'trainset','maxBytes',default=50*1024*1024))
nTestSamples = testdata_set.x.shape[0]
nTestBatches = nTestSamples / testdata_set.batchSize
#nTmplSamples = tmpl_set.x.shape[0]
#nTmplBatches = nTmplSamples / tmpl_set.batchSize
batchSizeBytes = testdata_set.x.nbytes / nTestBatches
macroBatchSize = findNiceMacroBatchSize(macroBatchSize, maxBytes, batchSizeBytes, [nTestBatches])
dataManager = MacroBatchManager(dataManagerConfig,macroBatchSize,acceptPartialMacroBatches=False)
dataManager.setupVariables(testdata_set, doBorrow=True)
#################################################################################
# visualize testdata and traindata
# LOAD DATA
# create trainer for macrobatches descriptor computation
#rng = numpy.random.RandomState(23455)
#descrNetTrainer = NetworkTrainer(descrNet,descrNetTrainerParams,rng)
#descrNetTrainer.setData(train_set,val_set,macroBatchSize=1)
calcAndVis3DDescriptorsWithRot(descrNet,testdata_set,dataManager)
plt.title('test data')
######################################
# wait for quit
cv2.waitKey(10)
plt.show(block=True)
ch = 0
while ((ch & 0xFF) != ord('q')) and (ch >= 0):
print("waiting for 'q'")
ch = cv2.waitKey()
print(" ... got '{}' ({})".format(chr(ch & 0xFF),ch))
cv2.destroyAllWindows()
