def plot(self, select=''):
measureList = []
if select == '':
measureList = self._measureList
else:
selections = select.split(',')
for selection in selections:
measureList += tb.wildcardMatch(self._measureList, selection)
measureList = tb.unique(measureList)
plot = Plot("loss/accuracy for %s" % self._networkName)
def plotMeasure(name,label,color):
plot.plotList(label, self._measures[name], color)
measureList.remove(name)
for name, (label, style) in Config.plotMeasureStyles.iteritems():
if name in measureList:
plotMeasure(name, name, style)
styles = tb.styleList()
for name in measureList[:]:
if name.startswith('test_'):
plotMeasure(name, name, styles.pop(0))
for name in measureList[:]:
plotMeasure(name, name, styles.pop(0))
plot.finish()
def plot(self, select=''):
measureList = []
if select == '':
measureList = self._measureList
else:
selections = select.split(',')
for selection in selections:
measureList += tb.wildcardMatch(self._measureList, selection)
measureList = tb.unique(measureList)
plot = Plot("loss/accuracy for %s" % self._networkName)
def plotMeasure(name, label, color):
plot.plotList(label, self._measures[name], color)
measureList.remove(name)
for name, (label, style) in Config.plotMeasureStyles.iteritems():
if name in measureList:
plotMeasure(name, name, style)
styles = tb.styleList()
for name in measureList[:]:
if name.startswith('test_'):
plotMeasure(name, name, styles.pop(0))
for name in measureList[:]:
plotMeasure(name, name, styles.pop(0))
plot.finish()
def compare(self, networks, losses):
folders = [
dir for dir in os.listdir('.')
if os.path.isdir(dir) and not dir.startswith('.')
]
networks = tb.wildcardMatch(folders, networks)
logs = []
measureNames = []
for net in networks:
logfile = '%s/training/log.txt' % net
print 'reading %s' % logfile
logs.append(Log(net, logfile))
for name in logs[-1].measureNames():
if name not in measureNames: measureNames.append(name)
if losses is not None:
selectedNames = tb.unique(tb.wildcardMatch(measureNames, losses))
else:
selectedNames = tb.unique(measureNames)
print 'comparing networks:'
for net in networks:
print " ", net
print 'comparing losses: '
for name in selectedNames:
print " ", name
Log.plotComparison(selectedNames, logs)
def makeComputeJob(self, job):
job.setLog(self.file("log.txt"))
tb.noticeVerbose("creating compute job for <%s>" % self.img1Path(), "run")
job.addCommand(self.path(), "echo type=%s > info" % self.type())
self.computeJobImplementation(job)
self.epeJobImplementation(job)
def makeComputeJob(self,job):
job.setLog(self.file('log.txt'))
tb.noticeVerbose('creating compute job for <%s>' % self.img1Path(),'run')
job.addCommand(self.path(), "echo type=%s > info" % self.type())
self.computeJobImplementation(job)
self.epeJobImplementation(job)
def slice(net, input_blob, slice_point, axis=1):
return Layers.Slice(net,
input_blob,
nout=len(slice_point)+1 if tb.isList(slice_point) else 2,
slice_param={
'slice_point': slice_point if tb.isList(slice_point) else (slice_point,) ,
'axis': axis
})
def notice(self, message, type=None):
if self._silent:
return
if type is None:
print '%s\n' % message
else:
tb.notice(message, type)
def notice(self, message, type=None):
if self._silent:
return
if type is None:
print '%s\n' % message
else:
tb.notice(message, type)
def viewFilters(self, iter):
self.prepareTraining()
prototxt = self._trainDir + '/train.prototxt'
modelFile, iter = self.getModelFile(iter)
os.environ['LD_LIBRARY_PATH']="/misc/lmbraid17/sceneflownet/common/programs/torch/install/lib:/usr/lib/x86_64-linux-gnu:/misc/lmbraid17/sceneflownet/common/software-root/lib:/home/ilge/dev/hackathon-caffe2/build/lib:/misc/software-lin/Qt-5.3.2/5.3/gcc_64/lib:/misc/lmbraid17/sceneflownet/common/programs/torch/install/lib:/usr/lib/x86_64-linux-gnu:/misc/lmbraid17/sceneflownet/common/software-root/lib:/home/ilge/dev/hackathon-caffe2/build/lib:/misc/software-lin/Qt-5.3.2/5.3/gcc_64/lib::/home/ilge/lib:/misc/software-lin/lmbsoft/openni-1.5.2.23-x86_64/usr/lib:/misc/software-lin/lmbsoft/glog/lib:/misc/software-lin/lmbsoft/mkl/lib:/misc/software-lin/lmbsoft/mkl/lib/intel64:/misc/software-lin/lmbsoft/cuda-6.5.14-x86_64/lib64:/misc/software-lin/lmbsoft/cuda-6.0.37-x86_64/lib64:/misc/student/mayern/OpenNI-Bin-Dev-Linux-x64-v1.5.4.0/Lib:/home/ilge/lib:/misc/software-lin/lmbsoft/openni-1.5.2.23-x86_64/usr/lib:/misc/software-lin/lmbsoft/glog/lib:/misc/software-lin/lmbsoft/mkl/lib:/misc/software-lin/lmbsoft/mkl/lib/intel64:/misc/software-lin/lmbsoft/cuda-6.5.14-x86_64/lib64:/misc/software-lin/lmbsoft/cuda-6.0.37-x86_64/lib64:/misc/student/mayern/OpenNI-Bin-Dev-Linux-x64-v1.5.4.0/Lib"
os.environ['PATH']="/home/ilge/bin:/home/ilge/dev/pymill/bin:/misc/lmbraid17/sceneflownet/common/programs/torch/install/bin:/misc/lmbraid17/sceneflownet/common/software-root/bin:/misc/lmbraid17/sceneflownet/ilge/hackathon-caffe2/python/pymill/bin:/misc/software-lin/Qt-5.3.2/5.3/gcc_64/bin:/home/ilge/bin:/home/ilge/dev/pymill/bin:/misc/lmbraid17/sceneflownet/common/programs/torch/install/bin:/misc/lmbraid17/sceneflownet/common/software-root/bin:/misc/lmbraid17/sceneflownet/ilge/hackathon-caffe2/python/pymill/bin:/misc/software-lin/Qt-5.3.2/5.3/gcc_64/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/misc/software-lin/lmbsoft/cuda-6.5.14-x86_64/bin:/misc/software-lin/matlabR2013a/bin:/home/ilge/data/caffe/matching/bin:/misc/lmbraid15/hackathon/common/flo-results/bin:/misc/lmbraid17/sceneflownet/common/data_tools:/misc/software-lin/lmbsoft/cuda-6.5.14-x86_64/bin:/misc/software-lin/matlabR2013a/bin:/home/ilge/data/caffe/matching/bin:/misc/lmbraid15/hackathon/common/flo-results/bin:/misc/lmbraid17/sceneflownet/common/data_tools"
tb.system('/home/ilge/bin/weight-viewer %s %s' % (prototxt, modelFile))
def clean(self):
nCleaned = 0
for m in self._methods:
for ds in self._datasets:
for ent in ds.bents():
if ent.bind(m).clean():
nCleaned += 1
tb.notice('cleaned %d entries' % (nCleaned), 'passed')
def clean(self):
nCleaned = 0
for m in self._methods:
for ds in self._datasets:
for ent in ds.bents():
if ent.bind(m).clean():
nCleaned += 1
tb.notice("cleaned %d entries" % (nCleaned), "passed")
def slice(net, input_blob, slice_point, axis=1):
return Layers.Slice(
net,
input_blob,
nout=len(slice_point) + 1 if tb.isList(slice_point) else 2,
slice_param={
'slice_point':
slice_point if tb.isList(slice_point) else (slice_point, ),
'axis':
axis
})
def viewFilters(self, iter):
self.prepareTraining()
prototxt = self._trainDir + '/train.prototxt'
modelFile, iter = self.getModelFile(iter)
os.environ[
'LD_LIBRARY_PATH'] = "/misc/lmbraid17/sceneflownet/common/programs/torch/install/lib:/usr/lib/x86_64-linux-gnu:/misc/lmbraid17/sceneflownet/common/software-root/lib:/home/ilge/dev/hackathon-caffe2/build/lib:/misc/software-lin/Qt-5.3.2/5.3/gcc_64/lib:/misc/lmbraid17/sceneflownet/common/programs/torch/install/lib:/usr/lib/x86_64-linux-gnu:/misc/lmbraid17/sceneflownet/common/software-root/lib:/home/ilge/dev/hackathon-caffe2/build/lib:/misc/software-lin/Qt-5.3.2/5.3/gcc_64/lib::/home/ilge/lib:/misc/software-lin/lmbsoft/openni-1.5.2.23-x86_64/usr/lib:/misc/software-lin/lmbsoft/glog/lib:/misc/software-lin/lmbsoft/mkl/lib:/misc/software-lin/lmbsoft/mkl/lib/intel64:/misc/software-lin/lmbsoft/cuda-6.5.14-x86_64/lib64:/misc/software-lin/lmbsoft/cuda-6.0.37-x86_64/lib64:/misc/student/mayern/OpenNI-Bin-Dev-Linux-x64-v1.5.4.0/Lib:/home/ilge/lib:/misc/software-lin/lmbsoft/openni-1.5.2.23-x86_64/usr/lib:/misc/software-lin/lmbsoft/glog/lib:/misc/software-lin/lmbsoft/mkl/lib:/misc/software-lin/lmbsoft/mkl/lib/intel64:/misc/software-lin/lmbsoft/cuda-6.5.14-x86_64/lib64:/misc/software-lin/lmbsoft/cuda-6.0.37-x86_64/lib64:/misc/student/mayern/OpenNI-Bin-Dev-Linux-x64-v1.5.4.0/Lib"
os.environ[
'PATH'] = "/home/ilge/bin:/home/ilge/dev/pymill/bin:/misc/lmbraid17/sceneflownet/common/programs/torch/install/bin:/misc/lmbraid17/sceneflownet/common/software-root/bin:/misc/lmbraid17/sceneflownet/ilge/hackathon-caffe2/python/pymill/bin:/misc/software-lin/Qt-5.3.2/5.3/gcc_64/bin:/home/ilge/bin:/home/ilge/dev/pymill/bin:/misc/lmbraid17/sceneflownet/common/programs/torch/install/bin:/misc/lmbraid17/sceneflownet/common/software-root/bin:/misc/lmbraid17/sceneflownet/ilge/hackathon-caffe2/python/pymill/bin:/misc/software-lin/Qt-5.3.2/5.3/gcc_64/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/misc/software-lin/lmbsoft/cuda-6.5.14-x86_64/bin:/misc/software-lin/matlabR2013a/bin:/home/ilge/data/caffe/matching/bin:/misc/lmbraid15/hackathon/common/flo-results/bin:/misc/lmbraid17/sceneflownet/common/data_tools:/misc/software-lin/lmbsoft/cuda-6.5.14-x86_64/bin:/misc/software-lin/matlabR2013a/bin:/home/ilge/data/caffe/matching/bin:/misc/lmbraid15/hackathon/common/flo-results/bin:/misc/lmbraid17/sceneflownet/common/data_tools"
tb.system('/home/ilge/bin/weight-viewer %s %s' % (prototxt, modelFile))
def archive(self, src, target, iter_step=-1):
basename = os.path.basename(src)
targetPath = os.path.join(target, basename)
if os.path.exists(targetPath):
raise Exception("target path %s already exists" % targetPath)
env = Environment(src,backend=self._backend, unattended=self._unattended, silent=self._silent)
env.init()
env.shrink(iter_step=iter_step)
tb.notice('archiving %s to %s' % (src, targetPath), 'run')
os.system('mv %s %s' % (src, targetPath))
def update(self):
queue = tb.Queue()
for m in self._methods:
for ds in self._datasets:
tb.notice('creating jobs for <%s> on <%s>' % (m, ds))
for ent in ds.uents() if m.direction() == '' else ds.bents():
job = tb.Job()
ent.bind(m).makeUpdateJob(job)
queue.postJob(job)
queue.finishPacket()
queue.submit(local=self._args.local, cores=self._args.cores)
def check(self):
nTotal = 0
nOk = 0
for m in self._methods:
for ds in self._datasets:
for ent in ds.uents() if m.direction() == '' else ds.bents():
if ent.bind(m).checkOut(self._args.verbose):
nOk += 1
nTotal += 1
if nOk == nTotal:
tb.notice('(%d/%d) passed' % (nOk, nTotal), 'passed')
else:
tb.notice('(%d/%d) passed' % (nOk, nTotal), 'failed')
def prototxt(self, inFile, outDir, defs={}):
defs['name'] = self._name
if not os.path.isfile(inFile):
raise Exception('input file %s not file' % inFile)
if inFile.endswith('.prototxt'):
os.system('cp %s %s' % (inFile, outDir))
return '%s' % (inFile)
elif inFile.endswith('.prototmp'):
prototxt = '%s/%s.prototxt' % (
outDir, os.path.basename(inFile).replace('.prototmp', ''))
if not self._silent: tb.notice('preprocessing %s' % inFile, 'run')
tb.preprocessFile(inFile, prototxt, defs)
return prototxt
elif inFile.endswith('.py'):
prototxt = '%s/%s.prototxt' % (
outDir, os.path.basename(inFile).replace('.py', ''))
args = ''
for k, v in defs.iteritems():
if len(args): args += ' '
args += '%s=%s' % (k, v)
if not self._silent:
if not len(defs):
tb.notice('converting %s' % inFile, 'run')
else:
tb.notice('converting %s (%s)' % (inFile, args), 'run')
if os.system('python -B %s %s > %s' %
(inFile, args, prototxt)) != 0:
raise Exception('conversion of %s failed' % inFile)
return prototxt
else:
raise Exception('don\'t know how to convert file %s to prototxt' %
inFile)
def prototxt(self, inFile, outDir, defs={}):
defs['name'] = self._name
if not os.path.isfile(inFile):
raise Exception('input file %s not file' % inFile)
if inFile.endswith('.prototxt'):
os.system('cp %s %s' % (inFile, outDir))
return '%s' % (inFile)
elif inFile.endswith('.prototmp'):
prototxt = '%s/%s.prototxt' % (outDir, os.path.basename(inFile).replace('.prototmp', ''))
if not self._silent: tb.notice('preprocessing %s' % inFile, 'run')
tb.preprocessFile(inFile, prototxt, defs)
return prototxt
elif inFile.endswith('.py'):
prototxt = '%s/%s.prototxt' % (outDir, os.path.basename(inFile).replace('.py', ''))
args = ''
for k, v in defs.iteritems():
if len(args): args += ' '
args += '%s=%s' % (k, v)
if not self._silent:
if not len(defs):
tb.notice('converting %s' % inFile, 'run')
else:
tb.notice('converting %s (%s)' % (inFile, args), 'run')
if os.system('python -B %s %s > %s' % (inFile, args, prototxt)) != 0:
raise Exception('conversion of %s failed' % inFile)
return prototxt
else:
raise Exception('don\'t know how to convert file %s to prototxt' % inFile)
def update(self):
queue = tb.Queue()
for m in self._methods:
for ds in self._datasets:
tb.notice("creating jobs for <%s> on <%s>" % (m, ds))
for ent in ds.uents() if m.direction() == "" else ds.bents():
job = tb.Job()
ent.bind(m).makeUpdateJob(job)
queue.postJob(job)
queue.finishPacket()
queue.submit(local=self._args.local, cores=self._args.cores)
def check(self):
nTotal = 0
nOk = 0
for m in self._methods:
for ds in self._datasets:
for ent in ds.uents() if m.direction() == "" else ds.bents():
if ent.bind(m).checkOut(self._args.verbose):
nOk += 1
nTotal += 1
if nOk == nTotal:
tb.notice("(%d/%d) passed" % (nOk, nTotal), "passed")
else:
tb.notice("(%d/%d) passed" % (nOk, nTotal), "failed")
def archive(self, src, target, iter_step=-1):
basename = os.path.basename(src)
targetPath = os.path.join(target, basename)
if os.path.exists(targetPath):
raise Exception("target path %s already exists" % targetPath)
env = Environment(src,
backend=self._backend,
unattended=self._unattended,
silent=self._silent)
env.init()
env.shrink(iter_step=iter_step)
tb.notice('archiving %s to %s' % (src, targetPath), 'run')
os.system('mv %s %s' % (src, targetPath))
def readBents(bents):
list = tb.readTupleList(bents)
clips = []
clip = []
current = None
for ent in list:
if ent[2] != current:
if len(clip): clips.append(clip)
clip = []
current = ent[2]
clip.append(ent)
if len(clip): clips.append(clip)
data = OrderedDict()
for clip in clips:
sceneName = clip[0][2]
data[sceneName] = []
for frame in clip:
data[sceneName].append({'cleanImageL': frame[5], 'finalImageL': frame[5].replace('clean','final'), 'forwardFlowL': frame[7]})
data[sceneName].append({'cleanImageL': frame[6], 'finalImageL': frame[6].replace('clean','final')})
return data
def runProto(self, proto):
defFile = proto
modelFile, iter = self._env.getModelFile(self._iter)
print 'testing for iteration %d ...' % self._iter
if self._output:
dir = 'output_%s_%d' % (self._name, self._iter)
tb.system('mkdir -p %s' % dir)
self._variables['TEST_OUTPUT'] = 1
self._variables['TEST_OUTPUT_DIR'] = '"\\"%s\\""' % dir
self._env.makeScratchDir()
defPrototxt = self._env.prototxt(defFile, 'scratch', self._variables)
print defFile, defPrototxt
tb.system('%s test -weights %s -model %s -gpu 0 -iterations %d 2>&1' % (Environment.caffeBin(), modelFile, defPrototxt, self._iterations))
def epeStat(self):
if len(self._methods) > 1:
raise Exception('cannot make epe stat for more than one method')
method = self._methods[0]
entries = []
for ds in self._datasets:
for ent in ds.bents():
entries.append(ent.bind(method).flowStatParams())
cmd = 'FlowStat --epe-type=%s %s %s %s %s' % (
self._args.type, '--make-epe' if self._args.make_epe else '',
'--make-stat' if self._args.make_stat else '',
'--refresh' if self._args.refresh else '', '--stat')
print tb.run(cmd, '\n'.join(entries)),
def scale(net, image_blob, factor):
if tb.isList(factor):
return Layers.Convolution(net,
image_blob,
nout=1,
convolution_param={
'num_output': len(factor),
'pad': 0,
'kernel_size': 1,
'stride': 1,
'weight_filler': {
'type': 'diagonal',
'diag_val': factor
},
'bias_filler': {
'type': 'constant'
}
})
else:
return Layers.Eltwise(net,
image_blob,
nout=1,
eltwise_param={
'operation': Params.Eltwise.SUM,
'coeff': (factor, )
})
def testref(self, iter, definition=None, vars={}, num_iter=-1):
modelFile, iter = self.getModelFile(iter)
vars['output'] = True
vars['prefix'] = 'ref_%d' % iter
vars['lowres'] = True
self.makeScratchDir()
if definition is None: definition = 'test'
proto = self.findProto(definition)
outPath = '%s/output_%d_%s' % (self._path, iter, vars['dataset'])
if os.path.isdir(outPath):
if self._unattended or tb.queryYesNo(
'Output folder %s exists, do you want to delete it first?'
% os.path.basename(outPath)):
os.system('rm -rf %s' % outPath)
finalProto = self.makeScratchPrototxt(proto, vars)
solverProto = self.makeScratchPrototxt(self._solverProto, vars)
self.notice(
'testing snapshot iteration %d for %d iterations...' %
(iter, num_iter), 'notice')
os.chdir(self._path)
self._backend.test(caffemodelFilename=modelFile,
protoFilename=finalProto,
iterations=num_iter,
logFile=self._scratchLogFile)
def __init__(self, networkName, filename):
lines = open(filename, 'r').readlines()
iter = -1
self._networkName = networkName
self._lines = []
self._measures = {}
self._measureList = []
def appendMeasure(name, iter, value):
if name not in self._measures:
self._measures[name] = []
self._measureList.append(name)
self._measures[name].append((iter, float(value)))
for l in lines:
if ']' not in l:
self._lines.append((iter, l))
continue
msg = l.split(']')[1].strip()
if msg.startswith('Iteration'):
match = re.compile('Iteration ([0-9]+)').match(msg)
if match:
iter = int(match.group(1))
match = re.compile(
'Iteration [0-9]+, loss = (([0-9]|\.)+)').match(msg)
if match:
appendMeasure('train_loss', iter, match.group(1))
if msg.startswith('Test loss'):
value = re.compile('Test loss: (([0-9]|\.)+)').match(
msg).group(1)
appendMeasure('test_loss', iter, value)
if msg.startswith('Train net output'):
match = re.compile(
'Train net output ..: ([a-zA-Z0-9_-]+) = (([0-9]|\.)+)'
).match(msg)
if match:
name = match.group(1)
value = match.group(2)
appendMeasure('train_' + name, iter, value)
if msg.startswith('Test net output'):
match = re.compile(
'Test net output #[a-zA-Z0-9_-]+: ([a-zA-Z0-9_-]+) = (([0-9]|\.)+)'
).match(msg)
if match:
name = match.group(1)
value = match.group(2)
appendMeasure('test_' + name, iter, value)
self._lines.append((iter, l))
self._measureList = tb.unique(self._measureList)
def testFiles(self, filelist, iter, output=False, definition=None, vars={}):
modelFile, iter = self.getModelFile(iter)
assert(output)
vars['output'] = True
vars['prefix'] = iter
vars['dataset'] = os.path.splitext(os.path.basename(filelist))[0]
self.makeScratchDir()
if definition is None: definition = 'testsingle'
proto = self.findProto(definition)
if output and 'dataset' in vars:
outPath = '%s/output_%d_%s' % (self._path, iter, vars['dataset'])
if os.path.isdir(outPath):
if self._unattended or tb.queryYesNo('Output folder %s exists, do you want to delete it first?' % os.path.basename(outPath)):
os.system('rm -rf %s' % outPath)
data_list = IO.readTupleList(filelist)
print data_list
# Run a new net for every sample:
for idx, line in enumerate(data_list):
num_blobs = len(line)
input_data = []
for blob_idx in range(num_blobs):
img = IO.readImage(line[blob_idx])
#print(img.shape)
input_data.append(img[np.newaxis, :, :, :].transpose(0, 3, 1, 2)[:,[2,1,0],:,:])
#print(input_data[-1].shape)
vars['width'] = input_data[0].shape[3]
vars['height'] = input_data[0].shape[2]
vars['basename'] = 'b%03d' % idx
finalProto = self.makeScratchPrototxt(proto, vars)
caffe.set_logging_disabled()
caffe.set_mode_gpu()
net = caffe.Net(finalProto, modelFile,caffe.TEST)
print 'Network forward pass (%d of %d). %d inputs of shapes:' % (idx+1, len(data_list), num_blobs)
for blob_idx in range(num_blobs):
print(" " + str(input_data[blob_idx].shape))
if not len(net.inputs) == len(line):
raise Exception('Net has %d inputs and in file list there are %d' % (len(net.inputs), len(line)))
input_dict = {}
for blob_idx in range(num_blobs):
input_dict[net.inputs[blob_idx]] = input_data[blob_idx]
net.forward(**input_dict)
print 'Iteration was %d' %iter
def computeHistograms(resolution, subpath, collectionName, clips, skipIfExists=False, overwrite=True, numBins=5000, maxValue=1000):
dataPath = '/misc/lmbraid17/sceneflownet/common/data/4_bin-db'
savePath = '%s/hists/%s/%s' % (dataPath, resolution, subpath)
saveFile = '%s/%s.npz' % (savePath, collectionName)
completePath = os.path.dirname(saveFile)
os.system('mkdir -p %s' % completePath)
print 'savePath', savePath
print 'completePath', completePath
print 'saveFile', saveFile
histFlow = np.zeros(numBins)
histDisp = np.zeros(numBins)
histDispChange = np.zeros(numBins)
for clip in clips:
print 'processing', clip
for i in range(clip.startFrame(), clip.endFrame()+1):
frame = clip.frame(i)
if i%10 == 1:
print '%d/%d' % (i - clip.startFrame() + 1, clip.endFrame() - clip.startFrame() + 1)
# flow histogram
if frame.hasForwardFlowL():
values = tb.readFlow(frame.forwardFlowL())
mag = np.power(np.sum(np.power(values, 2), axis=2), 0.5)
curr_hist, _ = np.histogram(mag, bins=numBins, range=(0,maxValue))
histFlow += curr_hist
# disparity histogram
if frame.hasDispL():
values = tb.readDisparity(frame.dispL())
mag = np.abs(values)
curr_hist, _ = np.histogram(mag, bins=numBins, range=(0,maxValue))
histDisp += curr_hist
# disparity change histogram
if frame.hasBackwardDispChangeL():
values = tb.readDisparity(frame.forwardDispChangeL())
mag = np.abs(values)
curr_hist, _ = np.histogram(mag, bins=numBins, range=(0,maxValue))
histDispChange += curr_hist
np.savez(saveFile, histFlow=histFlow, histDisp=histDisp, histDispChange=histDispChange)
def runProto(self, proto):
defFile = proto
modelFile, iter = self._env.getModelFile(self._iter)
print 'testing for iteration %d ...' % self._iter
if self._output:
dir = 'output_%s_%d' % (self._name, self._iter)
tb.system('mkdir -p %s' % dir)
self._variables['TEST_OUTPUT'] = 1
self._variables['TEST_OUTPUT_DIR'] = '"\\"%s\\""' % dir
self._env.makeScratchDir()
defPrototxt = self._env.prototxt(defFile, 'scratch', self._variables)
print defFile, defPrototxt
tb.system(
'%s test -weights %s -model %s -gpu 0 -iterations %d 2>&1' %
(Environment.caffeBin(), modelFile, defPrototxt, self._iterations))
def epeStat(self):
if len(self._methods) > 1:
raise Exception("cannot make epe stat for more than one method")
method = self._methods[0]
entries = []
for ds in self._datasets:
for ent in ds.bents():
entries.append(ent.bind(method).flowStatParams())
cmd = "FlowStat --epe-type=%s %s %s %s %s" % (
self._args.type,
"--make-epe" if self._args.make_epe else "",
"--make-stat" if self._args.make_stat else "",
"--refresh" if self._args.refresh else "",
"--stat",
)
print tb.run(cmd, "\n".join(entries)),
def getSelectedCollections():
collectionNames = collections.keys()
selectedNames = []
if args.collections == '':
selectedNames = collectionNames
else:
exprs = args.collections.split(',')
for expr in exprs:
selectedNames += tb.wildcardMatch(collectionNames, expr)
selectedNames = tb.unique(selectedNames)
selectedCollections = {}
for name in selectedNames:
for collectionName, collection in collections.iteritems():
if collectionName == name:
selectedCollections[name] = collection
return selectedCollections
def getSelectedCollections():
collectionNames = collections.keys()
selectedNames = []
if args.collections == '':
selectedNames = collectionNames
else:
exprs = args.collections.split(',')
for expr in exprs:
selectedNames += tb.wildcardMatch(collectionNames, expr)
selectedNames = tb.unique(selectedNames)
selectedCollections = {}
for name in selectedNames:
for collectionName, collection in collections.iteritems():
if collectionName == name:
selectedCollections[name] = collection
return selectedCollections
def __init__(self, networkName, filename):
lines = open(filename,'r').readlines()
iter = -1
self._networkName = networkName
self._lines = []
self._measures = {}
self._measureList = []
def appendMeasure(name, iter, value):
if name not in self._measures:
self._measures[name]=[]
self._measureList.append(name)
self._measures[name].append((iter,float(value)))
for l in lines:
if ']' not in l:
self._lines.append((iter, l))
continue
msg = l.split(']')[1].strip()
if msg.startswith('Iteration'):
match = re.compile('Iteration ([0-9]+)').match(msg)
if match:
iter = int(match.group(1))
match = re.compile('Iteration [0-9]+, loss = (([0-9]|\.)+)').match(msg)
if match:
appendMeasure('train_loss', iter, match.group(1))
if msg.startswith('Test loss'):
value = re.compile('Test loss: (([0-9]|\.)+)').match(msg).group(1)
appendMeasure('test_loss', iter, value)
if msg.startswith('Train net output'):
match = re.compile('Train net output ..: ([a-zA-Z0-9_-]+) = (([0-9]|\.)+)').match(msg)
if match:
name = match.group(1)
value = match.group(2)
appendMeasure('train_'+name, iter, value)
if msg.startswith('Test net output'):
match = re.compile('Test net output #[a-zA-Z0-9_-]+: ([a-zA-Z0-9_-]+) = (([0-9]|\.)+)').match(msg)
if match:
name = match.group(1)
value = match.group(2)
appendMeasure('test_'+name, iter, value)
self._lines.append((iter, l))
self._measureList = tb.unique(self._measureList)
def compute(self, queue):
OpticalFlow.Method.compute(self, queue)
temp = Template(
'/misc/software-lin/matlabR2013a/bin/matlab -nodesktop -nojvm -r "gbDetect(\'$imagePath\',1); exit"'
)
command = temp.safe_substitute(imagePath=self.ent().img1Path())
job = tb.Job()
job.setPath(self.path())
job.addCommand(command)
queue.postJob(job)
def makeFlowCheck(resolution, rendertype, clip):
print clip.path()
checkPath = clip.path() + '/flow_check'
os.system('mkdir -p %s' % checkPath)
for i in range(clip.startFrame(), clip.endFrame()):
print 'frame %d' % i
frame = clip.frame(i, rendertype)
Image0 = misc.imread(frame.ImageL())
Image1 = misc.imread(frame.ImageL(+1))
flow = tb.readPFM(frame.forwardFlowL())[0][:,:,0:2]
warped = tb.flowWarp(Image1, flow)
h = Image0.shape[0]
w = Image0.shape[1]
ImageCheck = np.zeros((2*h, 2*w, 3)).astype(np.uint8)
ImageCheck[0:h, 0:w, :] = Image0
ImageCheck[0:h, w:2*w, :] = Image1
ImageCheck[h:2*h, 0:w, :] = warped
misc.imsave('%s/%s.png' % (checkPath, os.path.basename(frame.ImageL())), ImageCheck)
def compute(self, queue):
OpticalFlow.Method.compute(self, queue)
ent = self.ent()
input = self.spec().inputs()[0]
inputPath = ent.inFile(input, "lowres.flo")
if not os.path.exists(inputPath):
raise Exception('input does not exist: %s' % inputPath)
params = self.spec().params()
paramStr = ''
if 'b' in params and params['b'] == '1':
paramStr += ' --boundaries=' + ent.boundaryPath()
if 'delta' in params:
paramStr += ' --delta=' + params['delta']
if 'beta' in params:
paramStr += ' --beta=' + params['beta']
if 'sigma' in params:
paramStr += ' --sigma=' + params['sigma']
os.system("rm -rf %s" % ent.outFile(str(self.spec()), 'flow.flo'))
if ent.flowGTPath != None:
temp = Template(
'VarFlowRefine $img1 $img2 $flo $out $params --groundtruth=$groundtruth --make-epe --make-stat'
)
command = temp.safe_substitute(
img1=ent.img1Path(),
img2=ent.img2Path(),
flo=inputPath,
out=ent.outFile(str(self.spec()), 'flow.flo'),
groundtruth='%s:%s:%s' %
(ent.flowGTPath(), ent.occGTPath(), ent.mbdistGTPath()),
params=paramStr)
else:
temp = Template('VarFlowRefine $img1 $img2 $flo $out $params')
command = temp.safe_substitute(img1=ent.img1Path(),
img2=ent.img2Path(),
flo=inputPath,
out=ent.outFile(
str(self.spec()), 'flow.flo'),
params=paramStr)
job = tb.Job()
job.setPath(self.path())
job.addCommand(command)
queue.postJob(job)
def plotComparison(names, logs):
plot = Plot("loss/accuracy comparison")
styles = tb.styleList()
for log in logs:
style = styles.pop(0)
#lineStyles = tb.lineStyleCycle()
for name in names:
label = '%s for %s' % (name, log.networkName())
#style.setLineStyle(lineStyles.get())
if name in log.measures():
plot.plotList(label, log.measures()[name], style)
plot.finish()
def plotComparison(names, logs):
plot = Plot("loss/accuracy comparison")
styles = tb.styleList()
for log in logs:
style = styles.pop(0)
#lineStyles = tb.lineStyleCycle()
for name in names:
label = '%s for %s' %(name,log.networkName())
#style.setLineStyle(lineStyles.get())
if name in log.measures():
plot.plotList(label, log.measures()[name], style)
plot.finish()
def test(self,
iter,
output=False,
definition=None,
vars={},
num_iter=-1,
discard=False):
modelFile, iter = self.getModelFile(iter)
if output:
vars['output'] = True
vars['prefix'] = iter
self.makeScratchDir()
if definition is None: definition = 'test'
proto = self.findProto(definition)
if output and 'dataset' in vars:
outPath = '%s/output_%d_%s' % (self._path, iter, vars['dataset'])
if os.path.isdir(outPath):
if self._unattended or tb.queryYesNo(
'Output folder %s exists, do you want to delete it first?'
% os.path.basename(outPath)):
os.system('rm -rf %s' % outPath)
finalProto = self.makeScratchPrototxt(proto, vars)
solverProto = self.makeScratchPrototxt(self._solverProto, vars)
self.notice(
'testing snapshot iteration %d for %d iterations...' %
(iter, num_iter), 'notice')
os.chdir(self._path)
self._backend.test(caffemodelFilename=modelFile,
protoFilename=finalProto,
iterations=num_iter,
logFile=self._scratchLogFile)
if output and 'dataset' in vars and num_iter == -1:
outPath = '%s/output_%d_%s' % (self._path, iter, vars['dataset'])
if os.path.isdir(outPath):
logFile = '%s/log.txt' % outPath
print 'saving log to %s', logFile
os.system('cp %s %s' % (self._scratchLogFile, logFile))
if 'dataset' in vars:
self._saveTestResults(iter, vars['dataset'], not discard)
print 'Iteration was %d' % iter
def epeLists(methods,datasets,type='all'):
epe = {}
for m in methods:
entries = []
for ds in datasets:
for ent in ds.bents():
entries.append(ent.bind(m).flowStatParams())
cmd = 'FlowStat --epe-type=%s' % (
type,
)
epe[str(m)] = [line for line in tb.run(cmd, '\n'.join(entries)).split('\n') if line.strip() != '']
return epe
def compare(self, networks, losses):
folders = [dir for dir in os.listdir('.') if os.path.isdir(dir) and not dir.startswith('.')]
networks = tb.wildcardMatch(folders, networks)
logs = []
measureNames = []
for net in networks:
logfile = '%s/training/log.txt' % net
print 'reading %s' % logfile
logs.append(Log(net, logfile))
for name in logs[-1].measureNames():
if name not in measureNames: measureNames.append(name)
if losses is not None:
selectedNames = tb.unique(tb.wildcardMatch(measureNames, losses))
else:
selectedNames = tb.unique(measureNames)
print 'comparing networks:'
for net in networks: print " ", net
print 'comparing losses: '
for name in selectedNames: print " ", name
Log.plotComparison(selectedNames, logs)
def _callCopiedBin(self, cmd):
bin = './' + os.path.basename(caffeBin())
tb.notice('making a local copy of %s' % caffeBin())
os.system('cp %s .' % caffeBin())
ldd = tb.run('ldd %s' % caffeBin())
caffeLib = None
for line in ldd.split('\n'):
match = re.match('\\s*libcaffe.so => (.*\.so)', line)
if match:
caffeLib = match.group(1)
break
if caffeLib is None:
raise Exception('cannot find libcaffe.so dependency')
tb.notice('making a local copy of %s' % caffeLib)
os.system('cp %s .' % caffeLib)
cmd = 'GLOG_logtostderr=%d LD_LIBRARY_PATH=.:$LD_LIBRARY_PATH %s %s' % (not self._quiet, bin, cmd)
if not self._silent:
tb.notice('running "%s"' % cmd, 'run')
tb.system(cmd)
def resume(self, iter=-1):
if not len(self._stateFiles):
raise Exception('no .solverstate files to continue from')
stateFile, iter = self.getStateFile(iter)
if self.existingData(iter) and not self._unattended:
if not tb.queryYesNo(
'Existing data beyond iteration %d found. Do you want to delete it and continue?' % iter):
return
self.clean(iter)
solverFilename = self.prepareTraining()
self.notice('continuing from iteration %d ...' % iter, 'notice')
os.chdir(self._trainDir)
self._backend.resume(solverFilename=solverFilename, solverstateFilename=stateFile, logFile=self._logFile)
def train(self, weights=None, blobSummary=False):
if self.existingData() and not self._unattended:
if not tb.queryYesNo('Existing data found. Do you want to delete it and start from scratch?'):
return
self.clean()
solverFilename = self.prepareTraining()
self.notice('training...')
os.chdir(self._trainDir)
if weights != '':
self._backend.train(solverFilename=solverFilename, logFile=self._logFile, weights=weights)
else:
self._backend.train(solverFilename=solverFilename, logFile=self._logFile)
if blobSummary:
self.displayBlobSummary(self._logFile)
def _callCopiedBin(self, cmd):
bin = './' + os.path.basename(caffeBin())
tb.notice('making a local copy of %s' % caffeBin())
os.system('cp %s .' % caffeBin())
ldd = tb.run('ldd %s' % caffeBin())
caffeLib = None
for line in ldd.split('\n'):
match = re.match('\\s*libcaffe.so => (.*\.so)', line)
if match:
caffeLib = match.group(1)
break
if caffeLib is None:
raise Exception('cannot find libcaffe.so dependency')
tb.notice('making a local copy of %s' % caffeLib)
os.system('cp %s .' % caffeLib)
cmd = 'GLOG_logtostderr=%d LD_LIBRARY_PATH=.:$LD_LIBRARY_PATH %s %s' % (
not self._quiet, bin, cmd)
if not self._silent:
tb.notice('running "%s"' % cmd, 'run')
tb.system(cmd)
def preparePythonBackend():
os.system('mkdir -p training')
folder = os.path.dirname(caffe.__file__)
print 'copying %s to training' % folder
os.system('cp %s training -r' % folder)
ldd = tb.run('ldd %s' % caffe._caffe.__file__)
caffeLib = None
for line in ldd.split('\n'):
match = re.match('\\s*libcaffe.so => (.*\.so)', line)
if match:
caffeLib = match.group(1)
break
if caffeLib is None:
raise Exception('cannot find libcaffe.so dependency')
print 'copying %s to training' % caffeLib
os.system('cp %s %s' % (caffeLib, env.trainDir()))
def preparePythonBackend():
os.system("mkdir -p training")
folder = os.path.dirname(caffe.__file__)
print "copying %s to training" % folder
os.system("cp %s training -r" % folder)
ldd = tb.run("ldd %s" % caffe._caffe.__file__)
caffeLib = None
for line in ldd.split("\n"):
match = re.match("\\s*libcaffe.so => (.*\.so)", line)
if match:
caffeLib = match.group(1)
break
if caffeLib is None:
raise Exception("cannot find libcaffe.so dependency")
print "copying %s to training" % caffeLib
os.system("cp %s %s" % (caffeLib, env.trainDir()))
def concat(net, *args, **kwargs):
'''
@brief Setup a ConcatLayer that takes all ARGS and throws them together along the first dimension
@param net Current network
@returns A new blob (concatenation of all blobs in ARGS)
'''
axis = 1
if 'axis' in kwargs:
axis = int(kwargs['axis'])
del kwargs['axis']
if len(kwargs):
raise Exception('Cannot handle kwargs %s' % kwargs)
inputs = []
for arg in args:
if tb.isList(arg): inputs += arg
else: inputs.append(arg)
return Layers.Concat(net, inputs, nout=1, concat_param={'axis': axis})
def resume(self, iter=-1):
if not len(self._stateFiles):
raise Exception('no .solverstate files to continue from')
stateFile, iter = self.getStateFile(iter)
if self.existingData(iter) and not self._unattended:
if not tb.queryYesNo(
'Existing data beyond iteration %d found. Do you want to delete it and continue?'
% iter):
return
self.clean(iter)
solverFilename = self.prepareTraining()
self.notice('continuing from iteration %d ...' % iter, 'notice')
os.chdir(self._trainDir)
self._backend.resume(solverFilename=solverFilename,
solverstateFilename=stateFile,
logFile=self._logFile)
def scale(net, image_blob, factor):
if tb.isList(factor):
return Layers.Convolution(net,
image_blob,
nout=1,
convolution_param={
'num_output': len(factor),
'pad': 0,
'kernel_size': 1,
'stride': 1,
'weight_filler': {'type': 'diagonal', 'diag_val': factor},
'bias_filler': {'type': 'constant'}
})
else:
return Layers.Eltwise(net,
image_blob,
nout=1,
eltwise_param={
'operation': Params.Eltwise.SUM,
'coeff': (factor,)
})
def train(self, weights=None, blobSummary=False):
if self.existingData() and not self._unattended:
if not tb.queryYesNo(
'Existing data found. Do you want to delete it and start from scratch?'
):
return
self.clean()
solverFilename = self.prepareTraining()
self.notice('training...')
os.chdir(self._trainDir)
if weights != '':
self._backend.train(solverFilename=solverFilename,
logFile=self._logFile,
weights=weights)
else:
self._backend.train(solverFilename=solverFilename,
logFile=self._logFile)
if blobSummary:
self.displayBlobSummary(self._logFile)
def concat(net, *args, **kwargs):
'''
@brief Setup a ConcatLayer that takes all ARGS and throws them together along the first dimension
@param net Current network
@returns A new blob (concatenation of all blobs in ARGS)
'''
axis = 1
if 'axis' in kwargs:
axis = int(kwargs['axis'])
del kwargs['axis']
if len(kwargs):
raise Exception('Cannot handle kwargs %s' % kwargs)
inputs = []
for arg in args:
if tb.isList(arg): inputs += arg
else: inputs.append(arg)
return Layers.Concat(net,
inputs,
nout=1,
concat_param={'axis': axis})
def data(self, downsample = 1):
if self._type == "image":
Image = misc.imread(self._path)[:, :, 0:3]
if downsample != 1: Image = tb.downsampleImage(Image, downsample)
return Image[..., np.r_[2, 1, 0]].transpose((2, 0, 1))
elif self._type == "flow":
flow = tb.readFlow(self._path)
flow = flow[:, :, 0:2]
if downsample != 1: flow = OpticalFlow.downsampleMedian(flow, downsample)
flow = flow.transpose((2, 0, 1))
nanMask = np.isnan(flow)
data = (flow * 32.0).astype(np.int16)
data[nanMask] = np.iinfo(np.int16).max
return data
elif self._type == "leftdisparity":
disparity = tb.readDisparity(self._path)
nanMask = np.isnan(disparity)
disparity *= -1
if downsample != 1: raise Exception("no downsampling implemented for disparity")
data = (disparity * 32.0).astype(np.int16)
data[nanMask] = np.iinfo(np.int16).max
return data
elif self._type == "rightdisparity":
disparity = tb.readDisparity(self._path)
nanMask = np.isnan(disparity)
if downsample != 1: raise Exception("no downsampling implemented for disparity")
data = (disparity * 32.0).astype(np.int16)
data[nanMask] = np.iinfo(np.int16).max
return data
elif self._type == "leftdisparitychange":
disparityChange = tb.readDisparity(self._path)
nanMask = np.isnan(disparityChange)
disparityChange *= -1
if downsample != 1: raise Exception("no downsampling implemented for disparity")
data = (disparityChange * 32.0).astype(np.int16)
data[nanMask] = np.iinfo(np.int16).max
return data
elif self._type == "rightdisparitychange":
disparityChange = tb.readDisparity(self._path)
nanMask = np.isnan(disparityChange)
if downsample != 1: raise Exception("no downsampling implemented for disparity")
data = (disparityChange * 32.0).astype(np.int16)
data[nanMask] = np.iinfo(np.int16).max
return data
else:
raise Exception('unhandled data type')
def testref(self, iter, definition=None, vars={}, num_iter=-1):
modelFile, iter = self.getModelFile(iter)
vars['output'] = True
vars['prefix'] = 'ref_%d' % iter
vars['lowres'] = True
self.makeScratchDir()
if definition is None: definition = 'test'
proto = self.findProto(definition)
outPath = '%s/output_%d_%s' % (self._path, iter, vars['dataset'])
if os.path.isdir(outPath):
if self._unattended or tb.queryYesNo('Output folder %s exists, do you want to delete it first?' % os.path.basename(outPath)):
os.system('rm -rf %s' % outPath)
finalProto = self.makeScratchPrototxt(proto, vars)
solverProto = self.makeScratchPrototxt(self._solverProto, vars)
self.notice('testing snapshot iteration %d for %d iterations...' % (iter, num_iter), 'notice')
os.chdir(self._path)
self._backend.test(caffemodelFilename=modelFile, protoFilename=finalProto, iterations=num_iter, logFile=self._scratchLogFile)
def test(self, iter, output=False, definition=None, vars={}, num_iter=-1, discard=False):
modelFile, iter = self.getModelFile(iter)
if output:
vars['output'] = True
vars['prefix'] = iter
self.makeScratchDir()
if definition is None: definition = 'test'
proto = self.findProto(definition)
if output and 'dataset' in vars:
outPath = '%s/output_%d_%s' % (self._path, iter, vars['dataset'])
if os.path.isdir(outPath):
if self._unattended or tb.queryYesNo('Output folder %s exists, do you want to delete it first?' % os.path.basename(outPath)):
os.system('rm -rf %s' % outPath)
finalProto = self.makeScratchPrototxt(proto, vars)
solverProto = self.makeScratchPrototxt(self._solverProto, vars)
self.notice('testing snapshot iteration %d for %d iterations...' % (iter, num_iter), 'notice')
os.chdir(self._path)
self._backend.test(caffemodelFilename=modelFile, protoFilename=finalProto, iterations=num_iter, logFile=self._scratchLogFile)
if output and 'dataset' in vars and num_iter==-1:
outPath = '%s/output_%d_%s' % (self._path, iter, vars['dataset'])
if os.path.isdir(outPath):
logFile = '%s/log.txt' % outPath
print 'saving log to %s', logFile
os.system('cp %s %s' % (self._scratchLogFile, logFile))
if 'dataset' in vars:
self._saveTestResults(iter,vars['dataset'], not discard)
print 'Iteration was %d' %iter
def plotlr(self):
lrs = []
maxLr = 0
for l in self._lines:
iter = l[0]
if not ']' in l[1]: continue
msg = l[1].split(']')[1].strip()
if msg.startswith('Iteration'):
match = re.compile(
'Iteration [0-9]+, lr = (([0-9]|\.|-|e)+)').match(msg)
if match:
lr = float(match.group(1))
if lr > maxLr:
maxLr = lr
lrs.append((iter, lr))
plot = Plot("learning rate for %s" % self._networkName)
if len(lrs):
plot.plotList('LR', lrs, tb.PlotStyle('r-'))
plt.ylim((0, maxLr * 1.1))
plot.finish()
def shrink(self, iter_step):
self.notice('removing *.pyc', 'del')
os.system('rm -f %s/*.pyc' % (self._path))
self.notice('removing scratch', 'del')
os.system('rm -rf %s/scratch' % self._path)
if self.haveJobDir():
self.notice('removing jobs', 'del')
os.system('rm -rf %s' % self._jobDir)
self.sweep()
if self.haveTrainDir():
for file in self._stateFiles:
keep = False
if iter_step != -1:
if file.iteration() % iter_step == 0:
keep = True
if file == self._stateFiles[-1]:
keep = True
modelFile = None
for f in self._modelFiles:
if f.iteration() == file.iteration():
modelFile = f
if keep:
if modelFile is not None:
tb.notice(
'keeping file %s' %
(os.path.basename(file.filename())), 'passed')
tb.notice(
'keeping file %s' %
(os.path.basename(modelFile.filename())), 'passed')
else:
tb.notice('keeping file %s' % file, 'passed')
else:
if modelFile is not None:
modelFile.delete(True)
file.delete(True)
else:
file.delete(True)
def delete(self,verbose=False):
if verbose:
tb.notice('removing %s' % self._filename, 'del')
os.remove(self._filename)
