def net_prms(dbFile=DEF_DB % 'nyu2_net', **kwargs):
dArgs = mec.get_default_net_prms(dbFile, **kwargs)
del dArgs['expStr']
#The data NetDefProto
dArgs.dataNetDefProto = 'data_layer_nyu2'
#the basic network architecture: baseNetDefProto
dArgs.baseNetDefProto = 'doublefc-v1_window_fc6'
#the loss layers:
dArgs.lossNetDefProto = 'nyu2_loss_classify_layers'
if dArgs.batchSize is None:
dArgs.batchSize = 128
#The amount of jitter in both the images
dArgs.maxJitter = 0
#The size of crop that should be cropped from the image
dArgs.cropScale = 0.9
#the size to which the cropped image should be resized
dArgs.ipImSz = 101
##The mean file
dArgs.meanFile = ''
dArgs.meanType = None
dArgs.opLrMult = None
dArgs = mpu.get_defaults(kwargs, dArgs, False)
allKeys = dArgs.keys()
dArgs['expStr'] = mec.get_sql_id(dbFile, dArgs)
return dArgs, allKeys
def get_default_solver_prms(dbFile=DEF_DB, **kwargs): ''' Refer to caffe.proto for a description of the variables. ''' dArgs = edict() dArgs.baseSolDefFile = None dArgs.iter_size = 1 dArgs.max_iter = 250000 dArgs.base_lr = 0.001 dArgs.lr_policy = 'step' dArgs.stepsize = 20000 dArgs.gamma = 0.5 dArgs.weight_decay = 0.0005 dArgs.clip_gradients = -1 #Momentum dArgs.momentum = 0.9 #Other dArgs.regularization_type = 'L2' dArgs.random_seed = -1 #Testing info dArgs.test_iter = 100 dArgs.test_interval = 1000 dArgs.snapshot = 2000 dArgs.display = 20 #Update parameters dArgs = mpu.get_defaults(kwargs, dArgs, False) dArgs.expStr = 'solprms' + get_sql_id(dbFile, dArgs, ignoreKeys=['test_iter', 'test_interval', 'snapshot', 'display']) return dArgs
def get_default_solver_prms(dbFile=DEF_DB, **kwargs):
'''
Refer to caffe.proto for a description of the
variables.
'''
dArgs = edict()
dArgs.baseSolDefFile = None
dArgs.iter_size = 1
dArgs.max_iter = 250000
dArgs.base_lr = 0.001
dArgs.lr_policy = 'step'
dArgs.stepsize = 20000
dArgs.gamma = 0.5
dArgs.weight_decay = 0.0005
dArgs.clip_gradients = -1
#Momentum
dArgs.momentum = 0.9
#Other
dArgs.regularization_type = 'L2'
dArgs.random_seed = -1
#Testing info
dArgs.test_iter = 100
dArgs.test_interval = 1000
dArgs.snapshot = 2000
dArgs.display = 20
#Update parameters
dArgs = mpu.get_defaults(kwargs, dArgs, False)
dArgs.expStr = 'solprms' + get_sql_id(
dbFile,
dArgs,
ignoreKeys=['test_iter', 'test_interval', 'snapshot', 'display'])
return dArgs
def net_prms(dbFile=DEF_DB % 'net', **kwargs):
dArgs = mec.get_siamese_net_prms(dbFile, **kwargs)
del dArgs['expStr']
#The data NetDefProto
dArgs.dataNetDefProto = 'data_layer_groups'
#the basic network architecture: baseNetDefProto
dArgs.baseNetDefProto = 'smallnet-v5_window_siamese_fc5'
#the loss layers:
dArgs.lossNetDefProto = 'pose_loss_log_l1_layers'
if dArgs.batchSize is None:
dArgs.batchSize = 128
#The amount of jitter in both the images
dArgs.maxJitter = 0
#The amount of roll jitter to apply to the images
dArgs.maxRollJitter = None
#The size of crop that should be cropped from the image
dArgs.crpSz = 192
#the size to which the cropped image should be resized
dArgs.ipImSz = 101
#The size of the fc layer if present
dArgs.fcSz = None
dArgs.fcName = 'fc5'
##The mean file
dArgs.meanFile = ''
dArgs.meanType = None
dArgs.ncpu = 3
dArgs.readSingleGrp = False
dArgs = mpu.get_defaults(kwargs, dArgs, False)
allKeys = dArgs.keys()
dArgs['expStr'] = mec.get_sql_id(dbFile,
dArgs,
ignoreKeys=['ncpu', 'readSingleGrp'])
return dArgs, allKeys
def get_lr_prms(**kwargs):
#return mec.get_lr_prms()
dArgs = edict()
dArgs.batchsize = 128
dArgs.stepsize = 20000
dArgs.base_lr = 0.001
dArgs.max_iter = 250000
dArgs.gamma = 0.5
dArgs.weight_decay = 0.0005
dArgs.clip_gradients = -1
dArgs.debug_info = False
dArgs = mpu.get_defaults(kwargs, dArgs)
#Make the solver
debugStr = '%s' % dArgs.debug_info
debugStr = debugStr.lower()
del dArgs['debug_info']
solArgs = edict({
'test_iter': 100,
'test_interval': 1000,
'snapshot': 2000,
'debug_info': debugStr
})
print dArgs.keys()
for k in dArgs.keys():
if k in ['batchsize']:
continue
solArgs[k] = copy.deepcopy(dArgs[k])
dArgs.solver = mpu.make_solver(**solArgs)
expStr = 'batchSz%d_stepSz%.0e_blr%.5f_mxItr%.1e_gamma%.2f_wdecay%.6f'\
% (dArgs.batchsize, dArgs.stepsize, dArgs.base_lr,
dArgs.max_iter, dArgs.gamma, dArgs.weight_decay)
if not (dArgs.clip_gradients == -1):
expStr = '%s_gradClip%.1f' % (expStr, dArgs.clip_gradients)
dArgs.expStr = expStr
return dArgs
def get_data_prms(dbFile=DEF_DB % 'nyu2_data', **kwargs):
dArgs = edict()
dArgs.dataset = 'nyu2'
allKeys = dArgs.keys()
dArgs = mpu.get_defaults(kwargs, dArgs)
dArgs['expStr'] = mec.get_sql_id(dbFile, dArgs)
dArgs['paths'] = snyu.get_paths()
return dArgs
def get_siamese_window_net_prms(dbFile=DEF_DB, **kwargs): dArgs = get_siamese_net_prms(dbFile) del dArgs['expStr'] #Size of input image dArgs.imSz = 227 #If random cropping is to be used dArgs.randCrop = False #If gray scale images need to be used dArgs.isGray = False dArgs = mpu.get_defaults(kwargs, dArgs, False) dArgs.expStr = get_sql_id(dbFile, dArgs) return dArgs
def get_siamese_window_net_prms(dbFile=DEF_DB, **kwargs):
dArgs = get_siamese_net_prms(dbFile)
del dArgs['expStr']
#Size of input image
dArgs.imSz = 227
#If random cropping is to be used
dArgs.randCrop = False
#If gray scale images need to be used
dArgs.isGray = False
dArgs = mpu.get_defaults(kwargs, dArgs, False)
dArgs.expStr = get_sql_id(dbFile, dArgs)
return dArgs
def get_siamese_net_prms(dbFile=DEF_DB, **kwargs): dArgs = get_default_net_prms(dbFile) del dArgs['expStr'] #Layers at which the nets are to be concatenated dArgs.concatLayer = 'fc6' #If dropouts should be used in the concatenation layer dArgs.concatDrop = False #Number of filters in concatenation layer dArgs.concatSz = None #If an extra FC layer needs to be added dArgs.extraFc = None dArgs = mpu.get_defaults(kwargs, dArgs, False) dArgs.expStr = get_sql_id(dbFile, dArgs) return dArgs
def get_siamese_net_prms(dbFile=DEF_DB, **kwargs):
dArgs = get_default_net_prms(dbFile)
del dArgs['expStr']
#Layers at which the nets are to be concatenated
dArgs.concatLayer = 'fc6'
#If dropouts should be used in the concatenation layer
dArgs.concatDrop = False
#Number of filters in concatenation layer
dArgs.concatSz = None
#If an extra FC layer needs to be added
dArgs.extraFc = None
dArgs = mpu.get_defaults(kwargs, dArgs, False)
dArgs.expStr = get_sql_id(dbFile, dArgs)
return dArgs
def get_default_net_prms(dbFile=DEF_DB, **kwargs): dArgs = edict() #Name of the net which will be constructed dArgs.netName = 'alexnet' #For layers below lrAbove, learning rate is set to 0 dArgs.lrAbove = None #If weights from a pretrained net are to be used dArgs.preTrainNet = None #The base proto from which net will be constructed dArgs.baseNetDefProto = None #Batch size dArgs.batchSize = None #runNum dArgs.runNum = 0 dArgs = mpu.get_defaults(kwargs, dArgs, False) dArgs.expStr = get_sql_id(dbFile, dArgs) return dArgs
def get_default_net_prms(dbFile=DEF_DB, **kwargs):
dArgs = edict()
#Name of the net which will be constructed
dArgs.netName = 'alexnet'
#For layers below lrAbove, learning rate is set to 0
dArgs.lrAbove = None
#If weights from a pretrained net are to be used
dArgs.preTrainNet = None
#The base proto from which net will be constructed
dArgs.baseNetDefProto = None
#Batch size
dArgs.batchSize = None
#runNum
dArgs.runNum = 0
dArgs = mpu.get_defaults(kwargs, dArgs, False)
dArgs.expStr = get_sql_id(dbFile, dArgs)
return dArgs
def get_data_prms(dbFile=DEF_DB % 'data', lbPrms=None, tvPrms=None, **kwargs):
if lbPrms is None:
lbPrms = slu.PosePrms()
if tvPrms is None:
tvPrms = get_trainval_split_prms()
dArgs = mec.edict()
dArgs.dataset = 'dc-v2'
dArgs.lbStr = lbPrms.get_lbstr()
dArgs.tvStr = tvPrms.pStr
dArgs.isAlign = True
allKeys = dArgs.keys()
dArgs = mpu.get_defaults(kwargs, dArgs)
dArgs['expStr'] = mec.get_sql_id(dbFile, dArgs)
dArgs['splitPrms'] = tvPrms
dArgs['lbPrms'] = lbPrms
dArgs['paths'] = get_paths(dArgs)
return dArgs
def get_finetune_prms(**kwargs):
'''
sourceModelIter: The number of model iterations of the source model to consider
fine_max_iter : The maximum iterations to which the target model should be trained.
lrAbove : If learning is to be performed some layer.
fine_base_lr : The base learning rate for finetuning.
fineRunNum : The run num for the finetuning.
fineNumData : The amount of data to be used for the finetuning.
fineMaxLayer : The maximum layer of the source n/w that should be considered.
'''
dArgs = edict()
dArgs.base_lr = 0.001
dArgs.runNum = 1
dArgs.maxLayer = None
dArgs.lrAbove = None
dArgs.dataset = 'sun'
dArgs.maxIter = 40000
dArgs.extraFc = False
dArgs.extraFcDrop = False
dArgs.sourceModelIter = 60000
dArgs = mpu.get_defaults(kwargs, dArgs)
return dArgs
def get_data_prms(dbFile=DEF_DB % 'pascal_data', **kwargs):
dArgs = edict()
dArgs.dataset = 'pascal'
dArgs.imCutSz = 256
dArgs.imPadSz = 36
dArgs.angleFormat = 'radian'
dArgs.anglePreProc = 'mean_sub'
dArgs.nAzBins = None
dArgs.nElBins = None
allKeys = dArgs.keys()
dArgs = mpu.get_defaults(kwargs, dArgs)
if dArgs.anglePreProc == 'classify':
assert dArgs.nAzBins is not None
assert dArgs.nElBins is not None
dArgs['expStr'] = mec.get_sql_id(dbFile, dArgs)
dArgs['paths'] = get_paths(dArgs)
dArgs.azBins = None
dArgs.elBins = None
if dArgs.nAzBins is not None:
dArgs.azBins = np.linspace(-np.pi, np.pi, dArgs.nAzBins + 1)
if dArgs.nElBins is not None:
dArgs.elBins = np.linspace(-np.pi, np.pi, dArgs.nElBins + 1)
return dArgs
def get_nw_prms(**kwargs):
#return mec.get_nw_prms(**kwargs)
dArgs = edict()
dArgs.netName = 'alexnet'
dArgs.concatLayer = 'fc6'
dArgs.concatDrop = False
dArgs.contextPad = 0
dArgs.imSz = 227
dArgs.imgntMean = True
dArgs.maxJitter = 11
dArgs.randCrop = False
dArgs.lossWeight = 1.0
dArgs.multiLossProto = None
dArgs.ptchStreamNum = 256
dArgs.poseStreamNum = 256
dArgs.isGray = False
dArgs.isPythonLayer = False
dArgs.extraFc = None
dArgs.numFc5 = None
dArgs.numConv4 = None
dArgs.numCommonFc = None
dArgs.lrAbove = None
dArgs = mpu.get_defaults(kwargs, dArgs)
if dArgs.numFc5 is not None:
assert (dArgs.concatLayer == 'fc5')
expStr = 'net-%s_cnct-%s_cnctDrp%d_contPad%d_imSz%d_imgntMean%d_jit%d'\
%(dArgs.netName, dArgs.concatLayer, dArgs.concatDrop,
dArgs.contextPad,
dArgs.imSz, dArgs.imgntMean, dArgs.maxJitter)
if dArgs.numFc5 is not None:
expStr = '%s_numFc5-%d' % (expStr, dArgs.numFc5)
if dArgs.numConv4 is not None:
expStr = '%s_numConv4-%d' % (expStr, dArgs.numConv4)
if dArgs.numCommonFc is not None:
expStr = '%s_numCommonFc-%d' % (expStr, dArgs.numCommonFc)
if dArgs.randCrop:
expStr = '%s_randCrp%d' % (expStr, dArgs.randCrop)
if not (dArgs.lossWeight == 1.0):
if type(dArgs.lossWeight) == list:
lStr = ''
for i, l in enumerate(dArgs.lossWeight):
lStr = lStr + 'lw%d-%.1f_' % (i, l)
lStr = lStr[0:-1]
print lStr
expStr = '%s_%s' % (expStr, lStr)
else:
assert isinstance(dArgs.lossWeight, (int, long, float))
expStr = '%s_lw%.1f' % (expStr, dArgs.lossWeight)
if dArgs.multiLossProto is not None:
expStr = '%s_mlpr%s-posn%d-ptsn%d' % (expStr, dArgs.multiLossProto,
dArgs.poseStreamNum,
dArgs.ptchStreamNum)
if dArgs.isGray:
expStr = '%s_grayIm' % expStr
if dArgs.isPythonLayer:
expStr = '%s_pylayers' % expStr
if dArgs.extraFc is not None:
expStr = '%s_extraFc%d' % (expStr, dArgs.extraFc)
if dArgs.lrAbove is not None:
expStr = '%s_lrAbove-%s' % (expStr, dArgs.lrAbove)
dArgs.expStr = expStr
return dArgs