def create_network(points, batchIds, features, numInputFeatures, batchSize, k, isTraining, multiConv = True, useMC = True):
############################################ Compute point hierarchy
mPointHierarchy = PointHierarchy(points, features, batchIds, [], "MCNormS_PH", batchSize)
############################################ Convolutions
mConvBuilder = ConvolutionBuilder(KDEWindow=0.2)
# Convolution 1
convFeatures1 = mConvBuilder.create_convolution(
convName="Conv_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
inFeatures=features,
inNumFeatures=numInputFeatures,
outNumFeatures=k,
convRadius=0.15,
multiFeatureConv=True)
#BatchNorm and RELU
convFeatures1 = batch_norm_RELU_drop_out("BN_RELU", convFeatures1, isTraining, False, False)
# Convolution 2
normals = mConvBuilder.create_convolution(
convName="Conv_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
inFeatures=convFeatures1,
inNumFeatures=k,
outNumFeatures=3,
convRadius=0.15,
multiFeatureConv=True)
return normals
def compute_initial_pts(points, batchIds, features, batchSize):
mPointHierarchy = PointHierarchy(points,
features,
batchIds, [0.01],
"MCGIInit_PH1",
batchSize,
relativeRadius=False)
return mPointHierarchy.sampledIndexs_[0], mPointHierarchy.points_[1]
def create_point_hierarchy_output(points,
batchIds,
features,
batchSize,
relRad=False):
############################################ Create the point hierarchy
mPointHierarchy = PointHierarchy(inPoints=points,
inFeatures=features,
inBatchIds=batchIds,
radiusList=[],
hierarchyName="MCPtDeNoise_MPH_2",
batchSize=batchSize,
relativeRadius=relRad)
return mPointHierarchy
def create_point_hierarchy_input(points,
batchIds,
features,
batchSize,
radiusList=[0.025, 0.05],
relRad=False,
hierarchyName="MCPtDeNoise_MPH_1"):
############################################ Create the point hierarchy
mPointHierarchy = PointHierarchy(inPoints=points,
inFeatures=features,
inBatchIds=batchIds,
radiusList=radiusList,
hierarchyName=hierarchyName,
batchSize=batchSize,
relativeRadius=relRad)
return mPointHierarchy
def ptH(radii, scope, batchSize, numFeatures=3):
inPts = tf.placeholder(tf.float32, [None, 3], name='ptH_points')
inFeatures = tf.placeholder(tf.float32, [None, numFeatures],
name='ptH_features')
inBatchIds = tf.placeholder(tf.int32, [None, 1], name='ptH_batchIds')
feed_ptH = [inPts, inFeatures, inBatchIds]
outPointHierarchy = PointHierarchy(inPts, inFeatures, inBatchIds, radii,
scope, batchSize)
PtHierPoints = [[]] + outPointHierarchy.points_[1:5]
PtHierSampledPts = outPointHierarchy.sampledIndexs_[0:4]
PtHierFeatures = [[]] + [outPointHierarchy.features_[1]]
PtHierBatchIds = [[]] + outPointHierarchy.batchIds_[1:5]
PtHierAabbMin = outPointHierarchy.aabbMin_
PtHierAabbMax = outPointHierarchy.aabbMax_
outPtH = [
PtHierPoints, PtHierSampledPts, PtHierFeatures, PtHierBatchIds,
PtHierAabbMin, PtHierAabbMax
]
return outPtH, feed_ptH
def create_network(points,
batchIds,
features,
numInputFeatures,
batchSize,
k,
isTraining,
multiConv=True,
useMC=True):
############################################ Compute point hierarchy
mPointHierarchy = PointHierarchy(points, features, batchIds, [0.1, 0.4],
"MCNorm_PH", batchSize)
############################################ Convolutions
mConvBuilder = ConvolutionBuilder(KDEWindow=0.25)
############################################ Encoder
# First Convolution
convFeatures1 = mConvBuilder.create_convolution(
convName="Conv_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
inFeatures=features,
inNumFeatures=numInputFeatures,
outNumFeatures=k,
convRadius=0.1,
multiFeatureConv=True)
# First Pooling
bnConvFeatures1 = batch_norm_RELU_drop_out("Reduce_Pool_1_In_BN",
convFeatures1, isTraining,
False, False)
bnConvFeatures1 = conv_1x1("Reduce_Pool_1", bnConvFeatures1, k, k * 2)
bnConvFeatures1 = batch_norm_RELU_drop_out("Reduce_Pool_1_Out_BN",
bnConvFeatures1, isTraining,
False, False)
poolFeatures1 = mConvBuilder.create_convolution(
convName="Pool_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
outPointLevel=1,
inFeatures=bnConvFeatures1,
inNumFeatures=k * 2,
convRadius=0.2,
KDEWindow=0.2)
# Second Convolution
bnPoolFeatures1 = batch_norm_RELU_drop_out("Conv_2_In_BN", poolFeatures1,
isTraining, False, False)
convFeatures2 = mConvBuilder.create_convolution(
convName="Conv_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
inFeatures=bnPoolFeatures1,
inNumFeatures=k * 2,
convRadius=0.4)
convFeatures2 = tf.concat([poolFeatures1, convFeatures2], 1)
# Second Pooling
bnConvFeatures2 = batch_norm_RELU_drop_out("Reduce_Pool_2_In_BN",
convFeatures2, isTraining,
False, False)
bnConvFeatures2 = conv_1x1("Reduce_Pool_2", bnConvFeatures2, k * 4, k * 4)
bnConvFeatures2 = batch_norm_RELU_drop_out("Reduce_Pool_2_Out_BN",
bnConvFeatures2, isTraining,
False, False)
poolFeatures2 = mConvBuilder.create_convolution(
convName="Pool_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=2,
inFeatures=bnConvFeatures2,
inNumFeatures=k * 4,
convRadius=0.8,
KDEWindow=0.2)
# Third Convolution
bnPoolFeatures2 = batch_norm_RELU_drop_out("Conv_3_In_BN", poolFeatures2,
isTraining, False, False)
convFeatures3 = mConvBuilder.create_convolution(
convName="Conv_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
inFeatures=bnPoolFeatures2,
inNumFeatures=k * 4,
convRadius=math.sqrt(3))
convFeatures3 = tf.concat([poolFeatures2, convFeatures3], 1)
##################################################### Multi-hierarchy sampling
# Second upsampling
bnFeatures3 = batch_norm_RELU_drop_out("Up_2_3_BN", convFeatures3,
isTraining, False, False)
upFeatures2_3 = mConvBuilder.create_convolution(
convName="Up_2_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=1,
inFeatures=bnFeatures3,
inNumFeatures=k * 8,
convRadius=0.8)
deConvFeatures2 = tf.concat([upFeatures2_3, convFeatures2], 1)
deConvFeatures2 = batch_norm_RELU_drop_out("DeConv_2_Reduce_In_BN",
deConvFeatures2, isTraining,
False, False)
deConvFeatures2 = conv_1x1("DeConv_2_Reduce", deConvFeatures2, k * 12,
k * 4)
deConvFeatures2 = batch_norm_RELU_drop_out("DeConv_2_Reduce_Out_BN",
deConvFeatures2, isTraining,
False, False)
deConvFeatures2 = mConvBuilder.create_convolution(
convName="DeConv_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
inFeatures=deConvFeatures2,
inNumFeatures=k * 4,
convRadius=0.4)
# First upsampling
bnDeConvFeatures2 = batch_norm_RELU_drop_out("Up_1_2_BN", deConvFeatures2,
isTraining, False, False)
upFeatures1_2 = mConvBuilder.create_convolution(
convName="Up_1_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=0,
inFeatures=bnDeConvFeatures2,
inNumFeatures=k * 4,
convRadius=0.2)
deConvFeatures1 = tf.concat([upFeatures1_2, convFeatures1], 1)
deConvFeatures1 = batch_norm_RELU_drop_out("DeConv_1_Reduce_In_BN",
deConvFeatures1, isTraining,
False, False)
deConvFeatures1 = conv_1x1("DeConv_1_Reduce", deConvFeatures1, k * 5,
k * 2)
deConvFeatures1 = batch_norm_RELU_drop_out("DeConv_1_Reduce_Out_BN",
deConvFeatures1, isTraining,
False, False)
normals = mConvBuilder.create_convolution(convName="DeConv_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
inFeatures=deConvFeatures1,
inNumFeatures=k * 2,
outNumFeatures=3,
convRadius=0.1,
multiFeatureConv=True)
return normals
def create_network(points,
batchIds,
features,
numInputFeatures,
batchSize,
k,
numOutCat,
isTraining,
keepProbConv,
keepProbFull,
useConvDropOut=False,
useDropOutFull=True):
############################################ Compute point hierarchy
mPointHierarchy = PointHierarchy(points, features, batchIds,
[0.1, 0.4, math.sqrt(3.0) + 0.1],
"MCClassS_PH", batchSize)
############################################ Convolutions
mConvBuilder = ConvolutionBuilder(KDEWindow=0.2)
#### Convolution 1
convFeatures1 = mConvBuilder.create_convolution(
convName="Conv_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
outPointLevel=1,
inFeatures=features,
inNumFeatures=numInputFeatures,
outNumFeatures=k,
convRadius=0.2,
multiFeatureConv=True)
#### Convolution 2
convFeatures1 = batch_norm_RELU_drop_out("Reduce_1_In_BN", convFeatures1,
isTraining, useConvDropOut,
keepProbConv)
convFeatures1 = conv_1x1("Reduce_1", convFeatures1, k, k * 2)
convFeatures1 = batch_norm_RELU_drop_out("Reduce_1_Out_BN", convFeatures1,
isTraining, useConvDropOut,
keepProbConv)
convFeatures2 = mConvBuilder.create_convolution(
convName="Conv_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=2,
inFeatures=convFeatures1,
inNumFeatures=k * 2,
convRadius=0.8)
#### Convolution 3
convFeatures2 = batch_norm_RELU_drop_out("Reduce_2_In_BN", convFeatures2,
isTraining, useConvDropOut,
keepProbConv)
convFeatures2 = conv_1x1("Reduce_2", convFeatures2, k * 2, k * 4)
convFeatures2 = batch_norm_RELU_drop_out("Reduce_2_Out_BN", convFeatures2,
isTraining, useConvDropOut,
keepProbConv)
convFeatures3 = mConvBuilder.create_convolution(
convName="Conv_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=3,
inFeatures=convFeatures2,
inNumFeatures=k * 4,
convRadius=math.sqrt(3.0) + 0.1)
#Fully connected MLP - Global features.
finalInput = batch_norm_RELU_drop_out("BNRELUDROP_final", convFeatures3,
isTraining, useConvDropOut,
keepProbConv)
finalLogits = MLP_2_hidden(finalInput, k * 4, k * 2, k, numOutCat,
"Final_Logits", keepProbFull, isTraining,
useDropOutFull)
return finalLogits
def create_network(points,
batchIds,
features,
catLabels,
numInputFeatures,
numCats,
numParts,
batchSize,
k,
isTraining,
keepProbConv,
keepProbFull,
useConvDropOut=False,
useDropOutFull=True):
############################################ Compute point hierarchy
mPointHierarchy = PointHierarchy(points, features, batchIds,
[0.025, 0.1, 0.4], "MCSeg_PH", batchSize)
############################################ Convolutions
mConvBuilder = ConvolutionBuilder(KDEWindow=0.25)
############################################ Encoder
# First Convolution
convFeatures1 = mConvBuilder.create_convolution(
convName="Conv_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
inFeatures=features,
inNumFeatures=numInputFeatures,
outNumFeatures=k,
convRadius=0.03,
multiFeatureConv=True)
# First Pooling
bnConvFeatures1 = batch_norm_RELU_drop_out("Reduce_Pool_1_In_BN",
convFeatures1, isTraining,
useConvDropOut, keepProbConv)
bnConvFeatures1 = conv_1x1("Reduce_Pool_1", bnConvFeatures1, k, k * 2)
bnConvFeatures1 = batch_norm_RELU_drop_out("Reduce_Pool_1_Out_BN",
bnConvFeatures1, isTraining,
useConvDropOut, keepProbConv)
poolFeatures1 = mConvBuilder.create_convolution(
convName="Pool_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
outPointLevel=1,
inFeatures=bnConvFeatures1,
inNumFeatures=k * 2,
convRadius=0.05,
KDEWindow=0.2)
# Second Convolution
bnPoolFeatures1 = batch_norm_RELU_drop_out("Conv_2_In_BN", poolFeatures1,
isTraining, useConvDropOut,
keepProbConv)
convFeatures2 = mConvBuilder.create_convolution(
convName="Conv_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
inFeatures=bnPoolFeatures1,
inNumFeatures=k * 2,
convRadius=0.1)
convFeatures2 = tf.concat([poolFeatures1, convFeatures2], 1)
# Second Pooling
bnConvFeatures2 = batch_norm_RELU_drop_out("Reduce_Pool_2_In_BN",
convFeatures2, isTraining,
useConvDropOut, keepProbConv)
bnConvFeatures2 = conv_1x1("Reduce_Pool_2", bnConvFeatures2, k * 4, k * 4)
bnConvFeatures2 = batch_norm_RELU_drop_out("Reduce_Pool_2_Out_BN",
bnConvFeatures2, isTraining,
useConvDropOut, keepProbConv)
poolFeatures2 = mConvBuilder.create_convolution(
convName="Pool_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=2,
inFeatures=bnConvFeatures2,
inNumFeatures=k * 4,
convRadius=0.2,
KDEWindow=0.2)
# Third Convolution
bnPoolFeatures2 = batch_norm_RELU_drop_out("Conv_3_In_BN", poolFeatures2,
isTraining, useConvDropOut,
keepProbConv)
convFeatures3 = mConvBuilder.create_convolution(
convName="Conv_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
inFeatures=bnPoolFeatures2,
inNumFeatures=k * 4,
convRadius=0.4)
convFeatures3 = tf.concat([poolFeatures2, convFeatures3], 1)
# Third Pooling
bnConvFeatures3 = batch_norm_RELU_drop_out("Reduce_Pool_3_In_BN",
convFeatures3, isTraining,
useConvDropOut, keepProbConv)
bnConvFeatures3 = conv_1x1("Reduce_Pool_3", bnConvFeatures3, k * 8, k * 8)
bnConvFeatures3 = batch_norm_RELU_drop_out("Reduce_Pool_3_Out_BN",
bnConvFeatures3, isTraining,
useConvDropOut, keepProbConv)
poolFeatures3 = mConvBuilder.create_convolution(
convName="Pool_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=3,
inFeatures=bnConvFeatures3,
inNumFeatures=k * 8,
convRadius=0.8,
KDEWindow=0.2)
# Fourth Convolution
bnPoolFeatures3 = batch_norm_RELU_drop_out("Conv_4_In_BN", poolFeatures3,
isTraining, useConvDropOut,
keepProbConv)
convFeatures4 = mConvBuilder.create_convolution(
convName="Conv_4",
inPointHierarchy=mPointHierarchy,
inPointLevel=3,
inFeatures=bnPoolFeatures3,
inNumFeatures=k * 8,
convRadius=math.sqrt(3.0) + 0.1)
convFeatures4 = tf.concat([poolFeatures3, convFeatures4], 1)
############################################ Decoder
# Third upsampling
bnConvFeatures4 = batch_norm_RELU_drop_out("Up_3_4_BN", convFeatures4,
isTraining, useConvDropOut,
keepProbConv)
upFeatures3_4 = mConvBuilder.create_convolution(
convName="Up_3_4",
inPointHierarchy=mPointHierarchy,
inPointLevel=3,
outPointLevel=2,
inFeatures=bnConvFeatures4,
inNumFeatures=k * 16,
convRadius=math.sqrt(3.0) + 0.1)
deConvFeatures3 = tf.concat([upFeatures3_4, convFeatures3], 1)
deConvFeatures3 = batch_norm_RELU_drop_out("DeConv_3_Reduce_In_BN",
deConvFeatures3, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures3 = conv_1x1("DeConv_3_Reduce", deConvFeatures3, k * 24,
k * 8)
deConvFeatures3 = batch_norm_RELU_drop_out("DeConv_3_Reduce_Out_BN",
deConvFeatures3, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures3 = mConvBuilder.create_convolution(
convName="DeConv_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
inFeatures=deConvFeatures3,
inNumFeatures=k * 8,
convRadius=0.4)
# Second upsampling
bnDeConvFeatures3 = batch_norm_RELU_drop_out("Up_2_3_BN", deConvFeatures3,
isTraining, useConvDropOut,
keepProbConv)
upFeatures2_3 = mConvBuilder.create_convolution(
convName="Up_2_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=1,
inFeatures=bnDeConvFeatures3,
inNumFeatures=k * 8,
convRadius=0.2)
deConvFeatures2 = tf.concat([upFeatures2_3, convFeatures2], 1)
deConvFeatures2 = batch_norm_RELU_drop_out("DeConv_2_Reduce_In_BN",
deConvFeatures2, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures2 = conv_1x1("DeConv_2_Reduce", deConvFeatures2, k * 12,
k * 4)
deConvFeatures2 = batch_norm_RELU_drop_out("DeConv_2_Reduce_Out_BN",
deConvFeatures2, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures2 = mConvBuilder.create_convolution(
convName="DeConv_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
inFeatures=deConvFeatures2,
inNumFeatures=k * 4,
convRadius=0.1)
# First multiple upsamplings
bnDeConvFeatures2 = batch_norm_RELU_drop_out("Up_1_2_BN", deConvFeatures2,
isTraining, useConvDropOut,
keepProbConv)
upFeatures1_2 = mConvBuilder.create_convolution(
convName="Up_1_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=0,
inFeatures=bnDeConvFeatures2,
inNumFeatures=k * 4,
convRadius=0.05)
bnDeConvFeatures3 = batch_norm_RELU_drop_out("Up_1_3_BN", deConvFeatures3,
isTraining, useConvDropOut,
keepProbConv)
upFeatures1_3 = mConvBuilder.create_convolution(
convName="Up_1_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=0,
inFeatures=bnDeConvFeatures3,
inNumFeatures=k * 8,
convRadius=0.2)
deConvFeatures1 = tf.concat([upFeatures1_2, upFeatures1_3, convFeatures1],
1)
deConvFeatures1 = batch_norm_RELU_drop_out("DeConv_1_Reduce_In_BN",
deConvFeatures1, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures1 = conv_1x1("DeConv_1_Reduce", deConvFeatures1, k * 13,
k * 4)
deConvFeatures1 = batch_norm_RELU_drop_out("DeConv_1_Reduce_Out_BN",
deConvFeatures1, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures1 = mConvBuilder.create_convolution(
convName="DeConv_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
inFeatures=deConvFeatures1,
inNumFeatures=k * 4,
convRadius=0.03)
# Fully connected MLP - Global features.
finalInput = batch_norm_RELU_drop_out("BNRELUDROP_hier_final",
deConvFeatures1, isTraining,
useConvDropOut, keepProbConv)
#Convert cat labels
catLabelOneHot = tf.one_hot(catLabels,
numCats,
on_value=1.0,
off_value=0.0)
catLabelOneHot = tf.reshape(catLabelOneHot, [-1, numCats])
finalInput = tf.concat([catLabelOneHot, finalInput], 1)
finalLogits = MLP_2_hidden(finalInput,
k * 4 + numCats,
k * 4,
k * 2,
numParts,
"Final_Logits",
keepProbFull,
isTraining,
useDropOutFull,
useInitBN=False)
return finalLogits
def create_network(points,
batchIds,
features,
numInputFeatures,
numSem,
batchSize,
k,
isTraining,
keepProbConv,
keepProbFull,
useConvDropOut=False,
useDropOutFull=True):
############################################ Compute point hierarchy
mPointHierarchy = PointHierarchy(points, features, batchIds,
[0.1, 0.2, 0.4, 0.8], "MCSegScanNet_PH",
batchSize, False)
############################################ Convolutions
mConvBuilder = ConvolutionBuilder(KDEWindow=0.25, relativeRadius=False)
############################################ Encoder
# Init pooling
poolFeatures0 = mConvBuilder.create_convolution(
convName="Pool_0",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
outPointLevel=1,
inFeatures=features,
inNumFeatures=numInputFeatures,
outNumFeatures=k,
convRadius=0.1,
KDEWindow=0.2,
multiFeatureConv=True)
# First Convolution
bnPoolFeatures0 = batch_norm_RELU_drop_out("Conv_1_In_BN", poolFeatures0,
isTraining, useConvDropOut,
keepProbConv)
convFeatures1 = mConvBuilder.create_convolution(
convName="Conv_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
inFeatures=bnPoolFeatures0,
inNumFeatures=k,
convRadius=0.4)
convFeatures1 = tf.concat([poolFeatures0, convFeatures1], 1)
# First Pooling
bnConvFeatures1 = batch_norm_RELU_drop_out("Reduce_Pool_1_In_BN",
convFeatures1, isTraining,
useConvDropOut, keepProbConv)
bnConvFeatures1 = conv_1x1("Reduce_Pool_1", bnConvFeatures1, k * 2, k * 2)
bnConvFeatures1 = batch_norm_RELU_drop_out("Reduce_Pool_1_Out_BN",
bnConvFeatures1, isTraining,
useConvDropOut, keepProbConv)
poolFeatures1 = mConvBuilder.create_convolution(
convName="Pool_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=2,
inFeatures=bnConvFeatures1,
inNumFeatures=k * 2,
convRadius=0.4,
KDEWindow=0.2)
# Second Convolution
bnPoolFeatures1 = batch_norm_RELU_drop_out("Conv_2_In_BN", poolFeatures1,
isTraining, useConvDropOut,
keepProbConv)
convFeatures2 = mConvBuilder.create_convolution(
convName="Conv_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
inFeatures=bnPoolFeatures1,
inNumFeatures=k * 2,
convRadius=0.8)
convFeatures2 = tf.concat([poolFeatures1, convFeatures2], 1)
# Second Pooling
bnConvFeatures2 = batch_norm_RELU_drop_out("Reduce_Pool_2_In_BN",
convFeatures2, isTraining,
useConvDropOut, keepProbConv)
bnConvFeatures2 = conv_1x1("Reduce_Pool_2", bnConvFeatures2, k * 4, k * 4)
bnConvFeatures2 = batch_norm_RELU_drop_out("Reduce_Pool_2_Out_BN",
bnConvFeatures2, isTraining,
useConvDropOut, keepProbConv)
poolFeatures2 = mConvBuilder.create_convolution(
convName="Pool_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=3,
inFeatures=bnConvFeatures2,
inNumFeatures=k * 4,
convRadius=0.8,
KDEWindow=0.2)
# Third Convolution
bnPoolFeatures2 = batch_norm_RELU_drop_out("Conv_3_In_BN", poolFeatures2,
isTraining, useConvDropOut,
keepProbConv)
convFeatures3 = mConvBuilder.create_convolution(
convName="Conv_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=3,
inFeatures=bnPoolFeatures2,
inNumFeatures=k * 4,
convRadius=1.6)
convFeatures3 = tf.concat([poolFeatures2, convFeatures3], 1)
# Third Pooling
bnConvFeatures3 = batch_norm_RELU_drop_out("Reduce_Pool_3_In_BN",
convFeatures3, isTraining,
useConvDropOut, keepProbConv)
bnConvFeatures3 = conv_1x1("Reduce_Pool_3", bnConvFeatures3, k * 8, k * 8)
bnConvFeatures3 = batch_norm_RELU_drop_out("Reduce_Pool_3_Out_BN",
bnConvFeatures3, isTraining,
useConvDropOut, keepProbConv)
poolFeatures3 = mConvBuilder.create_convolution(
convName="Pool_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=3,
outPointLevel=4,
inFeatures=bnConvFeatures3,
inNumFeatures=k * 8,
convRadius=1.6,
KDEWindow=0.2)
# Fourth Convolution
bnPoolFeatures3 = batch_norm_RELU_drop_out("Conv_4_In_BN", poolFeatures3,
isTraining, useConvDropOut,
keepProbConv)
convFeatures4 = mConvBuilder.create_convolution(
convName="Conv_4",
inPointHierarchy=mPointHierarchy,
inPointLevel=4,
inFeatures=bnPoolFeatures3,
inNumFeatures=k * 8,
convRadius=5.0)
convFeatures4 = tf.concat([poolFeatures3, convFeatures4], 1)
############################################ Decoder
# Third upsampling
bnConvFeatures4 = batch_norm_RELU_drop_out("Up3_4_Reduce_In_BN",
convFeatures4, isTraining,
useConvDropOut, keepProbConv)
bnConvFeatures4 = conv_1x1("Up3_4_Reduce", bnConvFeatures4, k * 16, k * 8)
bnConvFeatures4 = batch_norm_RELU_drop_out("Up3_4_Reduce_Out_BN",
bnConvFeatures4, isTraining,
useConvDropOut, keepProbConv)
upFeatures3_4 = mConvBuilder.create_convolution(
convName="Up_3_4",
inPointHierarchy=mPointHierarchy,
inPointLevel=4,
outPointLevel=3,
inFeatures=bnConvFeatures4,
inNumFeatures=k * 8,
convRadius=1.6)
upFeatures3_4 = tf.concat([upFeatures3_4, convFeatures3], 1)
deConvFeatures3 = batch_norm_RELU_drop_out("DeConv_3_Reduce_In_BN",
upFeatures3_4, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures3 = conv_1x1("DeConv_3_Reduce", deConvFeatures3, k * 16,
k * 8)
deConvFeatures3 = batch_norm_RELU_drop_out("DeConv_3_Reduce_Out_BN",
deConvFeatures3, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures3 = mConvBuilder.create_convolution(
convName="DeConv_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=3,
inFeatures=deConvFeatures3,
inNumFeatures=k * 8,
convRadius=1.6)
# Second upsampling
bnDeConvFeatures3 = batch_norm_RELU_drop_out("Up2_3_Reduce_In_BN",
deConvFeatures3, isTraining,
useConvDropOut, keepProbConv)
bnDeConvFeatures3 = conv_1x1("Up2_3_Reduce", bnDeConvFeatures3, k * 8,
k * 4)
bnDeConvFeatures3 = batch_norm_RELU_drop_out("Up2_3_Reduce_Out_BN",
bnDeConvFeatures3, isTraining,
useConvDropOut, keepProbConv)
upFeatures2_3 = mConvBuilder.create_convolution(
convName="Up_2_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=3,
outPointLevel=2,
inFeatures=bnDeConvFeatures3,
inNumFeatures=k * 4,
convRadius=0.8)
upFeatures2_3 = tf.concat([upFeatures2_3, convFeatures2], 1)
deConvFeatures2 = batch_norm_RELU_drop_out("DeConv_2_Reduce_In_BN",
upFeatures2_3, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures2 = conv_1x1("DeConv_2_Reduce", deConvFeatures2, k * 8,
k * 4)
deConvFeatures2 = batch_norm_RELU_drop_out("DeConv_2_Reduce_Out_BN",
deConvFeatures2, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures2 = mConvBuilder.create_convolution(
convName="DeConv_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
inFeatures=deConvFeatures2,
inNumFeatures=k * 4,
convRadius=0.8)
# First multiple upsamplings
bnDeConvFeatures2 = batch_norm_RELU_drop_out("Up1_2_Reduce_In_BN",
deConvFeatures2, isTraining,
useConvDropOut, keepProbConv)
bnDeConvFeatures2 = conv_1x1("Up1_2_Reduce", bnDeConvFeatures2, k * 4,
k * 2)
bnDeConvFeatures2 = batch_norm_RELU_drop_out("Up1_2_Reduce_Out_BN",
bnDeConvFeatures2, isTraining,
useConvDropOut, keepProbConv)
upFeatures1_2 = mConvBuilder.create_convolution(
convName="Up_1_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=1,
inFeatures=bnDeConvFeatures2,
inNumFeatures=k * 2,
convRadius=0.4)
bnDeConvFeatures3 = batch_norm_RELU_drop_out("Up1_3_Reduce_In_BN",
deConvFeatures3, isTraining,
useConvDropOut, keepProbConv)
bnDeConvFeatures3 = conv_1x1("Up1_3_Reduce", bnDeConvFeatures3, k * 8,
k * 2)
bnDeConvFeatures3 = batch_norm_RELU_drop_out("Up1_3_Reduce_Out_BN",
bnDeConvFeatures3, isTraining,
useConvDropOut, keepProbConv)
upFeatures1_3 = mConvBuilder.create_convolution(
convName="Up_1_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=3,
outPointLevel=1,
inFeatures=bnDeConvFeatures3,
inNumFeatures=k * 2,
convRadius=0.8)
bnDeConvFeatures4 = batch_norm_RELU_drop_out("Up1_4_Reduce_In_BN",
convFeatures4, isTraining,
useConvDropOut, keepProbConv)
bnDeConvFeatures4 = conv_1x1("Up1_4_Reduce", bnDeConvFeatures4, k * 16,
k * 2)
bnDeConvFeatures4 = batch_norm_RELU_drop_out("Up1_4_Reduce_Out_BN",
bnDeConvFeatures4, isTraining,
useConvDropOut, keepProbConv)
upFeatures1_4 = mConvBuilder.create_convolution(
convName="Up_1_4",
inPointHierarchy=mPointHierarchy,
inPointLevel=4,
outPointLevel=1,
inFeatures=bnDeConvFeatures4,
inNumFeatures=k * 2,
convRadius=1.6)
upFeatures1 = tf.concat(
[upFeatures1_4, upFeatures1_3, upFeatures1_2, convFeatures1], 1)
deConvFeatures1 = batch_norm_RELU_drop_out("DeConv_1_Reduce_In_BN",
upFeatures1, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures1 = conv_1x1("DeConv_1_Reduce", deConvFeatures1, k * 8,
k * 4)
deConvFeatures1 = batch_norm_RELU_drop_out("DeConv_1_Reduce_Out_BN",
deConvFeatures1, isTraining,
useConvDropOut, keepProbConv)
deConvFeatures1 = mConvBuilder.create_convolution(
convName="DeConv_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
inFeatures=deConvFeatures1,
inNumFeatures=k * 4,
convRadius=0.4)
deConvFeatures1 = tf.concat([
upFeatures1_4, upFeatures1_3, upFeatures1_2, convFeatures1,
deConvFeatures1
], 1)
# Final upsampling
upFeaturesFinal = batch_norm_RELU_drop_out("Up_Final_Reduce_In_BN",
deConvFeatures1, isTraining,
useConvDropOut, keepProbConv)
upFeaturesFinal = conv_1x1("Up_Final_Reduce", upFeaturesFinal, k * 12,
k * 4)
upFeaturesFinal = batch_norm_RELU_drop_out("Up_Final_Reduce_Out_BN",
upFeaturesFinal, isTraining,
useConvDropOut, keepProbConv)
finalFeatures = mConvBuilder.create_convolution(
convName="Up_0_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=0,
inFeatures=upFeaturesFinal,
inNumFeatures=k * 4,
convRadius=0.1)
# Fully connected MLP - Global features.
finalInput = batch_norm_RELU_drop_out("BNRELUDROP_hier_final",
finalFeatures, isTraining,
useConvDropOut, keepProbConv)
finalLogits = MLP_2_hidden(finalInput,
k * 4,
k * 4,
k * 2,
numSem,
"Final_Logits",
keepProbFull,
isTraining,
useDropOutFull,
useInitBN=False)
return finalLogits
def create_network(points,
features,
batchIds,
batchSize,
numInputFeatures,
k,
numOutCat,
numOutputs,
isTraining,
keepProbConv,
keepProbFull,
useConvDropOut=False,
useDropOutFull=True,
useRenorm=False,
BNMomentum=0.99,
activation='relu'):
if activation == 'relu':
from MCNetworkUtils import batch_norm_RELU_drop_out as batch_norm_activation_drop_out
elif activation == 'leakyrelu':
from MCNetworkUtils import batch_norm_leakyRELU_drop_out as batch_norm_activation_drop_out
elif activation == 'mish':
from MCNetworkUtils import batch_norm_mish_drop_out as batch_norm_activation_drop_out
else:
print('\nError: Unknown activation function: %s\n' % (activation))
mPointHierarchy = PointHierarchy(
points, features, batchIds,
[0.025, 0.1, 0.4, math.sqrt(3.0) + 0.1], "MCSphere", batchSize)
############################################ Convolutions
mConvBuilder = ConvolutionBuilder(KDEWindow=0.25)
# Zeroth Pooling
convFeatures0 = conv_1x1("Reduce_Pool_0", features, numInputFeatures,
k * 2)
convFeatures0 = batch_norm_activation_drop_out("Reduce_Pool_0_Out_BN",
convFeatures0, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
poolFeatures0 = mConvBuilder.create_convolution(
convName="Pool_0",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
outPointLevel=1,
inFeatures=convFeatures0,
inNumFeatures=k * 2,
convRadius=0.05,
KDEWindow=0.2)
# First Pooling
convFeatures1 = batch_norm_activation_drop_out("Reduce_Pool_1_In_BN",
poolFeatures0, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
convFeatures1 = conv_1x1("Reduce_Pool_1", convFeatures1, k * 2, k * 4)
convFeatures1 = batch_norm_activation_drop_out("Reduce_Pool_1_Out_BN",
convFeatures1, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
poolFeatures1 = mConvBuilder.create_convolution(
convName="Pool_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=2,
inFeatures=convFeatures1,
inNumFeatures=k * 4,
convRadius=0.2,
KDEWindow=0.2)
# Second Pooling
convFeatures2 = batch_norm_activation_drop_out("Reduce_Pool_2_In_BN",
poolFeatures1, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
convFeatures2 = conv_1x1("Reduce_Pool_2", convFeatures2, k * 4, k * 16)
convFeatures2 = batch_norm_activation_drop_out("Reduce_Pool_2_Out_BN",
convFeatures2, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
poolFeatures2 = mConvBuilder.create_convolution(
convName="Pool_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=3,
inFeatures=convFeatures2,
inNumFeatures=k * 16,
convRadius=0.3,
KDEWindow=0.2)
# Third Pooling
convFeatures3 = batch_norm_activation_drop_out("Reduce_Pool_3_In_BN",
poolFeatures2, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
convFeatures3 = conv_1x1("Reduce_Pool_3", convFeatures3, k * 16, k * 32)
convFeatures3 = batch_norm_activation_drop_out("Reduce_Pool_3_Out_BN",
convFeatures3, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
poolFeatures3 = mConvBuilder.create_convolution(
convName="Pool_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=3,
outPointLevel=4,
inFeatures=convFeatures3,
inNumFeatures=k * 32,
convRadius=math.sqrt(3.0) + 0.1,
KDEWindow=0.2)
#Image decoder - Global features.
encoderOutput = batch_norm_activation_drop_out("BNRELUDROP_finalencoder",
poolFeatures3, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
encoderOutput = tf.reshape(encoderOutput, [-1, 1, 1, k * 32])
finalPredictions = []
for vq_i in range(numOutputs):
deconvLayer0 = tf.keras.layers.Conv2DTranspose(
k * 16, [4, 4],
strides=(1, 1),
data_format='channels_last',
input_shape=[1, 1, k * 32],
name='deconvFeatures0' + str(vq_i))
deconvFeatures0 = deconvLayer0(encoderOutput)
deconvFeatures0 = resblock_2d(deconvFeatures0, k * 16, [3, 3], 2)
deconvFeatures0 = batch_norm_activation_drop_out(
"BNRELUDROP_deconv0" + str(vq_i), deconvFeatures0, isTraining,
useConvDropOut, keepProbConv, useRenorm, BNMomentum)
deconvLayer1 = tf.keras.layers.Conv2DTranspose(
k * 4, [4, 4],
strides=(4, 4),
data_format='channels_last',
input_shape=[4, 4, k * 16],
name='deconvFeatures0' + str(vq_i))
deconvFeatures1 = deconvLayer1(deconvFeatures0)
deconvFeatures1 = resblock_2d(deconvFeatures1, k * 4, [3, 3], 2)
deconvFeatures1 = batch_norm_activation_drop_out(
"BNRELUDROP_deconv1" + str(vq_i), deconvFeatures1, isTraining,
useConvDropOut, keepProbConv, useRenorm, BNMomentum)
deconvLayer2 = tf.keras.layers.Conv2DTranspose(
1, [2, 2],
strides=(2, 2),
data_format='channels_last',
input_shape=[16, 16, k * 4],
name='deconvFeatures0' + str(vq_i))
predicted_images = deconvLayer2(deconvFeatures1)
# predicted_images = tf.squeeze(predicted_images)
finalPredictions.append(tf.reshape(predicted_images, [-1, 32, 32]))
return finalPredictions
def create_network(points,
features,
batchIds,
batchSize,
numInputFeatures,
k,
numOutCat,
numOutputs,
isTraining,
keepProbConv,
keepProbFull,
useConvDropOut=False,
useDropOutFull=True,
useRenorm=False,
BNMomentum=0.99,
activation='relu'):
if activation == 'relu':
from MCNetworkUtils import batch_norm_RELU_drop_out as batch_norm_activation_drop_out
elif activation == 'leakyrelu':
from MCNetworkUtils import batch_norm_leakyRELU_drop_out as batch_norm_activation_drop_out
elif activation == 'mish':
from MCNetworkUtils import batch_norm_mish_drop_out as batch_norm_activation_drop_out
else:
print('\nError: Unknown activation function: %s\n' % (activation))
mPointHierarchy = PointHierarchy(
points, features, batchIds,
[0.025, 0.1, 0.4, math.sqrt(3.0) + 0.1], "MCSphere", batchSize)
############################################ Convolutions
mConvBuilder = ConvolutionBuilder(KDEWindow=0.25)
# Zeroth Pooling
convFeatures0 = conv_1x1("Reduce_Pool_0", features, numInputFeatures,
k * 2)
convFeatures0 = batch_norm_activation_drop_out("Reduce_Pool_0_Out_BN",
convFeatures0, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
poolFeatures0 = mConvBuilder.create_convolution(
convName="Pool_0",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
outPointLevel=1,
inFeatures=convFeatures0,
inNumFeatures=k * 2,
convRadius=0.05,
KDEWindow=0.2)
# First Pooling
convFeatures1 = batch_norm_activation_drop_out("Reduce_Pool_1_In_BN",
poolFeatures0, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
convFeatures1 = conv_1x1("Reduce_Pool_1", convFeatures1, k * 2, k * 4)
convFeatures1 = batch_norm_activation_drop_out("Reduce_Pool_1_Out_BN",
convFeatures1, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
poolFeatures1 = mConvBuilder.create_convolution(
convName="Pool_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=2,
inFeatures=convFeatures1,
inNumFeatures=k * 4,
convRadius=0.2,
KDEWindow=0.2)
# Second Pooling
convFeatures2 = batch_norm_activation_drop_out("Reduce_Pool_2_In_BN",
poolFeatures1, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
convFeatures2 = conv_1x1("Reduce_Pool_2", convFeatures2, k * 4, k * 16)
convFeatures2 = batch_norm_activation_drop_out("Reduce_Pool_2_Out_BN",
convFeatures2, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
poolFeatures2 = mConvBuilder.create_convolution(
convName="Pool_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=3,
inFeatures=convFeatures2,
inNumFeatures=k * 16,
convRadius=0.3,
KDEWindow=0.2)
# Third Pooling
convFeatures3 = batch_norm_activation_drop_out("Reduce_Pool_3_In_BN",
poolFeatures2, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
convFeatures3 = conv_1x1("Reduce_Pool_3", convFeatures3, k * 16, k * 32)
convFeatures3 = batch_norm_activation_drop_out("Reduce_Pool_3_Out_BN",
convFeatures3, isTraining,
useConvDropOut,
keepProbConv, useRenorm,
BNMomentum)
poolFeatures3 = mConvBuilder.create_convolution(
convName="Pool_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=3,
outPointLevel=4,
inFeatures=convFeatures3,
inNumFeatures=k * 32,
convRadius=math.sqrt(3.0) + 0.1,
KDEWindow=0.2)
#Fully connected MLP - Global features.
finalInput = batch_norm_activation_drop_out("BNRELUDROP_final",
poolFeatures3, isTraining,
useConvDropOut, keepProbConv,
useRenorm, BNMomentum)
final = []
for vq_i in range(numOutputs):
Predictions = MLP_2_hidden(finalInput,
k * 32,
k * 16,
k * 8,
numOutCat,
"Final_Predictions" + str(vq_i),
keepProbFull,
isTraining,
useDropOutFull,
useRenorm=useRenorm,
BNMomentum=BNMomentum)
final.append(Predictions)
finalPredictions = tf.reshape(tf.stack(final, axis=1),
[-1, 3 * numOutputs])
return finalPredictions
def create_network(points,
batchIds,
features,
points2,
batchIds2,
features2,
outputs,
batchSize,
k,
isTraining,
bnMomentum,
brnMomentum,
brnClipping,
keepProbConv,
keepProbFull,
useConvDropOut=False,
useDropOutFull=True,
dataset=0):
def BN_NL_DP_Conv(layerName, inFeatures):
inFeatures = tf.layers.batch_normalization(inputs=inFeatures,
momentum=bnMomentum,
trainable=True,
training=isTraining,
name=layerName + "_BN",
renorm=True,
renorm_clipping=brnClipping,
renorm_momentum=brnMomentum)
inFeatures = tf.nn.leaky_relu(inFeatures)
if useConvDropOut:
inFeatures = tf.nn.dropout(inFeatures, keepProbConv)
return inFeatures
def BN_NL_DP_F(layerName, inFeatures):
inFeatures = tf.layers.batch_normalization(inputs=inFeatures,
momentum=bnMomentum,
trainable=True,
training=isTraining,
name=layerName + "_BN",
renorm=True,
renorm_clipping=brnClipping,
renorm_momentum=brnMomentum)
inFeatures = tf.nn.leaky_relu(inFeatures)
if useDropOutFull:
inFeatures = tf.nn.dropout(inFeatures, keepProbFull)
return inFeatures
############################################ Compute point hierarchy
mPointHierarchy1 = PointHierarchy(points,
features,
batchIds, [0.05, 0.1, 0.2],
"MCGI_PH1",
batchSize,
relativeRadius=False)
mPointHierarchy2 = PointHierarchy(points2,
features2,
batchIds2, [],
"MCGI_PH2",
batchSize,
relativeRadius=False)
############################################ Convolutions
mConvBuilder = ConvolutionBuilder(KDEWindow=0.25, relativeRadius=False)
resultColor = []
with tf.variable_scope("feature_channel_scope", reuse=tf.AUTO_REUSE):
for i in range(3):
if dataset == 1:
numInFeatures = 5
inputFeatures = features[:, 0:3]
inputFeatures = tf.concat(
[inputFeatures, features[:, 3 + i:3 + i + 1]], axis=1)
inputFeatures = tf.concat(
[inputFeatures, features[:, 6 + i:6 + i + 1]], axis=1)
inputFeatures2 = features2[:, 0:3]
elif dataset == 2:
numInFeatures = 7
inputFeatures = features[:, 0:3]
inputFeatures = tf.concat(
[inputFeatures, features[:, 3 + i:3 + i + 1]], axis=1)
inputFeatures = tf.concat(
[inputFeatures, features[:, 6 + i:6 + i + 1]], axis=1)
inputFeatures = tf.concat(
[inputFeatures, features[:, 9 + i:9 + i + 1]], axis=1)
inputFeatures = tf.concat([inputFeatures, features[:, 12:13]],
axis=1)
inputFeatures2 = features2[:, 0:3]
inputFeatures2 = tf.concat(
[inputFeatures2, features2[:, 3 + i:3 + i + 1]], axis=1)
inputFeatures2 = tf.concat(
[inputFeatures2, features2[:, 6 + i:6 + i + 1]], axis=1)
inputFeatures2 = tf.concat(
[inputFeatures2, features2[:, 9 + i:9 + i + 1]], axis=1)
inputFeatures2 = tf.concat(
[inputFeatures2, features2[:, 12:13]], axis=1)
else:
numInFeatures = 3
inputFeatures = features
inputFeatures2 = features2
############################################ Encoder
# First convolution
convFeatures1 = mConvBuilder.create_convolution(
convName="Conv_1",
inPointHierarchy=mPointHierarchy1,
inPointLevel=0,
inFeatures=inputFeatures,
inNumFeatures=numInFeatures,
outNumFeatures=k,
convRadius=0.025,
multiFeatureConv=True)
# First pooling
bnConvFeatures1 = BN_NL_DP_Conv("Reduce_Pool_1_In_BN",
convFeatures1)
bnConvFeatures1 = conv_1x1("Reduce_Pool_1", bnConvFeatures1, k,
k * 2)
bnConvFeatures1 = BN_NL_DP_Conv("Reduce_Pool_1_Out_BN",
bnConvFeatures1)
poolFeatures1 = mConvBuilder.create_convolution(
convName="Pool_1",
inPointHierarchy=mPointHierarchy1,
inPointLevel=0,
outPointLevel=1,
inFeatures=bnConvFeatures1,
inNumFeatures=k * 2,
convRadius=0.05,
KDEWindow=0.2)
# Second convolution
bnPoolFeatures1 = BN_NL_DP_Conv("Reduce_Conv_2_In_BN",
poolFeatures1)
bnPoolFeatures1 = conv_1x1("Reduce_Conv_2", bnPoolFeatures1, k * 2,
k * 2)
bnPoolFeatures1 = BN_NL_DP_Conv("Reduce_Conv_2_Out_BN",
bnPoolFeatures1)
convFeatures2 = mConvBuilder.create_convolution(
convName="Conv_2",
inPointHierarchy=mPointHierarchy1,
inPointLevel=1,
inFeatures=bnPoolFeatures1,
inNumFeatures=k * 2,
convRadius=0.1)
convFeatures2 = tf.concat([poolFeatures1, convFeatures2], 1)
# Second pooling
bnConvFeatures2 = BN_NL_DP_Conv("Reduce_Pool_2_In_BN",
convFeatures2)
bnConvFeatures2 = conv_1x1("Reduce_Pool_2", bnConvFeatures2, k * 4,
k * 4)
bnConvFeatures2 = BN_NL_DP_Conv("Reduce_Pool_2_Out_BN",
bnConvFeatures2)
poolFeatures2 = mConvBuilder.create_convolution(
convName="Pool_2",
inPointHierarchy=mPointHierarchy1,
inPointLevel=1,
outPointLevel=2,
inFeatures=bnConvFeatures2,
inNumFeatures=k * 4,
convRadius=0.2,
KDEWindow=0.2)
# Third convolution
bnPoolFeatures2 = BN_NL_DP_Conv("Reduce_Conv_3_In_BN",
poolFeatures2)
bnPoolFeatures2 = conv_1x1("Reduce_Conv_3", bnPoolFeatures2, k * 4,
k * 4)
bnPoolFeatures2 = BN_NL_DP_Conv("Reduce_Conv_3_Out_BN",
bnPoolFeatures2)
convFeatures3 = mConvBuilder.create_convolution(
convName="Conv_3",
inPointHierarchy=mPointHierarchy1,
inPointLevel=2,
inFeatures=bnPoolFeatures2,
inNumFeatures=k * 4,
convRadius=0.2)
convFeatures3 = tf.concat([poolFeatures2, convFeatures3], 1)
# Third pooling
bnConvFeatures3 = BN_NL_DP_Conv("Reduce_Pool_3_In_BN",
convFeatures3)
bnConvFeatures3 = conv_1x1("Reduce_Pool_3", bnConvFeatures3, k * 8,
k * 8)
bnConvFeatures3 = BN_NL_DP_Conv("Reduce_Pool_3_Out_BN",
bnConvFeatures3)
poolFeatures3 = mConvBuilder.create_convolution(
convName="Pool_3",
inPointHierarchy=mPointHierarchy1,
inPointLevel=2,
outPointLevel=3,
inFeatures=bnConvFeatures3,
inNumFeatures=k * 8,
convRadius=0.4,
KDEWindow=0.2)
# Fourth convolution
bnPoolFeatures3 = BN_NL_DP_Conv("Reduce_Conv_4_In_BN",
poolFeatures3)
bnPoolFeatures3 = conv_1x1("Reduce_Conv_4", bnPoolFeatures3, k * 8,
k * 8)
bnPoolFeatures3 = BN_NL_DP_Conv("Reduce_Conv_4_Out_BN",
bnPoolFeatures3)
convFeatures4 = mConvBuilder.create_convolution(
convName="Conv_4",
inPointHierarchy=mPointHierarchy1,
inPointLevel=3,
inFeatures=bnPoolFeatures3,
inNumFeatures=k * 8,
convRadius=1.0,
KDEWindow=0.2)
############################################ Decoder
# Third upsampling
bnConvFeatures4 = BN_NL_DP_Conv("Up_3_BN", convFeatures4)
upFeatures3 = mConvBuilder.create_convolution(
convName="Up_3",
inPointHierarchy=mPointHierarchy1,
inPointLevel=3,
outPointLevel=2,
inFeatures=bnConvFeatures4,
inNumFeatures=k * 8,
convRadius=0.4,
KDEWindow=0.2)
upFeatures3 = tf.concat([convFeatures3, upFeatures3], 1)
upFeatures3 = BN_NL_DP_Conv("Up_3_Reduce_BN", upFeatures3)
upFeatures3 = conv_1x1("Up_3_Reduce", upFeatures3, k * 16, k * 4)
# Second upsampling
bnUpFeatures3 = BN_NL_DP_Conv("Up_2_BN", upFeatures3)
upFeatures2 = mConvBuilder.create_convolution(
convName="Up_2",
inPointHierarchy=mPointHierarchy1,
inPointLevel=2,
outPointLevel=1,
inFeatures=bnUpFeatures3,
inNumFeatures=k * 4,
convRadius=0.2,
KDEWindow=0.2)
upFeatures2 = tf.concat([convFeatures2, upFeatures2], 1)
upFeatures2 = BN_NL_DP_Conv("Up_2_Reduce_BN", upFeatures2)
upFeatures2 = conv_1x1("Up_2_Reduce", upFeatures2, k * 8, k * 2)
# First upsampling
bnUpFeatures2 = BN_NL_DP_Conv("Up_1_2_BN", upFeatures2)
upFeatures1 = mConvBuilder.create_convolution(
convName="Up_1_2",
inPointHierarchy=mPointHierarchy1,
inPointLevel=1,
outPointLevel=0,
inFeatures=bnUpFeatures2,
inNumFeatures=k * 2,
convRadius=0.1,
KDEWindow=0.2)
upFeatures1 = tf.concat([convFeatures1, upFeatures1], 1)
upFeatures1 = BN_NL_DP_Conv("Up_1_Reduce_BN", upFeatures1)
upFeatures1 = conv_1x1("Up_1_Reduce", upFeatures1, k * 3, k)
bnUpFeatures1 = BN_NL_DP_Conv("Up_1_BN", upFeatures1)
finalFeatures = mConvBuilder.create_convolution(
convName="Final_Conv",
inPointHierarchy=mPointHierarchy1,
outPointHierarchy=mPointHierarchy2,
inPointLevel=0,
outPointLevel=0,
inFeatures=bnUpFeatures1,
inNumFeatures=k,
convRadius=0.05,
KDEWindow=0.2)
finalFeatures = BN_NL_DP_F("Final_MLP1_BN", finalFeatures)
finalFeatures = tf.concat([finalFeatures, inputFeatures2], 1)
finalFeatures = conv_1x1("Final_MLP1", finalFeatures,
k + numInFeatures, k)
finalFeatures = BN_NL_DP_F("Final_MLP2_BN", finalFeatures)
predVals = conv_1x1("Final_MLP2", finalFeatures, k, 1)
resultColor.append(predVals)
return tf.concat(resultColor, axis=1)
def create_network(points,
batchIds,
features,
numInputFeatures,
batchSize,
k,
numOutCat,
isTraining,
keepProbConv,
keepProbFull,
useConvDropOut=False,
useDropOutFull=True):
############################################ Compute point hierarchy
mPointHierarchy = PointHierarchy(points, features, batchIds,
[0.1, 0.4, math.sqrt(3.0) + 0.1],
"MCClassH_PH", batchSize)
############################################ Convolutions
mConvBuilder = ConvolutionBuilder(KDEWindow=0.25)
############################################ LOGITS 1
############################################ First level
# Convolution
convFeatures1 = mConvBuilder.create_convolution(
convName="Conv_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
inFeatures=features,
inNumFeatures=numInputFeatures,
outNumFeatures=k,
convRadius=0.1,
multiFeatureConv=True)
# Pooling
convFeatures1 = batch_norm_RELU_drop_out("Reduce_Pool_1_In_BN",
convFeatures1, isTraining,
useConvDropOut, keepProbConv)
convFeatures1 = conv_1x1("Reduce_Pool_1", convFeatures1, k, k * 2)
convFeatures1 = batch_norm_RELU_drop_out("Reduce_Pool_1_Out_BN",
convFeatures1, isTraining,
useConvDropOut, keepProbConv)
poolFeatures1 = mConvBuilder.create_convolution(
convName="Pool_1",
inPointHierarchy=mPointHierarchy,
inPointLevel=0,
outPointLevel=1,
inFeatures=convFeatures1,
inNumFeatures=k * 2,
convRadius=0.2,
KDEWindow=0.2)
############################################ Second level convolutions
#### Convolution
bnPoolFeatures1 = batch_norm_RELU_drop_out("Reduce_Conv_2_In_BN",
poolFeatures1, isTraining,
useConvDropOut, keepProbConv)
bnPoolFeatures1 = conv_1x1("Reduce_Conv_2", bnPoolFeatures1, k * 2, k * 2)
bnPoolFeatures1 = batch_norm_RELU_drop_out("Reduce_Conv_2_Out_BN",
bnPoolFeatures1, isTraining,
useConvDropOut, keepProbConv)
convFeatures2 = mConvBuilder.create_convolution(
convName="Conv_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
inFeatures=bnPoolFeatures1,
inNumFeatures=k * 2,
convRadius=0.4)
convFeatures2 = tf.concat([poolFeatures1, convFeatures2], 1)
# Pooling
convFeatures2 = batch_norm_RELU_drop_out("Reduce_Pool_2_In_BN",
convFeatures2, isTraining,
useConvDropOut, keepProbConv)
convFeatures2 = conv_1x1("Reduce_Pool_2", convFeatures2, k * 4, k * 8)
convFeatures2 = batch_norm_RELU_drop_out("Reduce_Pool_2_Out_BN",
convFeatures2, isTraining,
useConvDropOut, keepProbConv)
poolFeatures2 = mConvBuilder.create_convolution(
convName="Pool_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=2,
inFeatures=convFeatures2,
inNumFeatures=k * 8,
convRadius=0.8,
KDEWindow=0.2)
############################################ Third level convolutions
# Convolution
bnPoolFeatures2 = batch_norm_RELU_drop_out("Reduce_Conv_3_In_BN",
poolFeatures2, isTraining,
useConvDropOut, keepProbConv)
bnPoolFeatures2 = conv_1x1("Reduce_Conv_3", bnPoolFeatures2, k * 8, k * 8)
bnPoolFeatures2 = batch_norm_RELU_drop_out("Reduce_Conv_3_Out_BN",
bnPoolFeatures2, isTraining,
useConvDropOut, keepProbConv)
convFeatures3 = mConvBuilder.create_convolution(
convName="Conv_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
inFeatures=bnPoolFeatures2,
inNumFeatures=k * 8,
convRadius=1.2)
convFeatures3 = tf.concat([poolFeatures2, convFeatures3], 1)
# Pooling
convFeatures3 = batch_norm_RELU_drop_out("Reduce_Pool_3_In_BN",
convFeatures3, isTraining,
useConvDropOut, keepProbConv)
convFeatures3 = conv_1x1("Reduce_Pool_3", convFeatures3, k * 16, k * 32)
convFeatures3 = batch_norm_RELU_drop_out("Reduce_Pool_3_Out_BN",
convFeatures3, isTraining,
useConvDropOut, keepProbConv)
poolFeatures3 = mConvBuilder.create_convolution(
convName="Pool_3",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=3,
inFeatures=convFeatures3,
inNumFeatures=k * 32,
convRadius=math.sqrt(3.0) + 0.1,
KDEWindow=0.2)
#Fully connected MLP - Global features.
finalInput = batch_norm_RELU_drop_out("BNRELUDROP_final", poolFeatures3,
isTraining, useConvDropOut,
keepProbConv)
finalLogits1 = MLP_2_hidden(finalInput, k * 32, k * 16, k * 8, numOutCat,
"Final_Logits", keepProbFull, isTraining,
useDropOutFull)
############################################ LOGITS 2
############################################ Second level convolutions
#### Convolution
convFeatures22 = mConvBuilder.create_convolution(
convName="Conv_2_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
inFeatures=mPointHierarchy.features_[1],
inNumFeatures=numInputFeatures,
outNumFeatures=k * 2,
convRadius=0.4,
multiFeatureConv=True)
# Pooling
convFeatures22 = batch_norm_RELU_drop_out("Reduce_Pool_2_2_In_BN",
convFeatures22, isTraining,
useConvDropOut, keepProbConv)
convFeatures22 = conv_1x1("Reduce_Pool_2_2", convFeatures22, k * 2, k * 8)
convFeatures22 = batch_norm_RELU_drop_out("Reduce_Pool_2_2_Out_BN",
convFeatures22, isTraining,
useConvDropOut, keepProbConv)
poolFeatures22 = mConvBuilder.create_convolution(
convName="Pool_2_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=1,
outPointLevel=2,
inFeatures=convFeatures22,
inNumFeatures=k * 8,
convRadius=0.8,
KDEWindow=0.2)
############################################ Third level convolutions
# Convolution
bnPoolFeatures22 = batch_norm_RELU_drop_out("Reduce_Conv_3_2_In_BN",
poolFeatures22, isTraining,
useConvDropOut, keepProbConv)
bnPoolFeatures22 = conv_1x1("Reduce_Conv_3_2", bnPoolFeatures22, k * 8,
k * 8)
bnPoolFeatures22 = batch_norm_RELU_drop_out("Reduce_Conv_3_2_Out_BN",
bnPoolFeatures22, isTraining,
useConvDropOut, keepProbConv)
convFeatures32 = mConvBuilder.create_convolution(
convName="Conv_3_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
inFeatures=bnPoolFeatures22,
inNumFeatures=k * 8,
convRadius=1.2)
convFeatures32 = tf.concat([poolFeatures22, convFeatures32], 1)
# Pooling
convFeatures32 = batch_norm_RELU_drop_out("Reduce_Pool_3_2_In_BN",
convFeatures32, isTraining,
useConvDropOut, keepProbConv)
convFeatures32 = conv_1x1("Reduce_Pool_3_2", convFeatures32, k * 16,
k * 32)
convFeatures32 = batch_norm_RELU_drop_out("Reduce_Pool_3_2_Out_BN",
convFeatures32, isTraining,
useConvDropOut, keepProbConv)
poolFeatures32 = mConvBuilder.create_convolution(
convName="Pool_3_2",
inPointHierarchy=mPointHierarchy,
inPointLevel=2,
outPointLevel=3,
inFeatures=convFeatures32,
inNumFeatures=k * 32,
convRadius=math.sqrt(3.0) + 0.1,
KDEWindow=0.2)
#Fully connected MLP - Global features.
finalInput2 = batch_norm_RELU_drop_out("2BNRELUDROP_final", poolFeatures32,
isTraining, useConvDropOut,
keepProbConv)
finalLogits2 = MLP_2_hidden(finalInput2, k * 32, k * 16, k * 8, numOutCat,
"2Final_Logits", keepProbFull, isTraining,
useDropOutFull)
############################################ PATH DROPOUT
counter = tf.constant(0.0, dtype=tf.float32)
probability = tf.random_uniform([1])
mask1 = tf.less_equal(probability[0], tf.constant(0.66))
mask1 = tf.maximum(tf.cast(mask1, tf.float32),
tf.cast(tf.logical_not(isTraining), tf.float32))
counter = tf.add(counter, mask1)
finalLogits1 = tf.scalar_mul(mask1, finalLogits1)
mask2 = tf.greater_equal(probability[0], tf.constant(0.33))
mask2 = tf.maximum(tf.cast(mask2, tf.float32),
tf.cast(tf.logical_not(isTraining), tf.float32))
counter = tf.add(counter, mask2)
finalLogits2 = tf.scalar_mul(mask2, finalLogits2)
counter = tf.multiply(tf.constant((2.0), dtype=tf.float32),
tf.reciprocal(counter))
return tf.scalar_mul(counter, tf.add(finalLogits1, finalLogits2))