def TruncatedNormal(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict,
extraNodeInfoDict: dict):
curNodeDataType = curNode.getAttrMapRef()["\"dtype\""].getDataType()
assert (curNodeDataType is not Graph.DataTypeEnum.DT_INVALID)
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 1)
curNodeOutputShape = extraNodeInfoDict[curNode.getName()][0]
return (None,
AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.TruncatedNormal.name,
[AST.ID(curNodeDataType.name)] +
list(map(lambda x: AST.Int(x), curNodeOutputShape)))
) # TODO
def Reshape(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 2)
return (None,
AST.Reshape(AST.ID(dictNodeNameToOutVarStr[inputsRef[0]]),
extraNodeInfoDict[curNode.getName()][0], None))
def ArgMax(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 2)
return (None,
AST.ArgMax(extraNodeInfoDict[curNode.getName()][0],
AST.ID(dictNodeNameToOutVarStr[inputsRef[0]]),
AST.ID(dictNodeNameToOutVarStr[inputsRef[1]]),
extraNodeInfoDict[inputsRef[0]][0]))
def ExpandDims(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 2)
retAST = AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.ExpandDims.name,
list(map(lambda x: AST.ID(dictNodeNameToOutVarStr[x]), inputsRef)))
return (None, retAST)
def Slice(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 3)
curNodeDataType = curNode.getAttrMapRef()["\"T\""].getDataType()
curNodeShapeASTLi = list(
map(lambda x: AST.Int(x), extraNodeInfoDict[curNode.getName()][0]))
retAST = AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.CreateCopy.name,
[
AST.ID(dictNodeNameToOutVarStr[inputsRef[0]]), # of this
# begin idx
AST.ID(dictNodeNameToOutVarStr[inputsRef[1]]),
# size
AST.ID(dictNodeNameToOutVarStr[inputsRef[2]])
])
return (None, retAST)
def Placeholder(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
#curNodeShapeLi = curNode.getAttrMapRef()["\"shape\""].getShape().getDimRef()
curNodeShapeLi = extraNodeInfoDict[curNode.getName()][0]
curNodeInputType = curNode.getAttrMapRef()["\"dtype\""].getDataType()
assert (curNodeInputType is not Graph.DataTypeEnum.DT_INVALID)
# TODO : There has to be some way to take range, understand the dimensions for SeeDot
# CHANGESRI
# return (None, AST.Input(curNodeShapeLi, curNodeInputType.name))
return (None, AST.Decl(curNodeShapeLi, (-0.1, 0.1)))
def MaxPoolGrad(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
return (None,
AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.MaxPoolGrad.name,
list(
map(lambda x: AST.ID(dictNodeNameToOutVarStr[x]),
inputsRef))))
def Tile(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 2)
return (None,
AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.Tile.name, [
AST.ID(dictNodeNameToOutVarStr[inputsRef[0]]),
AST.ID(dictNodeNameToOutVarStr[inputsRef[1]])
]))
def ZerosLike(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 1)
curNodeOutputType = curNode.getAttrMapRef()["\"T\""].getDataType()
assert (curNodeOutputType is not Graph.DataTypeEnum.DT_INVALID)
retAST = AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.CreateTensor.name, [AST.Int(0)],
isSecret=False)
return (None, retAST)
def Pack(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
N = curNode.getAttrMapRef()["\"N\""].getI()
axis = curNode.getAttrMapRef()["\"axis\""].getI()
assert (len(inputsRef) == N)
retAST = AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.Pack.name,
list(map(lambda x: AST.ID(dictNodeNameToOutVarStr[x]),
inputsRef)) + [AST.Int(axis)])
return (None, retAST)
def RandomUniform(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
curNodeDataType = curNode.getAttrMapRef()["\"dtype\""].getDataType()
assert (curNodeDataType is not Graph.DataTypeEnum.DT_INVALID)
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 1)
curNodeOutputShape = extraNodeInfoDict[curNode.getName()][0]
return (None,
AST.UninterpFuncCall(
curNodeOutputShape,
TFNodesAST.UninterpFuncCallNames.RandomUniform.name,
[AST.ID(curNodeDataType.name)]))
def Cast(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 1)
sourceType = curNode.getAttrMapRef()["\"SrcT\""].getDataType()
destType = curNode.getAttrMapRef()["\"DstT\""].getDataType()
return (None,
AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.Cast.name, [
AST.ID(dictNodeNameToOutVarStr[inputsRef[0]]),
AST.ID(sourceType.name),
AST.ID(destType.name)
]))
def ConcatV2(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
N = curNode.getAttrMapRef()["\"N\""].getI()
assert (len(inputsRef) == N + 1) # One extra for axis
# TODO : Since the axis of concat is constant, therefore, its known here - the input's sizes along that dim should be
# passed as input to the below function.
# For now hardcoding.
retAST = AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.Concat.name + str(N) + 'T',
list(map(lambda x: AST.ID(dictNodeNameToOutVarStr[x]), inputsRef)),
outputDiffInpDims=1)
return (None, retAST)
def Assign(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 2)
curNodeShape = extraNodeInfoDict[curNode.getName()][0]
# TODO_TAB : for inference, have commented the copyTensor function calls.
# TODO : Hack -- fix this later after discussing with Aseem
# return (None, AST.UninterpFuncCall(curNodeShape,
# TFNodesAST.UninterpFuncCallNames.CopyTensor.name,
# [AST.ID(dictNodeNameToOutVarStr[inputsRef[0]]),
# AST.ID(dictNodeNameToOutVarStr[inputsRef[1]])]))
return (None, None)
def Identity(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
# In SeeDot, J2=J1 creates a new reference for J1 -- so
# the corresponding code in Seedot cannot simply be J2 = J1.
# Instead create a new tensor first and then assign the old one to the new one.
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 1)
curNodeDataType = curNode.getAttrMapRef()["\"T\""].getDataType()
assert (curNodeDataType is not Graph.DataTypeEnum.DT_INVALID)
curNodeShape = extraNodeInfoDict[curNode.getName()][0]
retAST = AST.UninterpFuncCall(
curNodeShape, TFNodesAST.UninterpFuncCallNames.CreateIdentity.name,
[AST.ID(dictNodeNameToOutVarStr[inputsRef[0]])])
return (None, retAST)
def Squeeze(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
# TODO : Do this in somewhat better way
inputsRef = curNode.getInputsRef()
inputTensorShape = extraNodeInfoDict[inputsRef[0]][0]
inputTensorRank = len(inputTensorShape)
squeezeDims = curNode.getAttrMapRef()["\"squeeze_dims\""].getList(
).getILi()
squeezeDimsRank = len(squeezeDims)
return (None,
AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.Squeeze.name,
list(map(lambda x: AST.Int(x, 32), squeezeDims)) +
[AST.ID(dictNodeNameToOutVarStr[inputsRef[0]])]))
def Fill(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 2)
curNodeOutputShape = extraNodeInfoDict[inputsRef[0]][0]
# inputsRef[0] denotes a shape and should have a rank of 1
assert (len(curNodeOutputShape) == 1)
curNodeOutputType = curNode.getAttrMapRef()["\"T\""].getDataType()
assert (curNodeOutputType is not Graph.DataTypeEnum.DT_INVALID)
retAST = AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.CreateTensor.name,
[AST.ID(dictNodeNameToOutVarStr[inputsRef[1]])],
isSecret=False)
return (None, retAST)
def AvgPool(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
inputsRef = curNode.getInputsRef()
assert (len(inputsRef) == 1)
options = {}
stridesUsed = curNode.getAttrMapRef()["\"strides\""].getList().getILi()
assert ((stridesUsed[0] == 1) and (stridesUsed[3] == 1))
strideH = stridesUsed[1]
strideW = stridesUsed[2]
kSizeUsed = curNode.getAttrMapRef()["\"ksize\""].getList().getILi()
assert ((kSizeUsed[0] == 1) and (kSizeUsed[3] == 1))
kSizeH = kSizeUsed[1]
kSizeW = kSizeUsed[2]
paddingUsedStr = curNode.getAttrMapRef()["\"padding\""].getS()
zPadH = zPadW = -1
if (paddingUsedStr == "\"SAME\""):
zPadH = int((kSizeH - 1) / 2)
zPadW = int((kSizeW - 1) / 2)
elif (paddingUsedStr == "\"VALID\""):
zPadH = zPadW = 0
else:
zPadH = zPadW = -1
inputShape = extraNodeInfoDict[inputsRef[0]][0]
imgH = inputShape[1]
imgW = inputShape[2]
return (None,
AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.AvgPool.name, [
AST.Int(kSizeH, 32),
AST.Int(kSizeW, 32),
AST.Int(zPadH, 32),
AST.Int(zPadW, 32),
AST.Int(strideH, 32),
AST.Int(strideW, 32),
AST.ID(dictNodeNameToOutVarStr[inputsRef[0]])
]))
def FusedBatchNorm(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
# NOTE : Since the weights to this layer will be scaled appropriately, this op will become identity.
inputsRef = curNode.getInputsRef()
# TODO : This below thing is the right way of implementing the operator
# For now using uninterpreted function call.
# tempAst = AST.BOp(AST.ID(dictNodeNameToOutVarStr[inputsRef[0]]),
# TFNodesAST.getOperatorsIdx('*'),
# AST.ID(dictNodeNameToOutVarStr[inputsRef[1]])
# )
# return (None, AST.BOp(tempAst,
# TFNodesAST.getOperatorsIdx('+'),
# AST.ID(dictNodeNameToOutVarStr[inputsRef[2]])
# ))
return (None,
AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.TempFusedBatchNorm.name, [
AST.ID(dictNodeNameToOutVarStr[inputsRef[0]]),
AST.ID(dictNodeNameToOutVarStr[inputsRef[1]]),
AST.ID(dictNodeNameToOutVarStr[inputsRef[2]]),
]))
def Pad(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
# Mode refers to 'CONSTANT', 'REFLECT' or 'SYMMETRIC'
mode = 0
if ("\"mode\"" in curNode.getAttrMapRef()):
mode = curNode.getAttrMapRef()["\"mode\""].getI()
constant_values = 0
if ("\"constant_values\"" in curNode.getAttrMapRef()):
constant_values = curNode.getAttrMapRef(
)["\"constant_values\""].getI()
# For now to make life easy - deal with SYMMETRIC AND REFLECT when time comes
assert (mode == 0 and constant_values == 0)
inputsRef = curNode.getInputsRef()
inputTensorShapeLi = extraNodeInfoDict[inputsRef[0]][0]
return (None,
AST.UninterpFuncCall(
extraNodeInfoDict[curNode.getName()][0],
TFNodesAST.UninterpFuncCallNames.Pad.name, [
AST.ID(dictNodeNameToOutVarStr[inputsRef[0]]),
AST.ID(dictNodeNameToOutVarStr[inputsRef[1]])
],
outputDiffInpDims=1))