def visitDecl(self, ctx: SeeDotParser.DeclContext):
shape = [
int(IntConst.getText())
for IntConst in ctx.intConstList().IntConst()
]
range = float(ctx.FloatConst(0).getText()), float(
ctx.FloatConst(1).getText())
return AST.Decl(shape, range)
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 VariableV2(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
curNodeShapeLi = curNode.getAttrMapRef()["\"shape\""].getShape(
).getDimRef()[:]
curNodeInputType = curNode.getAttrMapRef()["\"dtype\""].getDataType()
# TODO_TAB : for inference, have commented out decl and inserted input nodes.
# TODO : Right now in the current implementation, the dataType being passed to the node is being ignored by SeeDot.
# Fix this later.
# return (None, AST.Decl(curNodeShapeLi, curNodeInputType.name, None, None))
# NOTE : since this becomes an input node right now, i have also added to be prefixed at top in ProcessTFGraph::prefixAllPlaceHolderNodes()
# CHANGESRI
# return (None, AST.Input(curNodeShapeLi, curNodeInputType.name))
return (None, AST.Decl(curNodeShapeLi, [0.1, 0.1]))
def Const(graph: Graph.Graph, curNode: Graph.Node,
dictNodeNameToOutVarStr: dict, extraNodeInfoDict: dict):
assert (len(curNode.getInputsRef()) == 0)
tensor = curNode.getAttrMapRef()["\"value\""].getTensor()
curNodeDataType = curNode.getAttrMapRef()["\"dtype\""].getDataType()
# create a different copy to not change the original copy
curNodeShape = tensor.getShapeRef()[:]
tensorConstantVal = tensor.getConstantVal()
if tensorConstantVal is not None:
# Use uinterpreted call of CreateTensor to create the tensor and fill it with a constant value
dataPassed = None
if curNodeDataType == Graph.DataTypeEnum.DT_INT32:
dataPassed = AST.Int(tensorConstantVal, 32)
elif curNodeDataType == Graph.DataTypeEnum.DT_FLOAT:
dataPassed = AST.Float(tensorConstantVal)
else:
assert False
if (len(curNodeShape) == 0):
# This is a constant element
retAST = dataPassed
else:
retAST = AST.UninterpFuncCall(
curNodeShape,
TFNodesAST.UninterpFuncCallNames.CreateTensor.name,
[dataPassed],
isSecret=False)
else:
# The tensor content is given as byte array. Extract val array from the byte array and create ast.
if curNodeDataType == Graph.DataTypeEnum.DT_INT32:
dataPassed = list(
map(lambda x: AST.Int(x, 32),
tensor.getContentAsValArr()[:]))
elif curNodeDataType == Graph.DataTypeEnum.DT_FLOAT:
dataPassed = list(
map(lambda x: AST.Float(x),
tensor.getContentAsValArr()[:]))
else:
assert False
retAST = AST.Decl(curNodeShape,
None,
None,
dataPassed,
isSecret=False)
return (None, retAST)
def Input(node, value_info, node_name_to_out_var_dict, init_val=None):
if (DEBUG):
print(node.outputs[0])
# There are two types of inputs
dims = list(node.dims if hasattr(node, 'dims') else (
[val.dim_value for val in node.type.tensor_type.shape.dim]))
data_type = node.data_type if hasattr(
node, 'data_type') else node.type.tensor_type.elem_type
# return AST.Input(dims, onnx2seedot(data_type))
from onnx import numpy_helper
range = (-3, 3)
if init_val is not None:
arr = numpy_helper.to_array(init_val)
range = (np.min(arr), np.max(arr))
return AST.Decl(dims, range)