def parse(self):
logger.debug("Parsing %s...", self.type)
op = self.tflite
opcode = self.model.OperatorCodes(op.OpcodeIndex()).BuiltinCode()
assert (opcode in self.TypeMapping)
assert (op.InputsLength() == 1)
assert (op.OutputsLength() == 1)
ilayout = Layout('NHWC', 'NCHW')
self.parseInput(0, ilayout)
op_opt = op.BuiltinOptions()
option = tflite.Pool2DOptions()
option.Init(op_opt.Bytes, op_opt.Pos)
self.attrs['auto_pad'] = PaddingMapping[option.Padding()]
self.attrs['kernel_shape'] = [
option.FilterHeight(), option.FilterWidth()
]
self.attrs['strides'] = [option.StrideH(), option.StrideW()]
olayout = Layout('NHWC', 'NCHW')
ot = self.parseOutput(0, olayout)
handleFusedActivation(self, option, ot)
self.setParsed()
def parse(self):
logger.debug("Parsing %s...", self.type)
op = self.tflite
opcode = self.model.OperatorCodes(op.OpcodeIndex()).BuiltinCode()
assert (opcode in self.TypeMapping)
assert (op.InputsLength() == 2), "TFLite has only two inputs"
assert (op.OutputsLength() == 1)
# ONNX Resize doesn't have layout semantic, but TFLite requires NHWC
ilayout = Layout('NHWC', 'NCHW')
im = self.parseInput(0, ilayout)
# ROI and Scale are not optional until Resize v13,
# currently (v11) we create them as empty initializer.
# After v13, we can try to not include them in graph
empty_input = self.TFactory.createEmptyTensor()
empty_input.addConsumer(self)
self.inputs.append(empty_input) # ROI
self.inputs.append(empty_input) # Scale
# output size
sz = self.parseInput(1)
# TFLite sizes is (H_new, W_new) while ONNX needs (N, C, H_new,W_new)
assert len(sz.data) == 2
assert len(im.shape) == 4
sz.shape = [len(im.shape)]
sz.data = np.concatenate(
(np.array([im.shape[0], im.shape[-1]]), sz.data)).astype('int64')
sz.dtype = mapping.DTYPE_NAME2ONNX['int64']
# output
olayout = Layout('NHWC', 'NCHW')
self.parseOutput(0, olayout)
# options
if opcode is tflite.BuiltinOperator.RESIZE_BILINEAR:
self.attrs['mode'] = 'linear'
option = tflite.ResizeBilinearOptions()
elif opcode is tflite.BuiltinOperator.RESIZE_NEAREST_NEIGHBOR:
self.attrs['mode'] = 'nearest'
option = tflite.ResizeNearestNeighborOptions()
else:
assert False, "Unreachable path!"
op_opt = op.BuiltinOptions()
option.Init(op_opt.Bytes, op_opt.Pos)
if option.AlignCorners():
self.attrs['coordinate_transformation_mode'] = 'align_corners'
elif option.HalfPixelCenters():
self.attrs['coordinate_transformation_mode'] = 'half_pixel'
else:
raise NotImplementedError("This path has not been tried")
self.setParsed()
def preserveOutputSpatialSemantic(self):
# If the input (or output) tensor of Reshape
# holds a special data layout semantic,
# we need to insert Transpose before (or after) the Reshape operator
# https://github.com/jackwish/tflite2onnx/issues/28
# An example for inserting `Transpose` after `Reshape`
# -------
# | Reshape |
# -------
# |
# | to_transpose (New created tensor)
# | (e.g. NHWC)
# -------
# | Trans | e.g. perm: (0, 3, 1, 2)
# -------
# |
# | transposed (Original `Reshape` output)
# | (e.g. NCHW)
# ------
# | Conv |
# ------
#
assert (self.status.parsed)
transposed = self.outputs[0]
to_transpose_name = 'TFLITE2ONNX_ToTranspose_%s' % transposed.name
to_transpose = self.TFactory.getWithRef(transposed, to_transpose_name,
True)
# Construct a layout as original output of `Reshape`
# e.g. source: NCHW, target: NHWC
# Exchange the source and target layout for layout preservation
layout = Layout(transposed.layout.target, transposed.layout.source)
to_transpose.shape = layout.transform(transposed.shape)
to_transpose.setParsed()
# Construct the additional transpose after `Reshape`
trans = Transpose(self.TFactory, -1)
# e.g. From NHWC to NCHW
trans.attrs['perm'] = transposed.layout.perm
trans.inputs.append(to_transpose)
transposed.replaceProducer(self, trans)
self.replaceOutput(transposed, to_transpose)
trans.outputs.append(transposed)
to_transpose.addProducer(self)
to_transpose.addConsumer(trans)
trans.setParsed()
# Rename the new `Transpose` operator
# to avoid the same name with 'Reshape'
trans.name = 'TFLITE2ONNX_Transpose_%s' % transposed.name
self.post.append(trans)
def parse(self):
logger.debug("Parsing %s...", self.type)
op = self.tflite
opcode = self.model.OperatorCodes(op.OpcodeIndex()).BuiltinCode()
assert (opcode in self.TypeMapping)
assert (op.InputsLength() == 4)
assert (op.OutputsLength() == 1)
# input
ilayout = Layout('NHWC', 'NCHW')
it = self.parseInput(2, ilayout)
# weight
wlayout = Layout('CHWM', 'MCHW')
wt = self.parseInput(1, wlayout)
# bias
self.parseInput(3, is_bias=True)
# shape
osi = self.tflite.Inputs(0)
out_shape = self.TFactory.get(osi, None, False)
out_shape.parse()
# output
olayout = Layout('NHWC', 'NCHW')
ot = self.parseOutput(0, olayout)
assert (collections.Counter(ot.shape) == collections.Counter(
out_shape.data))
# options
op_opt = op.BuiltinOptions()
option = tflite.TransposeConvOptions()
option.Init(op_opt.Bytes, op_opt.Pos)
self.attrs['dilations'] = [1, 1]
self.attrs['group'] = 1
self.attrs['kernel_shape'] = wt.shape[1:3]
self.attrs['strides'] = [option.StrideW(), option.StrideH()]
# XXX Not enabled as ONNXRuntime has limitation to infer pads for non-1 dilation
# self.attrs['auto_pad'] = PaddingMapping[option.Padding()]
self.attrs['pads'] = computePaddingSize(option.Padding(),
it.shape[1:3],
self.attrs['kernel_shape'],
self.attrs['strides'],
self.attrs['dilations'])
# handleFusedActivation(self, option, ot)
self.setParsed()
def parse(self):
logger.debug("Parsing %s...", self.type)
op = self.tflite
opcode = self.model.OperatorCodes(op.OpcodeIndex()).BuiltinCode()
assert (opcode in self.TypeMapping)
assert (op.InputsLength() == 3), "TFLite Conv always has bias"
assert (op.OutputsLength() == 1)
# input
ilayout = Layout('NHWC', 'NCHW')
it = self.parseInput(0, ilayout)
# weight
wlayout = Layout('CHWM', 'MCHW') if self.isDepthwise else Layout(
'OHWI', 'OIHW')
wt = self.parseInput(1, wlayout)
# bias
self.parseInput(2, is_bias=True)
# output
olayout = Layout('NHWC', 'NCHW')
ot = self.parseOutput(0, olayout)
# options
op_opt = op.BuiltinOptions()
option = tflite.DepthwiseConv2DOptions(
) if self.isDepthwise else tflite.Conv2DOptions()
option.Init(op_opt.Bytes, op_opt.Pos)
self.attrs['dilations'] = [
option.DilationHFactor(),
option.DilationWFactor()
]
self.attrs['group'] = wt.shape[3] if self.isDepthwise else 1
self.attrs['kernel_shape'] = wt.shape[1:3]
self.attrs['strides'] = [option.StrideH(), option.StrideW()]
# XXX Not enabled as ONNXRuntime has limitation to infer pads for non-1 dilation
# self.attrs['auto_pad'] = PaddingMapping[option.Padding()]
if self.isDepthwise:
assert (option.DepthMultiplier() == 1)
self.attrs['pads'] = computePaddingSize(option.Padding(),
it.shape[1:3],
self.attrs['kernel_shape'],
self.attrs['strides'],
self.attrs['dilations'])
handleFusedActivation(self, option, ot)
self.setParsed()
def preserveOutputSpatialSemantic(self):
# https://github.com/jackwish/tflite2onnx/issues/28
# An example for inserting `Transpose` after `Reshape`
# -------
# | Reshape |
# -------
# |
# | to_transpose (New created tensor)
# | (e.g. NHWC)
# -------
# | Trans | e.g. perm: (0, 3, 1, 2)
# -------
# |
# | transposed (Original `Reshape` output)
# | (e.g. NCHW)
# ------
# | Conv |
# ------
assert (self.status.parsed)
transposed = self.outputs[0]
to_transpose_name = 'TFLITE2ONNX_ToTranspose_%s' % transposed.name
to_transpose = self.TFactory.getWithRef(transposed, to_transpose_name,
True)
# Construct a layout from the original output of `Reshape`
layout = Layout(transposed.layout.target, transposed.layout.source)
to_transpose.shape = layout.transform(transposed.shape)
to_transpose.setParsed()
# Construct the additional transpose after `Reshape`
trans = Transpose(self.TFactory, -1)
trans.attrs['perm'] = transposed.layout.perm
trans.inputs.append(to_transpose)
transposed.replaceProducer(self, trans)
self.replaceOutput(transposed, to_transpose)
trans.outputs.append(transposed)
to_transpose.addProducer(self)
to_transpose.addConsumer(trans)
trans.setParsed()
# Rename the new `Transpose` operator avoid the name conflict with 'Reshape'
trans.name = 'TFLITE2ONNX_Transpose_%s' % transposed.name
self.post.append(trans)
def preserveInputSpatialSemantic(self):
# https://github.com/jackwish/tflite2onnx/issues/28
# An example for inserting `Transpose` before `Reshape`
# ------
# | Conv |
# ------
# |
# | to_transpose (Original input of `Reshape`)
# | (e.g. NCHW)
# -------
# | Trans | e.g. perm: (0, 2, 3, 1)
# -------
# |
# | transposed (New created tensor)
# | (e.g. NHWC)
# --------
# | Reshape |
# --------
assert (self.status.parsed)
to_transpose = self.inputs[0]
transposed_name = 'TFLITE2ONNX_Transposed_%s' % to_transpose.name
transposed = self.TFactory.getWithRef(to_transpose, transposed_name,
True)
# Construct the layout from the original input of `Reshape`
layout = Layout(to_transpose.layout.target, to_transpose.layout.source)
transposed.shape = layout.transform(to_transpose.shape)
transposed.setParsed()
# Construct the additional transpose before `Reshape`
trans = Transpose(self.TFactory, -1)
trans.attrs['perm'] = layout.perm
trans.inputs.append(to_transpose)
to_transpose.replaceConsumer(self, trans)
self.replaceInput(to_transpose, transposed)
trans.outputs.append(transposed)
transposed.addProducer(trans)
transposed.addConsumer(self)
trans.setParsed()
self.pre.append(trans)
def convert(self, explicit_layouts):
logger.debug("Converting...")
logger.debug("Handling data layout...")
for op in self.ops:
for t in op.inputs + op.outputs:
if t.name in explicit_layouts:
assert (t.layout is None)
layouts = explicit_layouts[t.name]
assert (len(layouts) == 2)
t.layout = Layout(layouts[0], layouts[1])
self._propagateLayout()
self._collectOpAndTensor()
logger.debug("Translating quantization semantic...")
for t in self.value_info | self.initializer:
deqt = handleQuantizationTensor(self.TFactory, t)
for i, o in enumerate(self.outputs):
if o == t:
self.outputs[i] = deqt
self._collectOpAndTensor()
logger.debug("Graph:\n%s", str(self))
self.validate()
for op in self.op_all:
op.convert()
logger.debug("Making ONNX...")
onodes = [n.onnx for n in self.op_all]
oinputs = [t.onnx for t in self.inputs]
ooutputs = [t.onnx for t in self.outputs]
initializer = [t.onnx for t in self.initializer]
value_info = [t.onnx for t in self.value_info]
self.onnx = helper.make_graph(onodes,
'pre-alpha',
oinputs,
ooutputs,
initializer=initializer,
value_info=value_info)
self.setConverted()
def parse(self):
logger.debug("Parsing %s...", self.type)
op = self.tflite
opcode = self.model.OperatorCodes(op.OpcodeIndex()).BuiltinCode()
assert (opcode in self.TypeMapping)
assert (op.InputsLength() == 3)
assert (op.OutputsLength() == 1)
# oshape
osi = op.Inputs(0)
oshape = self.TFactory.getData(self.model, self.graph, osi, 'int32')
# X
ilayout = Layout('NHWC', 'NCHW')
self.parseInput(2, ilayout)
# weight
wlayout = Layout('OHWI', 'IOHW')
wt = self.parseInput(1, wlayout)
# FIXME: we don't have a model containing bias.
# output
olayout = Layout('NHWC', 'NCHW')
ot = self.parseOutput(0, olayout)
assert ((ot.shape == oshape).all())
# options
op_opt = op.BuiltinOptions()
option = tflite.TransposeConvOptions()
option.Init(op_opt.Bytes, op_opt.Pos)
self.attrs['kernel_shape'] = wt.shape[1:3]
self.attrs['strides'] = [option.StrideH(), option.StrideW()]
oslayout = Layout('NHWC', 'NCHW')
self.attrs['output_shape'] = oslayout.transform(oshape)
self.setParsed()