def to_keras(self):
"""Convert a Onnx model to KerasNet model.
"""
# parse network inputs, aka parameters
for init_tensor in self.graph.initializer:
if not init_tensor.name.strip():
raise ValueError("Tensor's name is required.")
self.initializer[init_tensor.name] = OnnxInput(
name=init_tensor.name, zvalue=OnnxHelper.to_numpy(init_tensor))
# converting GraphProto message
# i: ValueInfoProto
for i in self.graph.input:
if i.name in self.initializer:
# we should have added that via graph._initializer
self._all_tensors[i.name] = self.initializer[i.name]
else:
self._inputs[i.name] = OnnxInput(
name=i.name, zvalue=OnnxHelper.get_shape_from_node(i))
self._all_tensors[i.name] = self._inputs[i.name]
# constructing nodes, nodes are stored as directed acyclic graph
# converting NodeProto message
for node in self.graph.node:
inputs = []
for i in node.input:
if i == "":
continue
if i not in self._all_tensors:
raise Exception("Cannot find {}".format(i))
inputs.append(self._all_tensors[i])
mapper = OperatorMapper.of(node, self.initializer, inputs)
# update inputs and all_tensors
for input in mapper.model_inputs:
# Only update the original inputs here.
if input.name in self._inputs:
self._inputs[input.name] = input.zvalue
self._all_tensors[input.name] = input.zvalue
tensor = mapper.to_tensor()
output_ids = list(node.output)
assert len(output_ids) == 1 or node.op_type == "Dropout",\
"Only support single output for now"
self._all_tensors[output_ids[0]] = OnnxInput(name=tensor.name,
zvalue=tensor)
output_tensors = []
for i in self.graph.output:
if i.name not in self._all_tensors:
raise Exception("The output haven't been calculate")
output_tensors.append(self._all_tensors[i.name].zvalue)
model = zmodels.Model(input=list(self._inputs.values()),
output=output_tensors)
return model
def create_operator(self):
assert len(self.inputs) == 1, "Conv accept single input only"
rank = len(self.inputs[0].get_input_shape())
W_weights = self.params[0]
if (rank == 4): # NCHW
nb_filter = W_weights.shape[0]
nb_row = int(self.onnx_attr['kernel_shape'][0])
nb_col = int(self.onnx_attr['kernel_shape'][1])
subSample = [int(i) for i in self.onnx_attr['strides']]
dim_ordering = "th"
assert self.onnx_attr['dilations'] == (1, 1), "we only support dilations == (1, 1)"
assert self.onnx_attr['group'] == 1, "we only support group == 1"
bias = True if (len(self.params) > 1) else False
# TODO: activation?? init?? W_regularizer??
border_mode, pads = OnnxHelper.get_padds(self.onnx_attr)
conv = zlayers.Convolution2D(nb_filter=nb_filter,
nb_row=nb_row,
nb_col=nb_col,
subsample=subSample,
dim_ordering=dim_ordering,
bias=bias,
border_mode=border_mode,
pads=pads)
return conv
else:
raise Exception("not supported.")
def _to_tensor(self):
input = self.model_inputs[0]
W_weights = self.model_trainable_values[0]
rank = len(input.zvalue.get_input_shape())
if (rank == 4): # NCHW
nb_filter = W_weights.shape[0]
nb_row = int(self.onnx_attr['kernel_shape'][0])
nb_col = int(self.onnx_attr['kernel_shape'][1])
subSample = [int(i) for i in
self.onnx_attr['strides']] if "strides" in self.onnx_attr else (1, 1)
dim_ordering = "th"
assert 'dilations' not in self.onnx_attr or self.onnx_attr['dilations'] == (
1, 1), "we only support dilations == (1, 1)"
assert 'group' not in self.onnx_attr or self.onnx_attr[
'group'] == 1, "we only support group == 1"
bias = True if (len(self._input_list) > 2) else False
border_mode, pads = OnnxHelper.get_padds(self.onnx_attr)
conv = zlayers.Convolution2D(nb_filter=nb_filter,
nb_row=nb_row,
nb_col=nb_col,
subsample=subSample,
dim_ordering=dim_ordering,
bias=bias,
border_mode=border_mode,
pads=pads)
return conv(input.zvalue)
else:
raise Exception("not supported.")
def __init__(self, node, initializer, inputs):
self.node = node
self.op_name = node.op_type
node_name = node.name.strip()
# it would be None if user doesn't set it manually
self.node_name = node_name if node_name else None
self.onnx_attr = OnnxHelper.parse_attr(
node.attribute) # dict(name: value)
self._initializer = initializer
self._input_list = inputs
self.model_inputs = self._extract_model_inputs()
self.model_trainable_values = self._extract_trainable_values()
self.output = node.output[0]
def _to_tensor(self):
assert len(self.model_inputs) == 1, "MaxPool accept single input only"
rank = len(self.model_inputs[0].zvalue.shape)
if "storage_order" in self.onnx_attr.keys():
assert self.onnx_attr['storage_order'] == 0
if (rank == 4): # NCHW
pool_size = [int(i) for i in self.onnx_attr['kernel_shape']]
if "strides" in self.onnx_attr.keys():
strides = [int(i) for i in self.onnx_attr['strides']]
else:
strides = [1 for i in self.onnx_attr['kernel_shape']]
border_mode, pads = OnnxHelper.get_padds(self.onnx_attr)
maxpool = zlayers.MaxPooling2D(pool_size=pool_size,
strides=strides,
border_mode=border_mode,
pads=pads)
return maxpool(self.model_inputs[0].zvalue)
elif (rank == 3):
pool_length = int(self.onnx_attr['kernel_shape'][0])
if "strides" in self.onnx_attr.keys():
stride = int(self.onnx_attr['strides'][0])
else:
stride = 1
border_mode, pads = OnnxHelper.get_padds(self.onnx_attr)
if border_mode is None and pads is None:
border_mode = 'valid'
if pads is None:
pads = 0
permute = zlayers.Permute(dims=(2, 1))(self.model_inputs[0].zvalue)
maxpool = zlayers.MaxPooling1D(pool_length=pool_length,
stride=stride,
border_mode=border_mode,
pad=pads)(permute)
return zlayers.Permute(dims=(2, 1))(maxpool)
else:
raise Exception("not supported.")
def __init__(self, node, _params, _all_tensors):
self.node = node
self.op_name = node.op_type
node_name = node.name.strip()
# it would be None if user doesn't set it manually
self.node_name = node_name if node_name else None
self.onnx_attr = OnnxHelper.parse_attr(
node.attribute) # dict(name: value)
self.params = [_params[i] for i in node.input if i in _params]
self.inputs = [
_all_tensors[i] for i in node.input if i in _all_tensors
]
self.all_tensors = _all_tensors
assert len(node.output) == 1, "we only support single output for now"
self.output = node.output[0]
def _to_tensor(self):
assert len(self.model_inputs) == 1, "AveragePool accept single input only"
rank = len(self.model_inputs[0].zvalue.shape)
if (rank == 4): # NCHW
poolSize = [int(i) for i in self.onnx_attr['kernel_shape']]
strides = [int(i) for i in self.onnx_attr['strides']]
count_include_pad = bool(self.onnx_attr['count_include_pad'])\
if "count_include_pad" in self.onnx_attr else False
dim_ordering = "th"
border_mode, pads = OnnxHelper.get_padds(self.onnx_attr)
averagepool2d = zlayers.AveragePooling2D(pool_size=poolSize,
strides=strides,
dim_ordering=dim_ordering,
pads=pads,
count_include_pad=count_include_pad)
return averagepool2d(self.model_inputs[0].zvalue)
else:
raise Exception("not supported.")
def _to_tensor(self):
assert len(self.model_inputs) == 1, "Conv accept single input only"
rank = len(self.model_inputs[0].zvalue.get_input_shape())
if (rank == 4): # NCHW
pool_size = [int(i) for i in self.onnx_attr['kernel_shape']]
if "strides" in self.onnx_attr.keys():
strides = [int(i) for i in self.onnx_attr['strides']]
else:
strides = [1 for i in self.onnx_attr['kernel_shape']]
border_mode, pads = OnnxHelper.get_padds(self.onnx_attr)
maxpool = zlayers.MaxPooling2D(pool_size=pool_size,
strides=strides,
border_mode=border_mode,
pads=pads)
return maxpool(self.model_inputs[0].zvalue)
else:
raise Exception("not supported.")
def _extract_model_inputs(self):
"""
:return: list of OnnxInput
"""
input = OnnxInput(name=self.op_name, zvalue=OnnxHelper.to_numpy(self.onnx_attr['value']))
return [self._to_zoo_input(input, is_constant=True)]