def convert_blob_tensor_format(tensor: base_tensor.Tensor, src_framework: str,
dst_framework: str) -> base_tensor.Tensor:
if not isinstance(tensor, base_tensor.Tensor):
raise TypeError("'tensor' must be Tensor, but given {}".format(
type(tensor)))
if not tensor.is_complete_tensor():
return tensor
if tensor.ndim != 4:
return tensor
if src_framework == dst_framework:
return tensor
src_layout = DataFormatMap.blob_format(src_framework)
dst_layout = DataFormatMap.blob_format(dst_framework)
if src_layout == dst_layout:
return tensor
if src_layout == "NCHW" and dst_layout == "NHWC":
tensor.transpose((0, 2, 3, 1))
elif src_layout == "NHWC" and dst_layout == "NCHW":
tensor.transpose((0, 3, 1, 2))
else:
raise ValueError(
"Can not transpose data format of '{}' from '{}' to '{}'".format(
op_type, src_layout, dst_layout))
return tensor
def convert_blob_tensor_format(tensor: base_tensor.Tensor, src_framework: str,
dst_framework: str) -> base_tensor.Tensor:
if not isinstance(tensor, base_tensor.Tensor):
raise TypeError("'tensor' must be Tensor, but given {}".format(
type(tensor)))
if not tensor.is_complete_tensor():
return tensor
if src_framework == dst_framework:
return tensor
if tensor.ndim not in DataFormatMap._blob_format_map[src_framework].keys():
return tensor
src_layout = DataFormatMap.blob_format(src_framework, tensor.ndim)
dst_layout = DataFormatMap.blob_format(dst_framework, tensor.ndim)
if src_layout == dst_layout:
return tensor
tensor.transpose(layout_transformer(src_layout, dst_layout))
return tensor
def clone_from(self, src_op, local_map):
for attr_name, attr_value in src_op._attrs.items():
new_attr_value = NndctIrAttr.create_attr_from(attr_value)
new_attr_value.value = self.clone_attr_from(
attr_name, attr_value.value, local_map)
self._attr_value_mem[attr_name] = new_attr_value._value_mem
self._attrs[attr_name] = new_attr_value
for i, config in enumerate(src_op.configs):
value = src_op.get_config(config)
new_value = self.clone_attr_from(config, value, local_map)
self._configs[i] = config
setattr(self, config, new_value)
for param_name, param_tensor in src_op.params.items():
if param_tensor.name in local_map:
self._params[param_name] = local_map[param_tensor.name]
else:
tensor = Tensor(name=param_tensor.name)
tensor.clone_from(param_tensor)
local_map[param_tensor.name] = tensor
self._params[param_name] = tensor
self._is_custom_op = src_op._is_custom_op
def set_param_from_data(self,
name: Enum,
data: np.ndarray,
framework_name: Optional[str] = None):
new_tensor = Tensor()
new_tensor.from_ndarray(data)
if framework_name:
new_tensor.name = framework_name
else:
new_tensor.name = self._params[name].name
self._params[name] = new_tensor
def set_param_from_des(self,
param_name: Enum,
description: Dict[str, Any],
fw_name: Optional[str] = None):
if not isinstance(description, dict):
raise TypeError("'description' must be a dictionary")
new_tensor = Tensor()
new_tensor.from_des(**description)
if fw_name:
new_tensor.name = fw_name
else:
new_tensor.name = self._params[param_name].name
self._params[param_name] = new_tensor
def convert_parameter_tensor_format(tensor: base_tensor.Tensor,
src_framework: str,
dst_framework: str) -> base_tensor.Tensor:
if not isinstance(tensor, base_tensor.Tensor):
raise TypeError("'tensor' must be Tensor, but given {}".format(
type(tensor)))
if not tensor.is_complete_tensor():
return tensor
if tensor.ndim != 4 and tensor.ndim != 2:
return tensor
if src_framework == dst_framework:
return tensor
src_format = DataFormatMap.param_format(src_framework, tensor.ndim)
dst_format = DataFormatMap.param_format(dst_framework, tensor.ndim)
if src_format == dst_format:
return tensor
elif src_format == 'OIHW' and dst_format == 'HWIO':
tensor.transpose((2, 3, 1, 0))
elif src_format == 'HWIO' and dst_format == 'OIHW':
tensor.transpose((3, 2, 0, 1))
elif src_format == 'OIHW' and dst_format == 'OWHI' or \
(src_format == 'OWHI' and dst_format == 'OIHW'):
tensor.transpose((0, 3, 2, 1))
elif src_format == 'HWIO' and dst_format == 'OWHI':
tensor.transpose((3, 1, 0, 2))
elif src_format == 'OIHW' and dst_format == 'OHWI':
tensor.transpose((0, 2, 3, 1))
elif src_format == 'OHWI' and dst_format == 'OIHW':
tensor.transpose((0, 3, 1, 2))
elif (src_format == 'OI' and dst_format == 'IO') or \
(src_format == 'IO' and dst_format == 'OI'):
tensor.transpose((1, 0))
else:
raise ValueError(
"Can not transpose data format from '{}' to '{}'".format(
src_format, dst_format))
return tensor