def _exchange_dims_in_attrs(self):
self._attr_value_mem[self.AttrName.ORDER][transformed_axis(
src="NCHW", dst="NHWC", ndim=self._input_ndim,
dim=self._dim0)] = transformed_axis(
src="NCHW", dst="NHWC", ndim=self._input_ndim, dim=self._dim1)
self._attr_value_mem[self.AttrName.ORDER][transformed_axis(
src="NCHW", dst="NHWC", ndim=self._input_ndim,
dim=self._dim1)] = transformed_axis(
src="NCHW", dst="NHWC", ndim=self._input_ndim, dim=self._dim0)
def dim(self, value):
self._attr_value_mem[self.AttrName.AXIS][:] = [
transformed_axis(src="NCHW",
dst="NHWC",
ndim=self._input_ndim,
dim=value)
]
def start(self, start):
if self._input_ndim != 4:
begin_mask = 0
for dim, pos in enumerate(start):
if pos == 0:
begin_mask |= 1 << dim
self.set_attr(self.AttrName.BEGIN_MASK, begin_mask)
self.set_attr(self.AttrName.BEGIN, start)
else:
begin = [0] * 4
begin_mask = 0
for dim, pos in enumerate(start):
new_dim = transformed_axis(src="NCHW",
dst="NHWC",
ndim=self._input_ndim,
dim=dim)
begin[new_dim] = pos
for dim, pos in enumerate(begin):
if pos == 0:
begin_mask |= 1 << dim
self.set_attr(self.AttrName.BEGIN_MASK, begin_mask)
self.set_attr(self.AttrName.BEGIN, begin)
def squeeze(xgraph: XGraph, node: Node,
quant_config: NndctQuantInfo) -> NoReturn:
if node.in_tensors[0].ndim == 4 and len(
node.node_attr(node.op.AttrName.DIMS)) == 1:
attrs: Dict[str, Any] = {}
attrs["order"] = [0, 3, 1, 2]
# resume dimension to NCHW
input_ops: Dict[str, List[Op]] = {}
input_list = []
for input in node.in_nodes:
input_op = xgraph.get_op_by_name(input)
input_list.append(input_op)
input_ops["input"] = input_list
xgraph.create_fixed_normal_op(node.name + "_i0",
"transpose",
quant_config,
attrs=attrs,
input_ops=input_ops)
attrs: Dict[str, Any] = {}
dim = node.node_attr(node.op.AttrName.DIMS)[0]
dim = transformed_axis("NHWC", "NCHW", ndim=4, dim=dim)
attrs["axis"] = [dim]
input_ops: Dict[str, List[Op]] = {}
input_ops["input"] = [xgraph.get_op_by_name(node.name + "_i0")]
xgraph.create_fixed_normal_op(node.name,
"squeeze",
quant_config,
attrs=attrs,
input_ops=input_ops)
else:
to_xir("squeeze")(xgraph, node, quant_config)
def dims(self, value):
dims = []
for dim in value:
dims.append(
transformed_axis(
src="NCHW", dst="NHWC", ndim=self._input_ndim, dim=dim))
self._attr_value_mem[self.AttrName.ORDER][:] = dims[:]
def dim(self):
dims = []
for dim in self.get_attr(self.AttrName.DIMS):
dims.append(
transformed_axis(
src="NHWC", dst="NCHW", ndim=self._input_ndim, dim=dim))
return tuple(dims)
def dims(self):
dims = []
for dim in self._attr_value_mem[self.AttrName.ORDER]:
dims.append(
transformed_axis(
src="NHWC", dst="NCHW", ndim=self._input_ndim, dim=dim))
return dims
def step(self):
if self._input_ndim != 4:
return self.get_attr(self.AttrName.STRIDES)
else:
strides = [1] * 4
for dim, step in enumerate(self.get_attr(self.AttrName.STRIDES)):
new_dim = transformed_axis(
src="NHWC", dst="NCHW", ndim=self._input_ndim, dim=dim)
strides[new_dim] = step
return strides
def end(self):
if self._input_ndim != 4:
return self.get_attr(self.AttrName.END)
else:
end = [NNDCT_CONSTANT.INT_MAX] * 4
for dim, pos in enumerate(self.get_attr(self.AttrName.END)):
new_dim = transformed_axis(
src="NHWC", dst="NCHW", ndim=self._input_ndim, dim=dim)
end[new_dim] = pos
return end
def start(self):
if self._input_ndim != 4:
return self.get_attr(self.AttrName.BEGIN)
else:
begin = [0] * 4
for dim, pos in enumerate(self.get_attr(self.AttrName.BEGIN)):
new_dim = transformed_axis(
src="NHWC", dst="NCHW", ndim=self._input_ndim, dim=dim)
begin[new_dim] = pos
return begin
def dim(self, value):
if isinstance(value, (tuple, list)):
value = list(value)
else:
value = [value]
dims = []
for dim in value:
dims.append(
transformed_axis(
src="NCHW", dst="NHWC", ndim=self._input_ndim, dim=dim))
self.set_attr(self.AttrName.DIMS, [dims[:]])
def end(self, end):
if self._input_ndim != 4:
end_mask = 0
for dim, pos in enumerate(end):
if isinstance(pos, int) and pos >= NNDCT_CONSTANT.INT_MAX:
end_mask |= 1 << dim
self.set_attr(self.AttrName.END_MASK, end_mask)
self.set_attr(self.AttrName.END, end)
else:
new_end = [NNDCT_CONSTANT.INT_MAX] * 4
end_mask = 0
for dim, pos in enumerate(end):
new_dim = transformed_axis(
src="NCHW", dst="NHWC", ndim=self._input_ndim, dim=dim)
new_end[new_dim] = pos
for dim, pos in enumerate(new_end):
if isinstance(pos, int) and pos >= NNDCT_CONSTANT.INT_MAX:
end_mask |= 1 << dim
self.set_attr(self.AttrName.END_MASK, end_mask)
self.set_attr(self.AttrName.END, new_end)
def dim(self):
return transformed_axis(
src="NHWC",
dst="NCHW",
ndim=self._input_ndim,
dim=self._attr_value_mem[self.AttrName.AXIS][0])
def dim1(self):
return self._attr_value_mem[self.AttrName.ORDER][transformed_axis(
src="NHWC", dst="NCHW", ndim=self._input_ndim, dim=self._dim0)]
