def Version_6(cls, ctx, node, **kwargs):
scalar_val = (node.get_attr_value("int_operand")
if node.get_attr_value("has_int_operand") else
node.get_attr_value("float_operand"))
np_dtype = util.Onnx2NumpyDtype(ctx.get_dtype(node.input[0]))
scalar_node = ctx.MakeConst(id_util.UniqueStr("scalar"),
np.array([scalar_val]).astype(np_dtype))
node.input.append(scalar_node.output[0])
def Version_6(cls, ctx, node, **kwargs):
scalar_val = (node.attrs["int_operand"]
if node.attrs["has_int_operand"] else
node.attrs["float_operand"])
np_dtype = util.Onnx2NumpyDtype(
ctx.get_dtype(node.input_tensor_names[0]))
scalar_node = ctx.MakeConst(id_util.UniqueStr("scalar"),
np.array([scalar_val]).astype(np_dtype))
node.input_tensor_names.append(scalar_node.output_tensor_names[0])
def get_saved_tensor(self, node):
tensor_name = node.output_tensor_names[0]
# TODO(daquexian): node.output_tensor_names[0] is "node_name/output_name", so this pathjoin doesn't work
# on windows (where path separator is "\")
path = pathjoin(self._model_save_dir, node.output_tensor_names[0])
tensor_value = np.fromfile(
path, dtype=util.Onnx2NumpyDtype(self.get_dtype(tensor_name))
).reshape(self.get_shape(tensor_name))
return tensor_value
def Version_11(cls, ctx, node, **kwargs):
node.set_attr("mode", "constant")
padding_before = node.get_attr_value("padding_before")
padding_after = node.get_attr_value("padding_after")
paddings = np.array(padding_before + padding_after).astype(np.int64)
padding_node = ctx.MakeConst(id_util.UniqueStr("const"), paddings)
node.input.append(padding_node.output[0])
dtype = ctx.get_dtype(node.input[0])
const_val = (node.get_attr_value("integral_constant_value")
if util.is_integral_onnx_dtype(dtype) else
node.get_attr_value("floating_constant_value"))
const_val = np.array(const_val).astype(util.Onnx2NumpyDtype(dtype))
const_val_node = ctx.MakeConst(id_util.UniqueStr("const"), const_val)
node.input.append(const_val_node.output[0])
def Version_6(cls, ctx, node, **kwargs):
node.op_type = "BatchNormalization"
# flow inputs: x, gamma, beta, moving_mean, moving_variance
# flow outputs: y, mean, inv_variance
# a: data_format, epsilon, is_training
# onnx inputs: X, scale, B, mean, variance, attributes: epsilon, momentum=0.9, spatial : 1
# output: y, mean, var, savedmean, savedvar,
# detach unused outputs. While we could let the unused outputs dangle,
# some runtimes like pytorch/caffe2 do complain about it.
if node.attrs["training"]:
raise NotImplementedError(
"We only support inference mode ONNX BatchNormalization now"
)
consumers = [
ctx.FindOutputConsumers(output_name)
for output_name in node.output_tensor_names[1:]
]
if not any(consumers):
new_output = [node.output_tensor_names[0]]
node.output_tensor_names = new_output
_ConvConvertInputs(ctx, node, with_kernel=False)
scale_shape = ctx.get_shape(node.input_tensor_names[1])
mean_shape = ctx.get_shape(node.input_tensor_names[3])
var_shape = ctx.get_shape(node.input_tensor_names[4])
val_type = util.Onnx2NumpyDtype(ctx.get_dtype(node.input_tensor_names[1]))
if mean_shape != scale_shape:
new_mean_value = np.array(
np.resize(
node.input_nodes[3].get_tensor_value(as_list=False), scale_shape
),
dtype=val_type,
)
new_mean_node_name = id_util.UniqueStr(node.name)
ctx.MakeConst(new_mean_node_name, new_mean_value)
node.input_tensor_names[3] = new_mean_node_name
if var_shape != scale_shape:
new_var_value = np.array(
np.resize(
node.input_nodes[4].get_tensor_value(as_list=False), scale_shape
),
dtype=val_type,
)
new_val_node_name = id_util.UniqueStr(node.name)
ctx.MakeConst(new_val_node_name, new_var_value)
node.input_tensor_names[4] = new_val_node_name