def onnx1_4to1_6(model: onnx.ModelProto) -> onnx.ModelProto:
"""Update ir_version from 4 to 6 and update opset from 9 to 11.
Args:
model (onnx.ModelProto): input onnx model.
Returns:
onnx.ModelProto: updated onnx model.
"""
graph = model.graph
if model.opset_import[0].version == 11:
print("(Stop) the input model is already opset 11, no need to upgrade")
exit(1)
# deal with empty node name issue
other.add_name_to_node(graph)
# simplify the node param type from initializer to constant
replacing.replace_initializer_with_Constant(graph)
# Modify the nodes.
replace_min_max_attribute_to_const_node_in_clip_node(graph)
replace_all_attribute_to_const_node_in_slice_node(graph)
replace_all_attribute_to_const_node_in_pad_node(graph)
upsampling_to_resize(graph)
other.topological_sort(graph)
# Change model properties.
model.ir_version = 6
model.opset_import[0].version = 11
model = onnx.utils.polish_model(model)
return model
def fuse_bias_in_consecutive_1x1_conv(g):
for second in g.node:
# Find two conv
if second.op_type != 'Conv':
continue
first = helper.find_node_by_output_name(g, second.input[0])
if first is None or first.op_type != 'Conv':
continue
# Check if the first one has only one folloing node
if len(
helper.find_following_nodes_by_input_value_name(
g, first.output[0])) != 1:
continue
# If first node has no bias, continue
if len(first.input) == 2:
continue
# Check their kernel size
first_kernel_shape = helper.get_list_attribute_by_name(
first, 'kernel_shape', 'int')
second_kernel_shape = helper.get_list_attribute_by_name(
second, 'kernel_shape', 'int')
prod = first_kernel_shape[0] * first_kernel_shape[
1] * second_kernel_shape[0] * second_kernel_shape[1]
if prod != 1:
continue
print('Found: ', first.name, ' ', second.name)
# Get bias of the nodes
first_bias_node = helper.find_node_by_output_name(g, first.input[2])
second_weight_node = helper.find_node_by_output_name(
g, second.input[1])
second_bias_node = helper.find_node_by_output_name(g, second.input[2])
first_bias = helper.constant_to_numpy(first_bias_node)
second_weight = helper.constant_to_numpy(second_weight_node)
second_bias = helper.constant_to_numpy(second_bias_node)
# Calculate the weight for second node
first_bias = np.reshape(first_bias, (1, first_bias.size))
second_weight = np.reshape(
second_weight, (second_weight.shape[0], second_weight.shape[1]))
second_weight = np.transpose(second_weight)
new_second_bias = second_bias + np.matmul(first_bias, second_weight)
new_second_bias = np.reshape(new_second_bias, (new_second_bias.size, ))
# Generate new weight
new_first_bias = np.reshape(first_bias, (first_bias.size, ))
for i in range(new_first_bias.shape[0]):
new_first_bias[i] = 0.0
new_first_bias_node = helper.numpy_to_constant(
first_bias_node.output[0], new_first_bias)
new_second_bias_node = helper.numpy_to_constant(
second_bias_node.output[0], new_second_bias)
# Delete old weight and add new weights
g.node.remove(first_bias_node)
g.node.remove(second_bias_node)
g.node.extend([new_first_bias_node, new_second_bias_node])
topological_sort(g)
def do_convert(m):
graph = m.graph
# Modify the nodes.
remove_BN_spatial(graph)
upsample_attribute_to_const(graph)
relu6_to_clip(graph)
PRelu_weight_reshape(graph)
other.topological_sort(graph)
# Change model properties.
m.ir_version = 4
m.opset_import[0].version = 9
return m
help='add nop bn using specific input')
parser.add_argument(
'--rename-output',
dest='rename_output',
type=str,
nargs='+',
help='Rename the specific output(e.g. --rename-output old_name new_name)')
args = parser.parse_args()
# Load model and polish
m = onnx.load(args.in_file)
m = onnx.utils.polish_model(m)
g = m.graph
replacing.replace_initializer_with_Constant(g)
other.topological_sort(g)
# Remove nodes according to the given arguments.
if args.delete_node is not None:
helper.delete_nodes(g, args.delete_node)
if args.delete_input is not None:
helper.delete_input(g, args.delete_input)
if args.delete_output is not None:
helper.delete_output(g, args.delete_output)
# Add do-nothing Conv node
if args.add_conv is not None:
other.add_nop_conv_after(g, args.add_conv)
other.topological_sort(g)
new_input = onnx.helper.make_tensor_value_info(
node.input[1], input_value.type.tensor_type.elem_type,
(1, slope.dims[1], 1, 1))
g.input.remove(input_value)
g.input.append(new_input)
value_info = helper.find_value_by_name(g, node.input[1])
if value_info is not None:
g.value_info.remove(value_info)
if __name__ == "__main__":
if len(sys.argv) != 3:
print("Usage:{} file_in file_out".format(sys.argv[0]))
exit(1)
model = onnx.load(sys.argv[1])
graph = model.graph
# Modify the nodes.
remove_BN_spatial(graph)
upsample_attribute_to_const(graph)
relu6_to_clip(graph)
PRelu_weight_reshape(graph)
other.topological_sort(graph)
# Change model properties.
model.ir_version = 4
model.opset_import[0].version = 9
onnx.save(model, sys.argv[2])