def inference_split_shape(g):
processed = False
for node in g.node:
if node.op_type != 'Split':
continue
input_val_info = helper.find_value_by_name(g, node.input[0])
if not input_val_info:
input_val_info = helper.find_input_by_name(g, node.input[0])
if not input_val_info:
continue
_, input_shape = helper.find_size_shape_from_value(input_val_info)
if not input_shape:
continue
output_val_names = list(node.output)
output_vals = [
helper.find_value_by_name(g, val_name)
for val_name in output_val_names
]
output_shapes = [
helper.find_size_shape_from_value(output_val)[1]
for output_val in output_vals
]
if not any([len(s) == 0 for s in output_shapes]):
continue
for att in node.attribute:
if att.name == 'axis':
axis = att.i
else:
split = list(att.ints)
new_output_vals = []
for i in range(len(output_val_names)):
new_shape = list(input_shape)
new_shape[axis] = split[i]
new_output_val = onnx.helper.make_tensor_value_info(
output_val_names[i], input_val_info.type.tensor_type.elem_type,
new_shape)
new_output_vals.append(new_output_val)
for val in output_vals:
if val is not None:
g.value_info.remove(val)
g.value_info.extend(new_output_vals)
processed = True
return processed
def inference_cov_shape(g):
processed = False
for node in g.node:
if node.op_type != 'Conv':
continue
input_value_info = helper.find_value_by_name(g, node.input[0])
if not input_value_info:
input_value_info = helper.find_input_by_name(g, node.input[0])
if not input_value_info:
continue
kernel_value_info = helper.find_value_by_name(g, node.input[1])
output_value_info = helper.find_value_by_name(g, node.output[0])
if not output_value_info:
output_value_info = helper.find_output_by_name(g, node.output[0])
if output_value_info and \
helper.get_shape_from_value_info(output_value_info):
continue
_, kernel_shape = helper.find_size_shape_from_value(kernel_value_info)
_, input_shape = helper.find_size_shape_from_value(input_value_info)
if not input_shape or not kernel_shape:
continue
strides = helper.get_attribute_by_name(node, 'strides').ints
pads = helper.get_attribute_by_name(node, 'pads').ints
dilation = helper.get_attribute_by_name(node, 'dilations').ints
# Pytorch model has the case where strides only have one number
if len(strides) == 1:
return strides.append(strides[0])
if len(dilation) == 1:
return dilation.append(dilation[0])
H = math.floor((input_shape[2]+pads[0]+pads[2]-\
dilation[0]*(kernel_shape[2]-1)-1)/strides[0]+1)
W = math.floor((input_shape[3]+pads[1]+pads[3]-\
dilation[1]*(kernel_shape[3]-1)-1)/strides[1]+1)
output_shape = [input_shape[0], kernel_shape[0], H, W]
new_output_value_info = onnx.helper.make_tensor_value_info(
node.output[0], input_value_info.type.tensor_type.elem_type,
output_shape)
processed = True
if output_value_info:
g.value_info.remove(output_value_info)
g.value_info.extend([new_output_value_info])
return processed
def replace_split_with_slices(g):
"""Replace split node with slice nodes.
:param g: input graph.
:return:
"""
node_to_remove = []
for node in g.node:
# Find a Split
if node.op_type != 'Split':
continue
input_value = helper.find_value_by_name(g, node.input[0])
if not input_value:
input_value = helper.find_input_by_name(g, node.input[0])
_, shape = helper.find_size_shape_from_value(input_value)
if len(shape) == 0:
continue
output_val_names = list(node.output)
axis = 0
split = []
for item in node.attribute:
if item.name == 'axis':
axis = item.i
if item.name == 'split':
split = item.ints
length = input_value.type.tensor_type.shape.dim[axis].dim_value
outputs = node.output
if split is not []:
n_out = len(node.attribute[1].ints)
pos = 0
for i in range(n_out):
pos += node.attribute[1].ints[i]
new_node_name = output_val_names[i]
new_node = onnx.helper.make_node(
op_type='Slice',
inputs=[node.input[0]],
outputs=[new_node_name],
name=new_node_name,
axes=[axis],
ends=[pos],
starts=[pos - node.attribute[1].ints[i]])
g.node.extend([new_node])
node_to_remove.append(node)
else:
n_out = len(outputs)
width = length // n_out
for i in range(n_out):
new_node = onnx.helper.make_node(op_type='Slice',
inputs=[node.input[0]],
outputs=[outputs[i]],
name=outputs[i],
axes=[axis],
ends=[(1 + i) * width],
starts=[i * width])
g.node.extend([new_node])
node_to_remove.append(node)
for old_node in node_to_remove:
g.node.remove(old_node)
topological_sort(g)
def replace_mul_to_bn(g):
"""Replace single Mul node with Batchnorm node.
:param g: input graph.
:return:
"""
node_to_del = []
for node in g.node:
if node.op_type != 'Mul':
continue
mul_op_node = node
# only support one input node
if len(mul_op_node.input) != 2: # OP node and value node
continue
input_op_node_name = mul_op_node.input[0]
mul_value_node = helper.find_node_by_output_name(
g, mul_op_node.input[1])
if not mul_value_node or mul_value_node.op_type != 'Constant':
continue
_, previous_node_output_shape = helper.find_size_shape_from_value(
helper.find_value_by_name(g, input_op_node_name))
scale_shape, scale_data = helper.constant_to_list(mul_value_node)
# only allow 4 dim data input due to the hardware limitation
if len(previous_node_output_shape) != 4:
continue
# channel dimension
c_dim = previous_node_output_shape[1]
# only allow channelwise mul or const mul
if scale_shape != [1, c_dim, 1, 1] and scale_shape != 1:
continue
ones = [1.0] * c_dim
zeros = [0.0] * c_dim
muls = scale_data * c_dim
bn_name = mul_op_node.output[0]
mean_value_node = helper.list_to_constant(bn_name + '_mean',
np.array(zeros).shape, zeros)
variance_value_node = helper.list_to_constant(bn_name + '_var',
np.array(ones).shape,
ones)
bias_value_node = helper.list_to_constant(bn_name + '_add',
np.array(zeros).shape, zeros)
new_mul_value_node = helper.list_to_constant(bn_name + '_mul',
np.array(muls).shape,
muls)
bn_node = onnx.helper.make_node('BatchNormalization', [
input_op_node_name, new_mul_value_node.output[0],
bias_value_node.output[0], mean_value_node.output[0],
variance_value_node.output[0]
], [mul_op_node.output[0]],
name=bn_name,
epsilon=0.00000001)
mid_val_info = helper.find_value_by_name(g, mul_op_node.output[0])
scale_val_info = helper.find_value_by_name(g, mul_value_node.output[0])
g.value_info.remove(mid_val_info)
g.value_info.remove(scale_val_info)
g.node.extend([bn_node])
g.node.extend([mean_value_node])
g.node.extend([variance_value_node])
g.node.extend([bias_value_node])
g.node.extend([new_mul_value_node])
node_to_del.extend([mul_op_node])
node_to_del.extend([mul_value_node])
while node_to_del:
g.node.remove(node_to_del.pop())
topological_sort(g)
def inference_cov_shape(g):
processed = False
for node in g.node:
# Check for Conv output shape need to be inferrenced.
if node.op_type != 'Conv':
continue
# Input shape is not ready yet. Skip.
input_value_info = helper.find_value_by_name(g, node.input[0])
if not input_value_info:
input_value_info = helper.find_input_by_name(g, node.input[0])
if not input_value_info:
continue
_, input_shape = helper.find_size_shape_from_value(input_value_info)
if not input_shape:
continue
# Output shape is already there. Skip.
output_value_info = helper.find_value_by_name(g, node.output[0])
if not output_value_info:
output_value_info = helper.find_output_by_name(g, node.output[0])
if output_value_info and \
helper.get_shape_from_value_info(output_value_info):
continue
# Now start the inference.
# If auto_pad is set, use the auto_pad.
auto_pad = helper.get_var_attribute_by_name(node, 'auto_pad', 'string')
pads = None
if auto_pad is not None and auto_pad != 'NOTSET':
if auto_pad == 'SAME_LOWER' or auto_pad == 'SAME_UPPER':
new_output_value_info = onnx.helper.make_tensor_value_info(
node.output[0],
input_value_info.type.tensor_type.elem_type,
input_shape
)
if output_value_info:
g.value_info.remove(output_value_info)
g.value_info.extend([new_output_value_info])
processed = True
continue
elif auto_pad == 'VALID':
pads = [0, 0, 0, 0]
else:
print("Unrecognized auto_pad value: " + str(auto_pad))
exit(1)
kernel_value_info = helper.find_value_by_name(g, node.input[1])
_, kernel_shape = helper.find_size_shape_from_value(kernel_value_info)
if not input_shape or not kernel_shape:
continue
strides = helper.get_attribute_by_name(node, 'strides').ints
if not pads:
pads = helper.get_attribute_by_name(node, 'pads').ints
dilation = helper.get_attribute_by_name(node, 'dilations').ints
# Pytorch model has the case where strides only have one number
if len(strides) == 1:
return strides.append(strides[0])
if len(dilation) == 1:
return dilation.append(dilation[0])
H = math.floor((input_shape[2]+pads[0]+pads[2]-\
dilation[0]*(kernel_shape[2]-1)-1)/strides[0]+1)
W = math.floor((input_shape[3]+pads[1]+pads[3]-\
dilation[1]*(kernel_shape[3]-1)-1)/strides[1]+1)
output_shape = [input_shape[0], kernel_shape[0], H, W]
new_output_value_info = onnx.helper.make_tensor_value_info(
node.output[0],
input_value_info.type.tensor_type.elem_type,
output_shape
)
processed = True
if output_value_info:
g.value_info.remove(output_value_info)
g.value_info.extend([new_output_value_info])
return processed
def fuse_mul_and_add_into_gemm(g):
node_to_del = []
for node in g.node:
if node.op_type != 'Add':
continue
add_node = node
mul_node = helper.find_node_by_output_name(g, add_node.input[0])
if not mul_node or mul_node.op_type != 'Mul':
continue
mul_const = helper.find_node_by_output_name(g, mul_node.input[1])
if not mul_const or mul_const.op_type != 'Constant':
continue
add_const = helper.find_node_by_output_name(g, add_node.input[1])
if not add_const or add_const.op_type != 'Constant':
continue
input_val = helper.find_value_by_name(g, mul_node.input[0])
if not input_val:
input_val = helper.find_input_by_name(g, mul_node.input[0])
if not input_val:
continue
_, input_shape = helper.find_size_shape_from_value(input_val)
if not input_shape:
continue
dim = int(np.prod(input_shape))
if input_shape != [1, dim]:
continue
mul_const_shape, mul_const_data = helper.constant_to_list(mul_const)
add_const_shape, __ = helper.constant_to_list(add_const)
if len(mul_const_shape) != 1 or mul_const_shape[0] != dim:
continue
if len(add_const_shape) != 1 or add_const_shape[0] != dim:
continue
b_data = np.zeros([dim, dim])
for i in range(dim):
b_data[i][i] = mul_const_data[i]
b_data = b_data.flatten().tolist()
b_tensor = onnx.helper.make_tensor(
name=mul_const.name + '_tensor',
data_type=mul_const.attribute[0].t.data_type,
dims=[dim, dim],
vals=b_data)
b_const_node = onnx.helper.make_node('Constant', [],
[mul_const.output[0]],
value=b_tensor,
name=mul_const.output[0])
add_const.attribute[0].t.dims.insert(0, 1)
gemm_node = onnx.helper.make_node(
'Gemm',
[mul_node.input[0], b_const_node.output[0], add_const.output[0]],
[add_node.output[0]],
name=add_node.output[0])
g.node.extend([gemm_node, b_const_node])
node_to_del.extend([mul_const, mul_node, add_node])
val_info_mid = helper.find_value_by_name(g, mul_node.output[0])
val_info_mul_const = helper.find_value_by_name(g, mul_const.output[0])
val_info_add_const = helper.find_value_by_name(g, add_const.output[0])
if val_info_mid:
g.value_info.remove(val_info_mid)
if val_info_mul_const:
g.value_info.remove(val_info_mul_const)
if val_info_add_const:
g.value_info.remove(val_info_add_const)
while node_to_del:
g.node.remove(node_to_del.pop())
topological_sort(g)
def fuse_mul_and_add_into_bn(g):
node_to_del = []
for node in g.node:
if node.op_type != 'Add':
continue
add_node = node
input_nodes_add = [
helper.find_node_by_output_name(g, input_name)
for input_name in add_node.input
]
if any([n == None for n in input_nodes_add]):
continue
mul_node, const_add = None, None
for input_node_add in input_nodes_add:
if input_node_add.op_type == 'Mul':
mul_node = input_node_add
elif input_node_add.op_type == 'Constant':
const_add = input_node_add
else:
pass
if not mul_node or not const_add:
continue
data_input_name, const_mul = None, None
for input_name in mul_node.input:
input_node = helper.find_node_by_output_name(g, input_name)
if not input_node:
data_input_name = input_name
elif input_node.op_type == 'Constant':
if not const_mul:
const_mul = input_node
else:
data_input_name = input_name
else:
data_input_name = input_name
if not const_mul:
continue
scale_shape, scale_data = helper.constant_to_list(const_mul)
bais_shape, __ = helper.constant_to_list(const_add)
c_dim = len(scale_data)
if scale_shape != bais_shape:
continue
_, previous_node_output_shape = helper.find_size_shape_from_value(
helper.find_value_by_name(g, data_input_name))
# only allow 4 dim data input due to the hardware limitation
if len(previous_node_output_shape) != 4:
continue
# check if mul's dim and input channel dimension are matched
if previous_node_output_shape[1] != c_dim:
continue
if scale_shape == [1, c_dim, 1, 1]:
# remove all '1'
for _ in range(3):
const_add.attribute[0].t.dims.remove(1)
const_mul.attribute[0].t.dims.remove(1)
elif scale_shape == [1, c_dim]:
# remove all '1'
const_add.attribute[0].t.dims.remove(1)
const_mul.attribute[0].t.dims.remove(1)
else:
continue
bn_name = add_node.output[0]
const_mean = helper.list_to_constant(bn_name + '_mean', [c_dim],
[0.0 for _ in range(c_dim)])
const_var = helper.list_to_constant(bn_name + '_var', [c_dim],
[1.0 for _ in range(c_dim)])
bn_node = onnx.helper.make_node(
'BatchNormalization',
[data_input_name, const_mul.output[0], const_add.output[0],\
const_mean.output[0], const_var.output[0]],
[add_node.output[0]],
name=bn_name,
epsilon=0.00000001
)
mid_val_info = helper.find_value_by_name(g, mul_node.output[0])
scale_val_info = helper.find_value_by_name(g, const_mul.output[0])
bais_val_info = helper.find_value_by_name(g, const_add.output[0])
g.value_info.remove(mid_val_info)
g.value_info.remove(scale_val_info)
g.value_info.remove(bais_val_info)
new_scale_val_info = onnx.helper.make_tensor_value_info(
const_mul.output[0], const_mul.attribute[0].t.data_type, [c_dim])
new_bais_val_info = onnx.helper.make_tensor_value_info(
const_add.output[0], const_add.attribute[0].t.data_type, [c_dim])
mean_val_info = onnx.helper.make_tensor_value_info(
const_mean.output[0], const_mean.attribute[0].t.data_type, [c_dim])
var_val_info = onnx.helper.make_tensor_value_info(
const_var.output[0], const_var.attribute[0].t.data_type, [c_dim])
g.value_info.extend([new_scale_val_info])
g.value_info.extend([new_bais_val_info])
g.value_info.extend([mean_val_info])
g.value_info.extend([var_val_info])
g.node.extend([bn_node])
g.node.extend([const_mean])
g.node.extend([const_var])
node_to_del.extend([mul_node, add_node])
while node_to_del:
g.node.remove(node_to_del.pop())
topological_sort(g)
def replace_add_to_bn(g):
"""Replace single Add node with Batchnorm node.
:param g: input graph.
:return:
"""
node_to_del = []
for node in g.node:
if node.op_type != 'Add':
continue
add_op_node = node
# only support one input node
if len(add_op_node.input) != 2: # OP node and value node
continue
input_op_node_name = add_op_node.input[0]
add_value_node = helper.find_node_by_output_name(
g, add_op_node.input[1])
if not add_value_node or add_value_node.op_type != 'Constant':
continue
prev_shape_value_info = helper.find_value_by_name(
g, input_op_node_name)
prev_shape_value_info = helper.find_input_by_name(
g, input_op_node_name
) if prev_shape_value_info is None else prev_shape_value_info
if prev_shape_value_info is None:
continue
_, previous_node_output_shape = helper.find_size_shape_from_value(
prev_shape_value_info)
bias_shape, bias_data = helper.constant_to_list(add_value_node)
# channel dimension
c_dim = previous_node_output_shape[1] if len(
previous_node_output_shape) > 1 else 1
# only allow channelwise mul or const mul
if bias_shape != [1, c_dim, 1, 1] and bias_shape != 1:
continue
ones = [1.0] * c_dim
zeros = [0.0] * c_dim
# If bias is a scaler, expand it.
if len(bias_data) == 1:
bias = bias_data * c_dim
else:
bias = bias_data
bn_name = add_op_node.output[0]
mean_value_node = helper.list_to_constant(bn_name + '_mean',
np.array(zeros).shape, zeros)
variance_value_node = helper.list_to_constant(bn_name + '_var',
np.array(ones).shape,
ones)
scale_value_node = helper.list_to_constant(bn_name + '_mul',
np.array(ones).shape, ones)
new_add_value_node = helper.list_to_constant(bn_name + '_add',
np.array(bias).shape,
bias)
bn_node = onnx.helper.make_node('BatchNormalization', [
input_op_node_name, scale_value_node.output[0],
new_add_value_node.output[0], mean_value_node.output[0],
variance_value_node.output[0]
], [add_op_node.output[0]],
name=bn_name,
epsilon=0.00000001)
add_val_info = helper.find_value_by_name(g, add_value_node.output[0])
g.value_info.remove(add_val_info)
g.node.extend([bn_node])
g.node.extend([mean_value_node])
g.node.extend([variance_value_node])
g.node.extend([scale_value_node])
g.node.extend([new_add_value_node])
node_to_del.extend([add_op_node])
node_to_del.extend([add_value_node])
while node_to_del:
g.node.remove(node_to_del.pop())
topological_sort(g)
