openvino/model-optimizer/mo/graph/perm_inputs.py

245 lines
13 KiB
Python

# Copyright (C) 2018-2021 Intel Corporation
# SPDX-License-Identifier: Apache-2.0
import networkx as nx
from extensions.ops.gather import Gather
from extensions.ops.transpose import Transpose
from mo.front.common.partial_infer.utils import int64_array
from mo.graph.graph import Node
from mo.ops.const import Const
def get_node_with_permutation(node: Node, port_info: str):
node_type, port = port_info.split(':')
port = int(port)
return node.in_node(port) if node_type == 'input' else node.out_node(port)
def axis(op_node: Node, port_info: str, input_port: int):
"""
Performs layout change related transformation of the data on the in_port_idx port of op_node.
Translates shape indexes from one layout to another according to inverse permutation
Transformation inserts Gather operation with
permutation as 0-port input data and
actual data to translate as 1-port input indexes of Gather
For example:
NHWC Reduce operation has 0-port input with data of shape [1, 2, 3, 4] and
1-port input with axis indices [0, 1].
After translating such operation to NCHW layout:
0-port input shape = [1, 4, 2, 3]
1-port input axis indices = [0, 2]
"""
graph = op_node.graph
permutation_data_node = get_node_with_permutation(op_node, port_info)
assert permutation_data_node.has_and_set('permutation'), 'Data node "{}" does not have permutation for node {}, ' \
'port_info "{}".'.format(permutation_data_node.id,
op_node.id, port_info)
permutation = permutation_data_node.permutation
if len(permutation.perm) == 0:
return
data_node = op_node.in_node(input_port)
gather_name = op_node.soft_get('name', op_node.id) + '/AxisGather'
const = Const(graph, {'value': permutation.inv, 'name': gather_name + '/const',
'need_shape_inference': True}).create_node_with_data()
axis_const = Const(graph, {'value': int64_array(0), 'name': gather_name + '/axis'}).create_node_with_data()
gather = Gather(graph, {'name': gather_name, 'need_shape_inference': True}).create_node_with_data(
[const, data_node, axis_const])
attrs = graph.get_edge_data(data_node.id, op_node.id, key=0).copy()
graph.add_edge(gather.id, op_node.id, **attrs)
graph.remove_edge(data_node.id, op_node.id)
op_node['need_shape_inference'] = True
def order(op_node: Node, port_info: str, input_port: int):
"""
Performs layout change related transformation of the data on the in_port_idx port of op_node.
Translates ordered shape indexes from one layout to another according to permutation
Transformation inserts two Gather operations
1 Gather reorders data to new layout according to direct permutation:
actual data to translate as 1-port input indexes of Gather and
permutation as 0-port input data
2 Gather translates shape indexes from one layout to another according to inverse permutation
permutation as 0-port input data and
actual data to translate as 1-port input indexes of Gather
For example:
NHWC Transpose operation has 0-port input with data of shape [1, 2, 3, 4] and
1-port input with new order indices [0, 1, 3, 2].
After translating such operation to NCHW layout:
0-port input shape = [1, 4, 2, 3]
1 phase (after first Gather insertion):
1-port input order indices = [0, 2, 1, 3]
2 phase (after second Gather insertion):
1-port input order indices = [0, 3, 2, 1]
"""
graph = op_node.graph
permutation_data_node = get_node_with_permutation(op_node, port_info)
assert permutation_data_node.has_and_set('permutation'), 'Data node "{}" does not have permutation for node {}, ' \
'port_info "{}".'.format(permutation_data_node.id,
op_node.id, port_info)
permutation = permutation_data_node.permutation
if len(permutation.perm) == 0:
return
data_node = op_node.in_node(input_port)
gather_name = op_node.soft_get('name', op_node.id) + '/OrderGather_1'
const = Const(graph, {'value': permutation.perm, 'name': gather_name + '/const',
'need_shape_inference': True}).create_node_with_data()
axis_const = Const(graph, {'value': int64_array(0), 'name': gather_name + '/axis'}).create_node_with_data()
gather = Gather(graph, {'name': gather_name,
'need_shape_inference': True}).create_node_with_data([data_node, const, axis_const])
gather_1_name = op_node.soft_get('name', op_node.id) + '/OrderGather_2'
const_1 = Const(graph, {'value': permutation.inv, 'name': gather_1_name + '/const',
'need_shape_inference': True}).create_node_with_data()
axis_const_1 = Const(graph, {'value': int64_array(0), 'name': gather_1_name + '/axis'}).create_node_with_data()
gather_1 = Gather(graph, {'name': gather_1_name,
'need_shape_inference': True}).create_node_with_data([const_1, gather, axis_const_1])
attrs = graph.get_edge_data(data_node.id, op_node.id, key=0).copy()
graph.add_edge(gather_1.id, op_node.id, **attrs)
graph.remove_edge(data_node.id, op_node.id)
op_node['need_shape_inference'] = True
def strided_slice(op_node: Node, port_info: str, input_port: int):
"""
StridedSLice must be permuted even if input or output tensors have rank lesser than 4
e.g. input_shape = (1, 10, 10), out = input[:, 0:10, :, new_axis], input_rank < 4
input_shape = (1, 10, 10, 3), out = input[:, 0:5, 0:4, 0], output_rank < 4
in both examples slice_rank is >= 4
slice_rank is defined by length of begin, end, strides (they all are of the same length)
"""
permutation_data_node = get_node_with_permutation(op_node, port_info)
assert permutation_data_node.has_and_set('permutation'), 'Data node "{}" does not have permutation for node {}, ' \
'port_info "{}".'.format(permutation_data_node.id,
op_node.id, port_info)
permute_indices_for_gather = permutation_data_node.permutation.perm
if len(permute_indices_for_gather) == 0:
return
from mo.ops.op import PermuteAttrs
slice_rank = op_node.in_port(input_port).data.get_shape()[0] # length of begin, end or strides
permute_indices_for_gather = PermuteAttrs.get_nhwc_to_nchw_permutation(slice_rank).perm
reorder_inputs_for_shape_or_slice(op_node, input_port, permute_indices_for_gather)
def shape(op_node: Node, port_info: str, input_port: int):
permutation_data_node = get_node_with_permutation(op_node, port_info)
assert permutation_data_node.has_and_set('permutation'), 'Data node "{}" does not have permutation for node {}, ' \
'port_info "{}".'.format(permutation_data_node.id,
op_node.id, port_info)
permute_indices_for_gather = permutation_data_node.permutation.perm
if len(permute_indices_for_gather) == 0:
return
reorder_inputs_for_shape_or_slice(op_node, input_port, permute_indices_for_gather)
def reorder_inputs_for_shape_or_slice(op_node: Node, input_port: int, permute_indices_for_gather: list):
"""
axis and slice permutations are almost the same the only difference is that for slice in general
case permutation depends from slice_rank not from input_rank or output_rank
"""
graph = op_node.graph
data_node = op_node.in_node(input_port)
gather_name = op_node.soft_get('name', op_node.id) + '/ShapeGather'
const = Const(graph, {'value': permute_indices_for_gather, 'name': gather_name + '/const',
'need_shape_inference': True}).create_node_with_data()
axis_const = Const(graph, {'value': int64_array(0), 'name': gather_name + '/axis'}).create_node_with_data()
gather = Gather(graph, {'name': gather_name,
'need_shape_inference': True}).create_node_with_data([data_node, const, axis_const])
attrs = graph.get_edge_data(data_node.id, op_node.id, key=0).copy()
graph.add_edge(gather.id, op_node.id, **attrs)
graph.remove_edge(data_node.id, op_node.id)
# need to run manually to override output shape value to resolve shape collision for nodes with
# 'correct_data_layout' output port attrs
op_node['need_shape_inference'] = True
def transpose(op_node: Node, port_info: str, input_port: int):
graph = op_node.graph
permutation_data_node = get_node_with_permutation(op_node, port_info)
assert permutation_data_node.has_and_set('permutation'), \
'Data node "{}" does not have permutation for node {}, port_info "{}".'.format(
permutation_data_node.id, op_node.id, port_info)
permutation = permutation_data_node.permutation
if len(permutation.perm) == 0:
return
transpose_name = op_node.soft_get('name', op_node.id) + '/Transpose'
from mo.front.tf.graph_utils import create_op_with_const_inputs # avoiding recursive imports
transpose = create_op_with_const_inputs(
graph, Transpose, {1: permutation.perm}, {'name': transpose_name, 'override_output_shape': True})
op_node.in_port(input_port).get_connection().insert_node(transpose)
transpose.infer(transpose)
def transpose_nchw_to_nhwc(op_node: Node, port_info: str, input_port: int):
graph = op_node.graph
permutation_data_node = get_node_with_permutation(op_node, port_info)
rank = len(permutation_data_node.shape)
assert rank >= 4, 'Rank must be 4D or higher for HCHW to HHWC permutation on node {}.'.format(op_node.id)
perm = list(range(rank))
perm.insert(1, perm.pop())
perm = int64_array(perm)
transpose_name = op_node.soft_get('name', op_node.id) + '/Transpose'
from mo.front.tf.graph_utils import create_op_with_const_inputs # avoiding recursive imports
transpose = create_op_with_const_inputs(
graph, Transpose, {1: perm}, {'name': transpose_name, 'override_output_shape': True})
op_node.in_port(input_port).get_connection().insert_node(transpose)
transpose.infer(transpose)
class PermuteInputs:
input_permutes = {
'axis': lambda node, port_info, input_port: axis(node, port_info, input_port),
'slice': lambda node, port_info, input_port: strided_slice(node, port_info, input_port),
'order': lambda node, port_info, input_port: order(node, port_info, input_port),
'shape': lambda node, port_info, input_port: shape(node, port_info, input_port),
'transpose': lambda node, port_info, input_port: transpose(node, port_info, input_port),
'transpose_nchw_to_nhwc': lambda node, port_info, input_port: transpose_nchw_to_nhwc(node, port_info,
input_port),
}
shape_check_rules = {
'rank': lambda port: bool(len(port.data.get_shape()) >= 4),
'dim_size': lambda port: bool(port.data.get_shape()[0] >= 4), # if input 'dim_size' >= 4 need to permute
}
def set_input_permutation(self, node1: Node, node2: Node, port_info: str, permutation_rule: str,
shape_check_rule: str = 'rank'):
"""
Sets input permutation attribute on the edge between node1 and node2.
Input permutation consists of function that perform input permutation and
input port info 'input' or 'output' + <port_number> that points on the input with PermuteAttr.Permutation which
current input depends on.
shape_check_rule defines the check rule if the op node inputs need to be permuted.
By default 'rank' rule is applied, 'dim_size' is used only for StridedSlice so far.
"""
assert permutation_rule in self.input_permutes, 'No `{}` permutation rule in {}'.format(permutation_rule,
__class__.__name__)
assert shape_check_rule in self.shape_check_rules, 'No `{}` permutation shape check rule ' \
'in {}'.format(shape_check_rule, __class__.__name__)
nx.set_edge_attributes(G=node1.graph,
values={(node1.id, node2.id, 0): (self.input_permutes[permutation_rule], port_info,
self.shape_check_rules[shape_check_rule])},
name='input_permutation')