IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Revert "Fix squeeze shape infer to not change axis value (#13975)" (#14430)

Browse Source 
This reverts commit 580c0c6b90.
This commit is contained in:
Maxim Vafin 2022-12-06 14:32:56 +01:00 committed by GitHub
parent a27d69db75
commit 3b86f12c37
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 10 additions and 148 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
tools/mo/openvino/tools/mo/middle/permute_tensor_iterator.py
View File

@ -7,7 +7,6 @@ from openvino.tools.mo.middle.LSTMRNNSequenceToTensorIterator import LSTMToTenso
from openvino.tools.mo.middle.ONNXRNNSequenceNormalize import ONNXRNNSequenceNormalize from openvino.tools.mo.middle.ONNXRNNSequenceNormalize import ONNXRNNSequenceNormalize
from openvino.tools.mo.middle.SwapAxesMiddleReplacer import SwapAxisMiddleReplacer from openvino.tools.mo.middle.SwapAxesMiddleReplacer import SwapAxisMiddleReplacer
from openvino.tools.mo.middle.TensorIteratorMerge import TensorIteratorMerge from openvino.tools.mo.middle.TensorIteratorMerge import TensorIteratorMerge
from openvino.tools.mo.ops.const import Const
from openvino.tools.mo.ops.gather import Gather from openvino.tools.mo.ops.gather import Gather
from openvino.tools.mo.front.common.partial_infer.utils import int64_array from openvino.tools.mo.front.common.partial_infer.utils import int64_array
from openvino.tools.mo.front.tf.graph_utils import create_op_with_const_inputs from openvino.tools.mo.front.tf.graph_utils import create_op_with_const_inputs
@ -198,14 +197,10 @@ class TransposeTensorIteratorLSTM(MiddleReplacementPattern):
isomorphism['input_unsqueezed_i'].shape = isomorphism['input_unsqueezed_i'].shape[[1, 0, 2]] isomorphism['input_unsqueezed_i'].shape = isomorphism['input_unsqueezed_i'].shape[[1, 0, 2]]
isomorphism['input_unsqueezed_i'].infer(isomorphism['input_unsqueezed_i']) isomorphism['input_unsqueezed_i'].infer(isomorphism['input_unsqueezed_i'])
squeeze_dim = Const(squeeze.graph, {'value': [ti.input_port_map[data_input_port]['axis']], isomorphism['squeeze_dim'].value = ti.input_port_map[data_input_port]['axis']
'need_shape_inference': True, isomorphism['squeeze_dim'].infer(isomorphism['squeeze_dim'])
'override_output_shape': True}).create_node() isomorphism['squeeze']['need_shape_inference'] = True
squeeze.in_port(1).get_connection().set_source(squeeze_dim.out_port(0))
squeeze['need_shape_inference'] = True
unsqueeze_dim = Const(unsqueeze.graph, {'value': [ti.output_port_map[data_output_port]['axis']], isomorphism['unsqueeze_dim'].value = ti.output_port_map[data_output_port]['axis']
'need_shape_inference': True, isomorphism['unsqueeze_dim'].infer(isomorphism['unsqueeze_dim'])
'override_output_shape': True}).create_node() isomorphism['unsqueeze'].infer(isomorphism['unsqueeze'])
unsqueeze.in_port(1).get_connection().set_source(unsqueeze_dim.out_port(0))
unsqueeze.infer(unsqueeze)

4
tools/mo/openvino/tools/mo/ops/squeeze.py
View File

@ -62,6 +62,10 @@ class Squeeze(Op):
output_shape = shape_delete(output_shape, real_squeeze_dims) output_shape = shape_delete(output_shape, real_squeeze_dims)
node.out_port(0).data.set_shape(output_shape) node.out_port(0).data.set_shape(output_shape)
# make dimensions positive to correctly translate from NHWC to NCHW layout
if node.in_port(1).get_source().node.op == 'Const':
node.in_port(1).data.set_value(real_squeeze_dims)
if node.in_port(0).data.get_value() is not None: if node.in_port(0).data.get_value() is not None:
node.out_port(0).data.set_value(node.in_port(0).data.get_value().reshape(output_shape)) node.out_port(0).data.set_value(node.in_port(0).data.get_value().reshape(output_shape))

137
tools/mo/unit_tests/mo/middle/permute_tensor_iterator_test.py
View File

@ -1,137 +0,0 @@
# Copyright (C) 2018-2022 Intel Corporation
# SPDX-License-Identifier: Apache-2.0
import unittest
from openvino.tools.mo.front.common.partial_infer.utils import int64_array, shape_array
from openvino.tools.mo.graph.graph import Node
from openvino.tools.mo.middle.permute_tensor_iterator import TransposeTensorIteratorLSTM
from openvino.tools.mo.ops.parameter import Parameter
from openvino.tools.mo.ops.squeeze import Squeeze
from openvino.tools.mo.ops.unsqueeze import Unsqueeze
from openvino.tools.mo.utils.ir_engine.compare_graphs import compare_graphs
from unit_tests.utils.graph import build_graph, connect, connect_data, regular_op_with_shaped_data, \
valued_const_with_data, regular_op_with_empty_data, result, shaped_const_with_data, empty_data
class PermuteTensorIteratorTest(unittest.TestCase):
def test_permute_ti(self):
nodes = {
**regular_op_with_shaped_data('input', [3, 1, 10],
{'type': 'Parameter', 'op': 'Parameter'}),
**valued_const_with_data('direct_permute_order', int64_array([1, 0, 2])),
**regular_op_with_empty_data('direct_permute', {'type': 'Transpose', 'op': 'Transpose'}),
**regular_op_with_empty_data('init_hidden', {'op': 'SomeOp'}),
**regular_op_with_empty_data('init_cell', {'op': 'SomeOp'}),
**regular_op_with_shaped_data('ti', [3, 1, 10],
{'type': 'TensorIterator', 'op': 'TensorIterator', 'body': None,
'output_port_map': [{'axis': 1, 'start': -1, 'end': 0, 'stride': -1,
'external_port_id': 0}],
'input_port_map': [{'axis': 1, 'start': 0, 'end': -1, 'stride': 1,
'external_port_id': 0}]}),
**valued_const_with_data('inverse_permute_order', int64_array([1, 0, 2])),
**regular_op_with_shaped_data('inverse_permute', [3, 1, 10],
{'type': 'Transpose', 'op': 'Transpose', 'shape': [3, 1, 10]}),
**result()
}
graph = build_graph(nodes,
[*connect('input', '0:direct_permute'),
*connect('direct_permute_order', '1:direct_permute'),
('direct_permute', 'direct_permute_d'),
('direct_permute_d', 'ti', {'in': 0, 'external_port_id': 0}),
*connect('init_hidden', '1:ti'),
*connect('init_cell', '2:ti'),
('ti', 'ti_d', {'external_port_id': 0}),
('ti_d', 'inverse_permute', {'in': 0}),
*connect('inverse_permute_order', '1:inverse_permute'),
*connect('inverse_permute', 'output')],
nodes_with_edges_only=True)
nodes_internal = {
**regular_op_with_shaped_data('input_unsqueezed', [3, 1, 10],
{'type': 'Parameter', 'op': 'Parameter', 'shape': shape_array([3, 1, 10]),
'infer': Parameter.infer}),
**regular_op_with_empty_data('input_hidden', {'type': 'Parameter', 'op': 'Parameter'}),
**regular_op_with_empty_data('input_cell', {'type': 'Parameter', 'op': 'Parameter'}),
**valued_const_with_data("squeeze_dim", int64_array([1])),
**valued_const_with_data("unsqueeze_dim", int64_array([1])),
**regular_op_with_shaped_data('squeeze', [3, 10], {'type': 'Squeeze', 'op': 'Squeeze',
'infer': Squeeze.infer}),
**regular_op_with_shaped_data('unsqueeze', [3, 1, 10],
{'type': 'Unsqueeze', 'op': 'Unsqueeze',
'infer': Unsqueeze.infer}),
**shaped_const_with_data("weights", [1, 2, 3]),
**shaped_const_with_data("biases", [1, 2, 3]),
**regular_op_with_shaped_data('lstm', [3, 10], {'type': 'LSTMCell', 'op': 'LSTMCell', 'shape': [3, 10]}),
**empty_data('lstm_d2'),
**result('op_output'),
**result('op_output_1'),
**result('op_output_2')
}
sub_graph = build_graph(nodes_internal,
[*connect('input_unsqueezed', '0:squeeze'),
*connect('squeeze_dim', '1:squeeze'),
*connect('squeeze', '0:lstm'),
*connect('input_hidden', '1:lstm'),
*connect('input_cell', '2:lstm'),
*connect('weights', '3:lstm'),
*connect('biases', '4:lstm'),
*connect('lstm:0', 'unsqueeze'),
('lstm', 'lstm_d2', {'out': 1}),
*connect_data('lstm:0', 'op_output_1'),
('lstm_d2', 'op_output_2'),
*connect('unsqueeze_dim', '1:unsqueeze'),
*connect('unsqueeze', 'op_output')],
nodes_with_edges_only=True)
ti_node = Node(graph, 'ti')
ti_node.body = sub_graph
TransposeTensorIteratorLSTM().find_and_replace_pattern(graph)
graph.clean_up()
graph_ref = build_graph(nodes,
[('input', 'input_d'),
('input_d', 'ti', {'in': 0, 'external_port_id': 0}),
*connect('init_hidden', '1:ti'),
*connect('init_cell', '2:ti'),
('ti', 'ti_d', {'external_port_id': 0}),
('ti_d', 'output', {'in': 0}), ],
{'input_d': {'shape': [3, 1, 10]},
'ti': {'output_port_map': [{'axis': 0, 'start': -1, 'end': 0, 'stride': -1,
'external_port_id': 0}],
'input_port_map': [{'axis': 0, 'start': 0, 'end': -1, 'stride': 1,
'external_port_id': 0}]},
'ti_d': {'shape': [1, 3, 10]}}, nodes_with_edges_only=True)
sub_graph_ref = build_graph(nodes_internal,
[*connect('input_unsqueezed', '0:squeeze'),
*connect('squeeze_dim', '1:squeeze'),
*connect('squeeze', '0:lstm'),
*connect('input_hidden', '1:lstm'),
*connect('input_cell', '2:lstm'),
*connect('weights', '3:lstm'),
*connect('biases', '4:lstm'),
*connect('lstm:0', 'unsqueeze'),
('lstm', 'lstm_d2', {'out': 1}),
*connect_data('lstm:0', 'op_output_1'),
('lstm_d2', 'op_output_2'),
*connect('unsqueeze_dim', '1:unsqueeze'),
*connect('unsqueeze', 'op_output')],
{'input_unsqueezed': {'shape': shape_array([1, 3, 10])},
'input_unsqueezed_d': {'shape': [1, 3, 10]},
'squeeze_dim': {'value': int64_array([0])},
'unsqueeze_dim': {'value': int64_array([1])},
}, nodes_with_edges_only=True)
flag, resp = compare_graphs(
sub_graph, sub_graph_ref, 'op_output', check_op_attrs=True)
self.assertTrue(flag, resp)
ti_node = Node(graph, 'ti')
ti_node.body = None # compare main graph without body
flag, resp = compare_graphs(
graph, graph_ref, 'output', check_op_attrs=True)
self.assertTrue(flag, resp)