[MO] Fix output shape bug in GatherNDDecomposition (#12110)
This commit is contained in:
Przemyslaw Wysocki
committed by
GitHub
GitHub
parent
e8310f7e0b
commit
603b56ec7f
@@ -748,7 +748,7 @@ openvino/tools/mo/middle/FusedBatchNormNonConstant.py
|
||||
openvino/tools/mo/middle/FusedBatchNormTraining.py
|
||||
openvino/tools/mo/middle/FuseReshapesSequence.py
|
||||
openvino/tools/mo/middle/fusings.py
|
||||
openvino/tools/mo/middle/GatherNdNormalizer.py
|
||||
openvino/tools/mo/middle/GatherNDDecomposition.py
|
||||
openvino/tools/mo/middle/GroupNorm.py
|
||||
openvino/tools/mo/middle/GRURNNSequenceToTensorIterator.py
|
||||
openvino/tools/mo/middle/InputCut.py
|
||||
|
||||
@@ -13,7 +13,7 @@ from openvino.tools.mo.middle.replacement import MiddleReplacementPattern
|
||||
from openvino.tools.mo.ops.reshape import Reshape
|
||||
|
||||
|
||||
class GatherNDNormalize(MiddleReplacementPattern):
|
||||
class GatherNDDecomposition(MiddleReplacementPattern):
|
||||
"""
|
||||
Hot fix for new speech-to-text model enabling while GatherND is not implemented in IE.
|
||||
We can replace GatherND to Reshape + Gather in case when GatherND indices have just one
|
||||
@@ -47,7 +47,7 @@ class GatherNDNormalize(MiddleReplacementPattern):
|
||||
n_dims = indices.shape[-1]
|
||||
non_zero = None
|
||||
for i in range(n_dims):
|
||||
if not all(np.take(indices, indices=[i], axis=-1) == 0):
|
||||
if not np.all(np.take(indices, indices=[i], axis=-1) == 0):
|
||||
if non_zero is None:
|
||||
non_zero = i
|
||||
else:
|
||||
@@ -69,7 +69,8 @@ class GatherNDNormalize(MiddleReplacementPattern):
|
||||
# 0. All needed checks that we can replace GatherND by Gather
|
||||
gather_idx = self.indices_check(indices, input_shape)
|
||||
if gather_idx is None:
|
||||
log.warning('Node {} with op=GatherND can\'t be normalized to op=Gather.'.format(gather_name))
|
||||
log.warning(
|
||||
'Node {} with op=GatherND can\'t be normalized to op=Gather.'.format(gather_name))
|
||||
return
|
||||
|
||||
# 1. Add Reshape and connect
|
||||
@@ -78,8 +79,9 @@ class GatherNDNormalize(MiddleReplacementPattern):
|
||||
{'name': gather_name + '/Reshape_for_GatherND/'})
|
||||
gather.in_port(0).get_connection().set_destination(reshape.in_port(0))
|
||||
|
||||
# 2. Change indices from Nd to 1d:
|
||||
new_indices = np.reshape(np.take(indices, indices=[gather_idx], axis=-1), [-1])
|
||||
# 2. Eliminate last dim (n_dims values) from indices shape:
|
||||
new_indices = np.reshape(
|
||||
np.take(indices, indices=[gather_idx], axis=-1), indices.shape[:-1])
|
||||
|
||||
rename_node(gather, gather_name + '/to_delete')
|
||||
|
||||
395
tools/mo/unit_tests/mo/middle/GatherNDDecomposition_test.py
Normal file
395
tools/mo/unit_tests/mo/middle/GatherNDDecomposition_test.py
Normal file
@@ -0,0 +1,395 @@
|
||||
# Copyright (C) 2018-2022 Intel Corporation
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
import unittest
|
||||
import numpy as np
|
||||
|
||||
from openvino.tools.mo.middle.GatherNDDecomposition import GatherNDDecomposition
|
||||
from openvino.tools.mo.utils.ir_engine.compare_graphs import compare_graphs
|
||||
from unit_tests.utils.graph import build_graph
|
||||
|
||||
|
||||
nodes = {
|
||||
'input': {'kind': 'op', 'op': 'Const'},
|
||||
'input_data': {'kind': 'data'},
|
||||
|
||||
'indices_input': {'kind': 'op', 'op': 'Const'},
|
||||
'indices_input_data': {'kind': 'data'},
|
||||
|
||||
'gathernd': {'kind': 'op', 'op': 'GatherND'},
|
||||
'gathernd_data': {'kind': 'data'},
|
||||
|
||||
'result': {'kind': 'op', 'op': 'Result'},
|
||||
}
|
||||
|
||||
edges = [
|
||||
('input', 'input_data'),
|
||||
('input_data', 'gathernd', {'in': 0}),
|
||||
|
||||
('indices_input', 'indices_input_data'),
|
||||
('indices_input_data', 'gathernd', {'in': 1}),
|
||||
|
||||
('gathernd', 'gathernd_data'),
|
||||
('gathernd_data', 'result'),
|
||||
]
|
||||
|
||||
nodes_expected = {
|
||||
'input': {'kind': 'op', 'op': 'Const'},
|
||||
'input_data': {'kind': 'data'},
|
||||
|
||||
'reshape_shape': {'kind': 'op', 'op': 'Const'},
|
||||
'reshape_shape_data': {'kind': 'data'},
|
||||
|
||||
'reshape': {'kind': 'op', 'op': 'Reshape'},
|
||||
'reshape_data': {'kind': 'data'},
|
||||
|
||||
'axis': {'kind': 'op', 'op': 'Const'},
|
||||
'axis_data': {'kind': 'data'},
|
||||
|
||||
'indices': {'kind': 'op', 'op': 'Const'},
|
||||
'indices_data': {'kind': 'data'},
|
||||
|
||||
'gather': {'kind': 'op', 'op': 'Gather'},
|
||||
'gather_data': {'kind': 'data'},
|
||||
|
||||
'result': {'kind': 'op', 'op': 'Result'},
|
||||
}
|
||||
|
||||
edges_expected = [
|
||||
('input', 'input_data'),
|
||||
('input_data', 'reshape', {'in': 0}),
|
||||
|
||||
('reshape_shape', 'reshape_shape_data'),
|
||||
('reshape_shape_data', 'reshape', {'in': 1}),
|
||||
|
||||
('reshape', 'reshape_data'),
|
||||
('reshape_data', 'gather', {'in': 0}),
|
||||
|
||||
('indices', 'indices_data'),
|
||||
('indices_data', 'gather', {'in': 1}),
|
||||
|
||||
('axis', 'axis_data'),
|
||||
('axis_data', 'gather', {'in': 2}),
|
||||
|
||||
('gather', 'gather_data'),
|
||||
('gather_data', 'result'),
|
||||
]
|
||||
|
||||
|
||||
class GatherNDDecompositionTest(unittest.TestCase):
|
||||
|
||||
def test_GatherNDDecomposition_2by2indices_validinputs(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 1]), 'value': np.array([1, 2]).reshape([2, 1])},
|
||||
'indices_input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 0, 0, 0]).reshape([2, 2])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = build_graph(nodes_expected,
|
||||
edges_expected,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 1]), 'value': np.array([1, 2]).reshape([2, 1])},
|
||||
'indices': {'shape': np.array([2]), 'value': np.array([1, 0])},
|
||||
'reshape_shape': {'shape': np.array([1]), 'value': np.array([-1])},
|
||||
'axis': {'shape': np.array([]), 'value': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_2by2indices_invalidinputs(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 2, 3, 4]).reshape([2, 2])},
|
||||
'indices_input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 0, 0, 0]).reshape([2, 2])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = graph
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_2by1indices_validinputs(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 2, 3, 4]).reshape([2, 2])},
|
||||
'indices_input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 0]).reshape([2, 1])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = build_graph(nodes_expected,
|
||||
edges_expected,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 2, 3, 4]).reshape([2, 2])},
|
||||
'indices': {'shape': np.array([2]), 'value': np.array([1, 0])},
|
||||
'reshape_shape': {'shape': np.array([2]), 'value': np.array([-1, 2])},
|
||||
'axis': {'shape': np.array([]), 'value': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_2by0indices_invalidinputs(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([1, 2]), 'value': np.array([1, 2]).reshape([1, 2])},
|
||||
'indices_input_data': {'shape': np.array([2]), 'value': np.array([1, 0])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = graph
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_2by0indices_validinputs(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 1]), 'value': np.array([1, 2])},
|
||||
'indices_input_data': {'shape': np.array([2]), 'value': np.array([1, 0])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = build_graph(nodes_expected,
|
||||
edges_expected,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 1]), 'value': np.array([1, 2])},
|
||||
'indices': {'shape': np.array([]), 'value': np.array([1])},
|
||||
'reshape_shape': {'shape': np.array([1]), 'value': np.array([-1])},
|
||||
'axis': {'shape': np.array([]), 'value': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_1leadingdim(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 2, 3, 4]).reshape([2, 2])},
|
||||
'indices_input_data': {'shape': np.array([2, 1, 1]), 'value': np.array([1, 0]).reshape([2, 1, 1])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = build_graph(nodes_expected,
|
||||
edges_expected,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 2, 3, 4]).reshape([2, 2])},
|
||||
'indices': {'shape': np.array([2, 1]), 'value': np.array([1, 0]).reshape([2, 1])},
|
||||
'reshape_shape': {'shape': np.array([2]), 'value': np.array([-1, 2])},
|
||||
'axis': {'shape': np.array([]), 'value': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_3leadingdims(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 2, 3, 4]).reshape([2, 2])},
|
||||
'indices_input_data': {'shape': np.array([2, 1, 1, 1, 1]), 'value': np.array([1, 0]).reshape([2, 1, 1, 1, 1])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = build_graph(nodes_expected,
|
||||
edges_expected,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 2, 3, 4]).reshape([2, 2])},
|
||||
'indices': {'shape': np.array([2, 1, 1, 1]), 'value': np.array([1, 0]).reshape([2, 1, 1, 1])},
|
||||
'reshape_shape': {'shape': np.array([2]), 'value': np.array([-1, 2])},
|
||||
'axis': {'shape': np.array([]), 'value': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_nonzerobatchdim(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 2]), 'value': np.array([1, 2, 3, 4]).reshape([2, 2])},
|
||||
'indices_input_data': {'shape': np.array([2, 1]), 'value': np.array([1, 0]).reshape([2, 1])},
|
||||
'gathernd': {'batch_dims': 1}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = graph
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_complexexample1_nonzerobatchdim(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 3, 4]), 'value': np.array([i for i in range(24)]).reshape([2, 3, 4])},
|
||||
'indices_input_data': {'shape': np.array([2, 1]), 'value': np.array([1, 0]).reshape([2, 1])},
|
||||
'gathernd': {'batch_dims': 1}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = graph
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_complexexample2(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 3, 4]), 'value': np.array([i for i in range(24)]).reshape([2, 3, 4])},
|
||||
'indices_input_data': {'shape': np.array([2, 1]), 'value': np.array([1, 0]).reshape([2, 1])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = build_graph(nodes_expected,
|
||||
edges_expected,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([2, 3, 4]), 'value': np.array([i for i in range(24)]).reshape([2, 3, 4])},
|
||||
'indices': {'shape': np.array([2]), 'value': np.array([1, 0]).reshape([2])},
|
||||
'reshape_shape': {'shape': np.array([3]), 'value': np.array([-1, 3, 4])},
|
||||
'axis': {'shape': np.array([]), 'value': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_complexexample3(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([1, 1, 5]), 'value': np.array([1, 2, 3, 4, 5]).reshape([1, 1, 5])},
|
||||
'indices_input_data': {'shape': np.array([2, 2, 3]), 'value': np.array([0, 0, 3, 0, 0, 1, 0, 0, 4, 0, 0, 2]).reshape([2, 2, 3])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = build_graph(nodes_expected,
|
||||
edges_expected,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([1, 1, 5]), 'value': np.array([1, 2, 3, 4, 5]).reshape([1, 1, 5])},
|
||||
'indices': {'shape': np.array([2, 2]), 'value': np.array([3, 1, 4, 2]).reshape([2, 2])},
|
||||
'reshape_shape': {'shape': np.array([1]), 'value': np.array([-1])},
|
||||
'axis': {'shape': np.array([]), 'value': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_complexexample4(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([1, 4, 1]), 'value': np.array([1, 2, 3, 4]).reshape([1, 4, 1])},
|
||||
'indices_input_data': {'shape': np.array([2, 2, 2]), 'value': np.array([0, 1, 0, 3, 0, 2, 0, 0]).reshape([2, 2, 2])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = build_graph(nodes_expected,
|
||||
edges_expected,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([1, 4, 1]), 'value': np.array([1, 2, 3, 4]).reshape([1, 4, 1])},
|
||||
'indices': {'shape': np.array([2, 2]), 'value': np.array([1, 3, 2, 0]).reshape([2, 2])},
|
||||
'reshape_shape': {'shape': np.array([2]), 'value': np.array([-1, 1])},
|
||||
'axis': {'shape': np.array([]), 'value': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_dynamic_data_shape(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([1, -1, 1]), 'value': np.array([1, 2, 3, 4]).reshape([1, 4, 1])},
|
||||
'indices_input_data': {'shape': np.array([2, 2, 2]), 'value': np.array([0, 1, 0, 3, 0, 2, 0, 0]).reshape([2, 2, 2])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = graph
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
|
||||
def test_GatherNDDecomposition_dynamic_indices_shape(self):
|
||||
|
||||
graph = build_graph(nodes,
|
||||
edges,
|
||||
update_attributes={
|
||||
'input_data': {'shape': np.array([1, 4, 1]), 'value': np.array([1, 2, 3, 4]).reshape([1, 4, 1])},
|
||||
'indices_input_data': {'shape': np.array([2, -1, 2]), 'value': np.array([0, 1, 0, 3, 0, 2, 0, 0]).reshape([2, 2, 2])},
|
||||
'gathernd': {'batch_dims': 0}
|
||||
},
|
||||
nodes_with_edges_only=True)
|
||||
graph_ref = graph
|
||||
|
||||
GatherNDDecomposition().find_and_replace_pattern(graph)
|
||||
|
||||
(flag, resp) = compare_graphs(
|
||||
graph, graph_ref, 'result', check_op_attrs=True)
|
||||
self.assertTrue(flag, resp)
|
||||
Reference in New Issue
Block a user