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

[TF FE] Support DynamicStitch operation (#13408)

Browse Source 
* Delete unneccessary changes

* codestyle

* skip dynamic stitch layer tests for legacy frontend

* apply review comments

* fix unit tests, apply review comments
This commit is contained in:
Ivan Tikhonov 2022-10-12 08:04:19 +04:00 committed by GitHub
parent 6bd099917a
commit 8a9890dbf1
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 172 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

77
src/frontends/tensorflow/src/op/parallel_dynamic_stitch.cpp Normal file
View File

@ -0,0 +1,77 @@
// Copyright (C) 2018-2022 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include "op_table.hpp"
#include "openvino/opsets/opset9.hpp"
using namespace std;
using namespace ov::opset9;
namespace ov {
namespace frontend {
namespace tensorflow {
namespace op {
OutputVector translate_parallel_dynamic_stitch_op(const NodeContext& node) {
// format for inputs: [indices1, indices2, ..., indicesN, data1, data2, ..., dataN]
// so we expect at least 2 input and the total number of inputs must be divisible by 2
default_op_checks(node, 2, {"ParallelDynamicStitch", "DynamicStitch"});
auto in_size = node.get_input_size();
TENSORFLOW_OP_VALIDATION(node,
in_size % 2 == 0,
"The total number of inputs to DynamicStitch or ParallelDynamicStitch operation "
"must be divisible by 2.");
int N = static_cast<int>(in_size / 2);
OutputVector indices_to_concat;
OutputVector data_to_concat;
auto data_element_type = node.get_input(N).get_element_type();
auto const_minus_one = make_shared<Constant>(ov::element::i32, Shape{1}, -1);
auto const_zero = make_shared<Constant>(ov::element::i32, Shape{1}, 0);
auto const_one = make_shared<Constant>(ov::element::i32, Shape{1}, 1);
for (int i = 0; i < N; ++i) {
auto indices = node.get_input(i);
auto data = node.get_input(N + i);
const auto& indices_pshape = indices.get_partial_shape();
auto rank = indices_pshape.rank();
TENSORFLOW_OP_VALIDATION(node,
indices_pshape.rank().is_static(),
"Only static rank for `indices` input is supported.");
auto rank_val = rank.get_length();
auto norm_indices = make_shared<Reshape>(indices, const_minus_one, false);
if (rank_val < 1) {
data = make_shared<Unsqueeze>(data, const_zero);
} else if (rank_val > 1) {
auto data_shape = make_shared<ShapeOf>(data, ov::element::i32);
auto start = make_shared<Constant>(ov::element::i32, Shape{1}, rank_val);
auto stop = make_shared<Constant>(ov::element::i32, Shape{1}, numeric_limits<int>::max());
auto shape_of_single_element = make_shared<Slice>(data_shape, start, stop, const_one);
auto new_shape = make_shared<Concat>(OutputVector{const_minus_one, shape_of_single_element}, 0);
data = make_shared<Reshape>(data, new_shape, false);
}
data_to_concat.push_back(data);
indices_to_concat.push_back(norm_indices);
}
auto update = make_shared<Concat>(data_to_concat, 0);
auto indices = make_shared<Concat>(indices_to_concat, 0);
auto data_shape = make_shared<ShapeOf>(update, ov::element::i32);
auto zero = make_shared<Constant>(data_element_type, Shape{}, 0);
auto zeros = make_shared<Broadcast>(zero, data_shape);
auto max_idx = make_shared<ReduceMax>(indices, Constant::create(element::i32, {1}, {0}), true);
auto stop = make_shared<Add>(max_idx->output(0), const_one);
auto start = make_shared<Constant>(ov::element::i32, Shape{1}, 0);
auto axis = make_shared<Constant>(ov::element::i32, Shape{1}, 0);
auto sliced_zeros = make_shared<Slice>(zeros, start, stop, const_one, axis);
auto result = make_shared<ScatterUpdate>(sliced_zeros, indices, update, const_zero);
set_node_name(node.get_name(), result);
return result->outputs();
}
} // namespace op
} // namespace tensorflow
} // namespace frontend
} // namespace ov

3
src/frontends/tensorflow/src/op_table.cpp
View File

@ -77,6 +77,7 @@ OP_CONVERTER(translate_max_pool_op);
OP_CONVERTER(translate_non_max_suppression_op);
OP_CONVERTER(translate_normalize_l2_op);
OP_CONVERTER(translate_pad_op);
OP_CONVERTER(translate_parallel_dynamic_stitch_op);
OP_CONVERTER(translate_placeholder_op);
OP_CONVERTER(translate_placeholder_with_default_op);
OP_CONVERTER(translate_no_op);
@ -257,6 +258,8 @@ const std::map<std::string, CreatorFunction> get_supported_ops() {
{"Pack", translate_pack_op},
{"Pad", translate_pad_op},
{"PadV2", translate_pad_op},
{"DynamicStitch", translate_parallel_dynamic_stitch_op},
{"ParallelDynamicStitch", translate_parallel_dynamic_stitch_op},
{"Placeholder", translate_placeholder_op},
{"PlaceholderWithDefault", translate_placeholder_with_default_op},
{"PreventGradient", translate_identity_op},

2
tests/layer_tests/common/utils/tf_utils.py
View File

@ -86,7 +86,7 @@ def summarize_graph(model_path, output_nodes_for_freeze=None, reshape_net=None):
node_dict['type'] = tf.DType(node.attr['dtype'].type).name
node_dict['shape'] = str(node.attr['shape'].shape.dim).replace('\n', '').replace(' ', '').replace(
'size:', '').replace('[', '').replace(']', '')
node_dict['shape'] = tuple(map(lambda x: int(x), node_dict['shape'].split(',')))
node_dict['shape'] = tuple(map(lambda x: int(x) if x else 0, node_dict['shape'].split(',')))
placeholders[node.name] = node_dict
if node.op == "Variable" or node.op == "VariableV2":
variables.append(node.name)

91
tests/layer_tests/tensorflow_tests/test_tf_ParallelDynamicStitch.py Normal file
View File

@ -0,0 +1,91 @@
# Copyright (C) 2018-2022 Intel Corporation
# SPDX-License-Identifier: Apache-2.0
import numpy as np
import pytest
import tensorflow as tf
from common.tf_layer_test_class import CommonTFLayerTest
class TestParallelDynamicStitch(CommonTFLayerTest):
def _prepare_input(self, inputs_info):
inputs_data = {}
num_elements = 0
assert len(inputs_info) % 2 == 0, "Number of inputs should be divisible by 2."
data_input_cnt = len(inputs_info)//2
for i in range(1, data_input_cnt + 1):
indices_in_name = "indices{}".format(i)
assert indices_in_name in inputs_info, "Test error: inputs_info must contain `{}`".format(indices_in_name)
indices_shape = inputs_info[indices_in_name]
num_elements = num_elements + np.prod(indices_shape, dtype=int)
indices_array = np.arange(np.random.randint(1, num_elements+1), dtype=np.intc)
np.random.shuffle(indices_array)
indices_array = np.resize(indices_array, num_elements)
idx = 0
for i in range(1, data_input_cnt + 1):
data_in_name = "data{}".format(i)
indices_in_name = "indices{}".format(i)
assert data_in_name in inputs_info, "Test error: inputs_info must contain `{}`".format(data_in_name)
data_shape = inputs_info[data_in_name]
indices_shape = inputs_info[indices_in_name]
inputs_data[data_in_name] = np.random.randint(-50, 50, data_shape)
num_elements_i = np.prod(indices_shape, dtype=int)
inputs_data[indices_in_name] = np.reshape(indices_array[idx:idx+num_elements_i], indices_shape)
idx = idx + num_elements_i
return inputs_data
def create_parallel_dynamic_stitch_net(self, data_input_cnt, shape_of_element, data_type):
tf.compat.v1.reset_default_graph()
# Create the graph and model
with tf.compat.v1.Session() as sess:
indices = []
data = []
data_shape = shape_of_element
indices_shape = []
for i in range(1, data_input_cnt + 1):
indices.append(tf.compat.v1.placeholder(tf.int32, indices_shape, 'indices{}'.format(i)))
data.append(tf.compat.v1.placeholder(data_type, data_shape, 'data{}'.format(i)))
data_shape.insert(0, i)
indices_shape.insert(0, i)
tf.dynamic_stitch(indices, data)
tf.compat.v1.global_variables_initializer()
tf_net = sess.graph_def
return tf_net, None
test_data_basic = [
dict(data_input_cnt=1, shape_of_element=[1], data_type=tf.float32),
dict(data_input_cnt=2, shape_of_element=[2, 2], data_type=tf.float32),
dict(data_input_cnt=3, shape_of_element=[2, 1, 2], data_type=tf.float32),
]
@pytest.mark.parametrize("params", test_data_basic)
@pytest.mark.precommit_tf_fe
def test_parallel_dynamic_stitch_basic(self, params, ie_device, precision, ir_version, temp_dir,
use_new_frontend, use_old_api):
if not use_new_frontend:
pytest.skip("DynamicStitch operation is not supported via legacy frontend.")
self._test(*self.create_parallel_dynamic_stitch_net(**params),
ie_device, precision, ir_version, temp_dir=temp_dir,
use_new_frontend=use_new_frontend, use_old_api=use_old_api)
test_data_different_types = [
dict(data_input_cnt=4, shape_of_element=[3, 2], data_type=tf.float64),
dict(data_input_cnt=2, shape_of_element=[2, 2, 1], data_type=tf.int64),
dict(data_input_cnt=3, shape_of_element=[2, 1, 2, 4], data_type=tf.int32),
]
@pytest.mark.parametrize("params", test_data_different_types)
@pytest.mark.nightly
def test_parallel_dynamic_stitch_different_types(self, params, ie_device, precision, ir_version, temp_dir,
use_new_frontend, use_old_api):
if not use_new_frontend:
pytest.skip("DynamicStitch operation is not supported via legacy frontend.")
self._test(*self.create_parallel_dynamic_stitch_net(**params),
ie_device, precision, ir_version, temp_dir=temp_dir,
use_new_frontend=use_new_frontend, use_old_api=use_old_api)