AUGRU op ov core shell and reference implementation (#12163)
This commit is contained in:
Katarzyna Mitrus authored and 
GitHub committed
1 parent
371e393581
commit
67f4b8ee66
19 files changed
+2267
-23
No files matched your search
@@ -0,0 +1,65 @@
|
||||
// Copyright (C) 2018-2022 Intel Corporation
|
||||
// SPDX-License-Identifier: Apache-2.0
|
||||
//
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <memory>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#include "openvino/op/util/activation_functions.hpp"
|
||||
#include "openvino/op/util/rnn_cell_base.hpp"
|
||||
#include "transformations_visibility.hpp"
|
||||
|
||||
namespace ov {
|
||||
namespace op {
|
||||
namespace internal {
|
||||
///
|
||||
/// \brief AUGRUCell operation.
|
||||
///
|
||||
/// \ingroup ov_ops_cpp_api
|
||||
class TRANSFORMATIONS_API AUGRUCell : public ov::op::util::RNNCellBase {
|
||||
public:
|
||||
OPENVINO_OP("AUGRUCell", "ie_internal_opset", ov::op::util::RNNCellBase);
|
||||
BWDCMP_RTTI_DECLARATION;
|
||||
|
||||
AUGRUCell();
|
||||
AUGRUCell(const Output<Node>& X,
|
||||
const Output<Node>& initial_hidden_state,
|
||||
const Output<Node>& W,
|
||||
const Output<Node>& R,
|
||||
const Output<Node>& B,
|
||||
const Output<Node>& A,
|
||||
size_t hidden_size);
|
||||
|
||||
void validate_and_infer_types() override;
|
||||
bool visit_attributes(AttributeVisitor& visitor) override;
|
||||
std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;
|
||||
|
||||
bool get_linear_before_reset() const {
|
||||
return m_linear_before_reset;
|
||||
}
|
||||
|
||||
private:
|
||||
///
|
||||
/// \brief The Activation function f.
|
||||
///
|
||||
util::ActivationFunction m_activation_f;
|
||||
///
|
||||
/// \brief The Activation function g.
|
||||
///
|
||||
util::ActivationFunction m_activation_g;
|
||||
|
||||
static constexpr std::size_t s_gates_count{3};
|
||||
///
|
||||
/// \brief Control whether or not apply the linear transformation.
|
||||
///
|
||||
/// \note The linear transformation may be applied when computing the output of
|
||||
/// hidden gate. It's done before multiplying by the output of the reset gate.
|
||||
///
|
||||
bool m_linear_before_reset = false;
|
||||
};
|
||||
} // namespace internal
|
||||
} // namespace op
|
||||
} // namespace ov
|
||||
@@ -0,0 +1,54 @@
|
||||
// Copyright (C) 2018-2022 Intel Corporation
|
||||
// SPDX-License-Identifier: Apache-2.0
|
||||
//
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <memory>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#include "openvino/op/util/rnn_cell_base.hpp"
|
||||
#include "transformations_visibility.hpp"
|
||||
|
||||
namespace ov {
|
||||
namespace op {
|
||||
namespace internal {
|
||||
///
|
||||
/// \brief AUGRUSequence operation.
|
||||
///
|
||||
/// \ingroup ov_ops_cpp_api
|
||||
class TRANSFORMATIONS_API AUGRUSequence : public ov::op::util::RNNCellBase {
|
||||
public:
|
||||
OPENVINO_OP("AUGRUSequence", "ie_internal_opset", ov::op::util::RNNCellBase);
|
||||
BWDCMP_RTTI_DECLARATION;
|
||||
|
||||
AUGRUSequence();
|
||||
AUGRUSequence(const Output<Node>& X,
|
||||
const Output<Node>& H_t,
|
||||
const Output<Node>& sequence_lengths,
|
||||
const Output<Node>& W,
|
||||
const Output<Node>& R,
|
||||
const Output<Node>& B,
|
||||
const Output<Node>& A,
|
||||
size_t hidden_size);
|
||||
|
||||
std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;
|
||||
|
||||
void validate_and_infer_types() override;
|
||||
|
||||
bool visit_attributes(AttributeVisitor& visitor) override;
|
||||
bool get_linear_before_reset() const {
|
||||
return m_linear_before_reset;
|
||||
}
|
||||
op::RecurrentSequenceDirection get_direction() const {
|
||||
return m_direction;
|
||||
}
|
||||
|
||||
protected:
|
||||
op::RecurrentSequenceDirection m_direction;
|
||||
bool m_linear_before_reset;
|
||||
};
|
||||
} // namespace internal
|
||||
} // namespace op
|
||||
} // namespace ov
|
||||
@@ -0,0 +1,151 @@
|
||||
// Copyright (C) 2018-2022 Intel Corporation
|
||||
// SPDX-License-Identifier: Apache-2.0
|
||||
//
|
||||
|
||||
#include "ngraph_ops/augru_cell.hpp"
|
||||
|
||||
#include <cmath>
|
||||
|
||||
#include "itt.hpp"
|
||||
|
||||
using namespace std;
|
||||
|
||||
BWDCMP_RTTI_DEFINITION(ov::op::internal::AUGRUCell);
|
||||
|
||||
ov::op::internal::AUGRUCell::AUGRUCell() : m_linear_before_reset(false) {
|
||||
m_activations = {"sigmoid", "tanh"};
|
||||
m_activation_f = get_activation_function(0);
|
||||
m_activation_g = get_activation_function(1);
|
||||
}
|
||||
|
||||
ov::op::internal::AUGRUCell::AUGRUCell(const Output<Node>& X,
|
||||
const Output<Node>& H_t,
|
||||
const Output<Node>& W,
|
||||
const Output<Node>& R,
|
||||
const Output<Node>& B,
|
||||
const Output<Node>& A,
|
||||
size_t hidden_size)
|
||||
: RNNCellBase({X, H_t, W, R, B, A}, hidden_size, 0.f, std::vector<std::string>{"sigmoid", "tanh"}, {}, {}),
|
||||
m_activation_f{get_activation_function(0)},
|
||||
m_activation_g{get_activation_function(1)},
|
||||
m_linear_before_reset{false} {
|
||||
constructor_validate_and_infer_types();
|
||||
}
|
||||
|
||||
bool ov::op::internal::AUGRUCell::visit_attributes(AttributeVisitor& visitor) {
|
||||
INTERNAL_OP_SCOPE(internal_AUGRUCell_visit_attributes);
|
||||
visitor.on_attribute("linear_before_reset", m_linear_before_reset);
|
||||
return op::util::RNNCellBase::visit_attributes(visitor);
|
||||
}
|
||||
|
||||
void ov::op::internal::AUGRUCell::validate_and_infer_types() {
|
||||
INTERNAL_OP_SCOPE(internal_AUGRUCell_validate_and_infer_types);
|
||||
for (const auto& input : inputs()) {
|
||||
if (input.get_partial_shape().rank().is_dynamic()) {
|
||||
set_output_type(0, get_input_element_type(0), PartialShape::dynamic(2));
|
||||
return;
|
||||
}
|
||||
}
|
||||
auto merged_batch_size = Dimension::dynamic();
|
||||
auto merged_hidden_size = Dimension::dynamic();
|
||||
auto result_et = element::dynamic;
|
||||
|
||||
// Get input partial shape for all inputs
|
||||
const auto& x_pshape = get_input_partial_shape(0);
|
||||
const auto& ht_pshape = get_input_partial_shape(1);
|
||||
const auto& w_pshape = get_input_partial_shape(2);
|
||||
const auto& r_pshape = get_input_partial_shape(3);
|
||||
const auto& b_pshape = get_input_partial_shape(4);
|
||||
const auto& a_pshape = get_input_partial_shape(5);
|
||||
|
||||
NODE_VALIDATION_CHECK(this, m_clip == 0.f, "AUGRUCell doesn't support clip other than 0.");
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
m_activations.size() == 2 && m_activations[0] == "sigmoid" && m_activations[1] == "tanh",
|
||||
"AUGRUCell supports only sigmoid for f and tanh for g activation functions.");
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
m_activations_alpha.empty() && m_activations_beta.empty(),
|
||||
"AUGRUCell doesn't support activations_alpha and activations_beta.");
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
m_linear_before_reset == false,
|
||||
"AUGRUCell supports only linear_before_reset equals false.");
|
||||
|
||||
validate_input_rank_dimension({x_pshape, ht_pshape, w_pshape, r_pshape, b_pshape});
|
||||
|
||||
// `A` input shape validation // [batch_size, 1]
|
||||
NODE_VALIDATION_CHECK(this, a_pshape.rank().compatible(2), "'A' input must be a 2D tensor.");
|
||||
if (a_pshape.rank().is_static()) {
|
||||
NODE_VALIDATION_CHECK(this, a_pshape[1].compatible(1), "The last dimension of `A` shape must be equal to `1`.");
|
||||
}
|
||||
|
||||
// Validate input types and save result for output type
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(0)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(1)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(2)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(3)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(4)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(5)),
|
||||
"Element types for inputs do not match.");
|
||||
|
||||
// Merge batch_size dimension across all inputs to evaluate output[0] dimension
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
Dimension::merge(merged_batch_size, merged_batch_size, ht_pshape[0]) &&
|
||||
Dimension::merge(merged_batch_size, merged_batch_size, a_pshape[0]) &&
|
||||
Dimension::merge(merged_batch_size, merged_batch_size, x_pshape[0]),
|
||||
"Dimension batch_size is not matched between inputs.");
|
||||
|
||||
// Merge hidden_size dimension across all inputs to evaluate output[1] dimension
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
Dimension::merge(merged_hidden_size, merged_hidden_size, ht_pshape[1]) &&
|
||||
Dimension::merge(merged_hidden_size, merged_hidden_size, r_pshape[1]),
|
||||
"Dimension hidden_size not matched for R and initial_hidden_state inputs.");
|
||||
|
||||
// Validate hidden_size value for W, B and R inputs
|
||||
if (merged_hidden_size.is_static()) {
|
||||
if (w_pshape[0].is_static()) {
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
w_pshape[0].compatible(merged_hidden_size * s_gates_count),
|
||||
"Parameter hidden_size mistmatched in W input. Current value is: ",
|
||||
w_pshape[0].get_length(),
|
||||
", expected: ",
|
||||
merged_hidden_size.get_length() * s_gates_count,
|
||||
".");
|
||||
}
|
||||
|
||||
if (r_pshape[0].is_static()) {
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
r_pshape[0].compatible(merged_hidden_size * s_gates_count),
|
||||
"Parameter hidden_size mistmatched in R input. Current value is: ",
|
||||
r_pshape[0].get_length(),
|
||||
", expected: ",
|
||||
merged_hidden_size.get_length() * s_gates_count,
|
||||
".");
|
||||
}
|
||||
|
||||
if (b_pshape[0].is_static()) {
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
b_pshape[0].compatible(merged_hidden_size * (s_gates_count + m_linear_before_reset)),
|
||||
"Parameter hidden_size mistmatched in B input. Current value is: ",
|
||||
b_pshape[0].get_length(),
|
||||
", expected: ",
|
||||
merged_hidden_size.get_length() * (s_gates_count + m_linear_before_reset),
|
||||
".");
|
||||
}
|
||||
}
|
||||
|
||||
// Set output size, type and shape
|
||||
set_output_size(1);
|
||||
set_output_type(0, result_et, {merged_batch_size, merged_hidden_size});
|
||||
}
|
||||
|
||||
shared_ptr<ov::Node> ov::op::internal::AUGRUCell::clone_with_new_inputs(const OutputVector& new_args) const {
|
||||
INTERNAL_OP_SCOPE(internal_AUGRUCell_clone_with_new_inputs);
|
||||
check_new_args_count(this, new_args);
|
||||
return make_shared<AUGRUCell>(new_args.at(0),
|
||||
new_args.at(1),
|
||||
new_args.at(2),
|
||||
new_args.at(3),
|
||||
new_args.at(4),
|
||||
new_args.at(5),
|
||||
get_hidden_size());
|
||||
}
|
||||
@@ -0,0 +1,189 @@
|
||||
// Copyright (C) 2018-2022 Intel Corporation
|
||||
// SPDX-License-Identifier: Apache-2.0
|
||||
//
|
||||
|
||||
#include "ngraph_ops/augru_sequence.hpp"
|
||||
|
||||
#include "itt.hpp"
|
||||
#include "ngraph/op/util/recurrent_sequence.hpp"
|
||||
|
||||
using namespace std;
|
||||
|
||||
BWDCMP_RTTI_DEFINITION(ov::op::internal::AUGRUSequence);
|
||||
|
||||
ov::op::internal::AUGRUSequence::AUGRUSequence()
|
||||
: m_direction(op::RecurrentSequenceDirection::FORWARD),
|
||||
m_linear_before_reset(false) {}
|
||||
|
||||
ov::op::internal::AUGRUSequence::AUGRUSequence(const Output<Node>& X,
|
||||
const Output<Node>& H_t,
|
||||
const Output<Node>& sequence_lengths,
|
||||
const Output<Node>& W,
|
||||
const Output<Node>& R,
|
||||
const Output<Node>& B,
|
||||
const Output<Node>& A,
|
||||
std::size_t hidden_size)
|
||||
: RNNCellBase({X, H_t, sequence_lengths, W, R, B, A},
|
||||
hidden_size,
|
||||
0.f,
|
||||
std::vector<std::string>{"sigmoid", "tanh"},
|
||||
{},
|
||||
{}),
|
||||
m_direction(op::RecurrentSequenceDirection::FORWARD),
|
||||
m_linear_before_reset(false) {
|
||||
constructor_validate_and_infer_types();
|
||||
}
|
||||
|
||||
void ov::op::internal::AUGRUSequence::validate_and_infer_types() {
|
||||
INTERNAL_OP_SCOPE(internal_AUGRUSequence_validate_and_infer_types);
|
||||
for (const auto& input : inputs()) {
|
||||
if (input.get_partial_shape().rank().is_dynamic()) {
|
||||
set_output_type(0, get_input_element_type(0), PartialShape::dynamic(4));
|
||||
set_output_type(1, get_input_element_type(0), PartialShape::dynamic(3));
|
||||
return;
|
||||
}
|
||||
}
|
||||
|
||||
auto gru_seq_gates_count = 3;
|
||||
auto merged_batch_size = Dimension::dynamic();
|
||||
auto merged_hidden_size = Dimension::dynamic();
|
||||
auto merged_num_directions = Dimension::dynamic();
|
||||
auto result_et = element::dynamic;
|
||||
|
||||
auto x_pshape = get_input_partial_shape(0);
|
||||
auto ht_pshape = get_input_partial_shape(1);
|
||||
auto sl_pshape = get_input_partial_shape(2);
|
||||
auto w_pshape = get_input_partial_shape(3);
|
||||
auto r_pshape = get_input_partial_shape(4);
|
||||
auto b_pshape = get_input_partial_shape(5);
|
||||
auto a_pshape = get_input_partial_shape(6);
|
||||
|
||||
NODE_VALIDATION_CHECK(this, m_clip == 0.f, "AUGRUSequence doesn't support clip other than 0.");
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
m_activations.size() == 2 && m_activations[0] == "sigmoid" && m_activations[1] == "tanh",
|
||||
"AUGRUSequence supports only sigmoid for f and tanh for g activation functions.");
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
m_activations_alpha.empty() && m_activations_beta.empty(),
|
||||
"AUGRUSequence doesn't support activations_alpha and activations_beta.");
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
m_direction == op::RecurrentSequenceDirection::FORWARD,
|
||||
"AUGRUSequence supports only forward direction.");
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
m_linear_before_reset == false,
|
||||
"AUGRUSequence supports only linear_before_reset equals false.");
|
||||
|
||||
ngraph::op::util::validate_seq_input_rank_dimension(
|
||||
{x_pshape, ht_pshape, sl_pshape, w_pshape, r_pshape, b_pshape, a_pshape});
|
||||
|
||||
// Validate input types and save result for output type
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(0)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(1)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(3)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(4)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(5)) &&
|
||||
element::Type::merge(result_et, result_et, get_input_element_type(6)),
|
||||
"Element types for inputs do not match.");
|
||||
|
||||
// Merge batch_size dimension across all inputs to evaluate output[0] dimension
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
Dimension::merge(merged_batch_size, merged_batch_size, ht_pshape[0]) &&
|
||||
Dimension::merge(merged_batch_size, merged_batch_size, a_pshape[0]) &&
|
||||
Dimension::merge(merged_batch_size, merged_batch_size, x_pshape[0]) &&
|
||||
Dimension::merge(merged_batch_size, merged_batch_size, sl_pshape[0]),
|
||||
"Dimension batch_size is not matched between inputs.");
|
||||
|
||||
// A input shape validation // [batch_size, seq_length, 1]
|
||||
NODE_VALIDATION_CHECK(this, a_pshape.rank().compatible(3), "'A' input must be a 3D tensor.");
|
||||
if (a_pshape.rank().is_static()) {
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
a_pshape[1].compatible(x_pshape[1]),
|
||||
"Dimension `seq_length` must be the same for `X` and `A` inputs.");
|
||||
NODE_VALIDATION_CHECK(this, a_pshape[2].compatible(1), "The last dimension of `A` shape must be equal to `1`.");
|
||||
}
|
||||
|
||||
// Merge hidden_size dimension across all inputs to evaluate output dimension
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
Dimension::merge(merged_hidden_size, merged_hidden_size, ht_pshape[2]) &&
|
||||
Dimension::merge(merged_hidden_size, merged_hidden_size, r_pshape[2]),
|
||||
"Parameter hidden_size not matched AUGRUSequence.");
|
||||
|
||||
// Merge num_directions dimension across all inputs to evaluate output dimension
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
Dimension::merge(merged_num_directions, merged_num_directions, ht_pshape[1]) &&
|
||||
Dimension::merge(merged_num_directions, merged_num_directions, w_pshape[0]) &&
|
||||
Dimension::merge(merged_num_directions, merged_num_directions, r_pshape[0]) &&
|
||||
Dimension::merge(merged_num_directions, merged_num_directions, b_pshape[0]),
|
||||
"Parameter num_directions not matched in AUGRUSequence.");
|
||||
|
||||
auto valid_num_directions = 1; // AUGRUSequence supports only Forward direction
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
Dimension::merge(merged_num_directions, merged_num_directions, valid_num_directions),
|
||||
"Parameter 'num_directions' doesn't match with direction '",
|
||||
m_direction,
|
||||
"' in AUGRUSequence. Expected ",
|
||||
valid_num_directions,
|
||||
", actual ",
|
||||
merged_num_directions);
|
||||
|
||||
// Validate hidden_size value for W, R, B inputs
|
||||
if (merged_hidden_size.is_static()) {
|
||||
if (w_pshape[1].is_static()) {
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
w_pshape[1].compatible(merged_hidden_size * gru_seq_gates_count),
|
||||
"Parameter hidden_size mistmatched in W input. Current value is: ",
|
||||
w_pshape[1].get_length(),
|
||||
", expected: ",
|
||||
merged_hidden_size.get_length() * gru_seq_gates_count,
|
||||
".");
|
||||
}
|
||||
|
||||
if (r_pshape[1].is_static()) {
|
||||
NODE_VALIDATION_CHECK(this,
|
||||
r_pshape[1].compatible(merged_hidden_size * gru_seq_gates_count),
|
||||
"Parameter hidden_size mistmatched in R input. Current value is: ",
|
||||
r_pshape[1].get_length(),
|
||||
", expected: ",
|
||||
merged_hidden_size.get_length() * gru_seq_gates_count,
|
||||
".");
|
||||
}
|
||||
|
||||
if (b_pshape[1].is_static()) {
|
||||
NODE_VALIDATION_CHECK(
|
||||
this,
|
||||
b_pshape[1].compatible(merged_hidden_size *
|
||||
(m_linear_before_reset ? (gru_seq_gates_count + 1) : gru_seq_gates_count)),
|
||||
"Parameter hidden_size mistmatched in B input. Current value is: ",
|
||||
b_pshape[1].get_length(),
|
||||
", expected: ",
|
||||
merged_hidden_size.get_length() *
|
||||
(m_linear_before_reset ? (gru_seq_gates_count + 1) : gru_seq_gates_count),
|
||||
".");
|
||||
}
|
||||
}
|
||||
|
||||
// Set output size, type and shape
|
||||
set_output_size(2);
|
||||
set_output_type(0, result_et, {merged_batch_size, merged_num_directions, x_pshape[1], merged_hidden_size});
|
||||
set_output_type(1, result_et, {merged_batch_size, merged_num_directions, merged_hidden_size});
|
||||
}
|
||||
|
||||
bool ov::op::internal::AUGRUSequence::visit_attributes(AttributeVisitor& visitor) {
|
||||
INTERNAL_OP_SCOPE(internal_AUGRUSequence_visit_attributes);
|
||||
visitor.on_attribute("direction", m_direction);
|
||||
visitor.on_attribute("linear_before_reset", m_linear_before_reset);
|
||||
return op::util::RNNCellBase::visit_attributes(visitor);
|
||||
}
|
||||
|
||||
shared_ptr<ov::Node> ov::op::internal::AUGRUSequence::clone_with_new_inputs(const OutputVector& new_args) const {
|
||||
INTERNAL_OP_SCOPE(internal_AUGRUSequence_clone_with_new_inputs);
|
||||
check_new_args_count(this, new_args);
|
||||
return make_shared<ov::op::internal::AUGRUSequence>(new_args.at(0),
|
||||
new_args.at(1),
|
||||
new_args.at(2),
|
||||
new_args.at(3),
|
||||
new_args.at(4),
|
||||
new_args.at(5),
|
||||
new_args.at(6),
|
||||
get_hidden_size());
|
||||
}
|
||||
Reference in new issue
Block a user