[ONNX] Handle optional outputs for Dropout and MaxPool (#4143)

inference-engine/src/transformations/src/transformations/convert_precision.cpp

@@ -404,9 +404,7 @@ bool fuse_type_to_constant(std::shared_ptr<Node> & node, element::Type to, const
         }
 
         new_const->validate_and_infer_types();
-        if (constant->get_output_target_inputs(0).size() == consumers.size()) {
-            new_const->set_friendly_name(constant->get_friendly_name());
-        }
+        new_const->set_friendly_name(constant->get_friendly_name());
     }
     return false;
 }

inference-engine/tests/functional/inference_engine/transformations/convert_precision.cpp

@@ -565,8 +565,10 @@ TEST(TransformationTests, ConvertPrecision_Variables) {
 template <typename From, typename To>
 void constant_convert_test(element::Type_t type_from, element::Type_t type_to, From value, To expected) {
     std::shared_ptr<ngraph::Function> f(nullptr);
+    std::string expected_friendly_name;
     {
         auto c = opset4::Constant::create(type_from, Shape{}, {value});
+        expected_friendly_name = c->get_friendly_name();
         f = std::make_shared<Function>(NodeVector{c}, ParameterVector{});
 
         pass::Manager manager;
@@ -576,6 +578,7 @@ void constant_convert_test(element::Type_t type_from, element::Type_t type_to, F
     auto ops = f->get_ordered_ops();
     auto c = std::dynamic_pointer_cast<opset4::Constant>(ops[0]);
     ASSERT_NE(c, nullptr);
+    ASSERT_EQ(c->get_friendly_name(), expected_friendly_name);
 
     auto actual = c->cast_vector<To>()[0];
     ASSERT_EQ(expected, actual);
@@ -622,3 +625,8 @@ TEST(TransformationTests, ConvertPrecision_ConstantConversion_U32MaxToI32) {
 TEST(TransformationTests, ConvertPrecision_ConstantConversion_U32ToI32) {
     constant_convert_test(element::Type_t::u32, element::Type_t::i32, 42, 42);
 }
+
+TEST(TransformationTests, ConvertPrecision_ConstantConversion_BoolToU8) {
+    constant_convert_test(element::Type_t::boolean, element::Type_t::u8, true, 1);
+    constant_convert_test(element::Type_t::boolean, element::Type_t::u8, false, 0);
+}

ngraph/frontend/onnx_import/src/core/graph.cpp

@@ -223,7 +223,11 @@ namespace ngraph
             OutputVector results;
             for (const auto& output : m_graph_proto->output())
             {
-                results.emplace_back(get_ng_node_from_cache(output.name()));
+                const auto& ng_output = get_ng_node_from_cache(output.name());
+                if (!ngraph::op::is_null(ng_output)) // ignore optional outputs
+                {
+                    results.emplace_back(ng_output);
+                }
             }
             return results;
         }

ngraph/frontend/onnx_import/src/op/dropout.cpp

@@ -0,0 +1,107 @@
+//*****************************************************************************
+// Copyright 2017-2021 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+#include <memory>
+
+#include "core/null_node.hpp"
+#include "default_opset.hpp"
+#include "exceptions.hpp"
+#include "ngraph/log.hpp"
+#include "ngraph/node.hpp"
+#include "op/dropout.hpp"
+
+namespace ngraph
+{
+    namespace onnx_import
+    {
+        namespace op
+        {
+            namespace
+            {
+                OutputVector build_dropout(const Node& node, bool training_mode)
+                {
+                    CHECK_VALID_NODE(
+                        node, !training_mode, "Training mode is not supported for Dropout op");
+
+                    const auto input_data = node.get_ng_inputs().at(0);
+                    const bool return_mask = node.get_outputs_size() > 1;
+
+                    if (return_mask)
+                    {
+                        const auto mask = std::make_shared<default_opset::Broadcast>(
+                            default_opset::Constant::create(
+                                ngraph::element::boolean, Shape{}, {true}),
+                            std::make_shared<default_opset::ShapeOf>(input_data));
+                        return {input_data, mask};
+                    }
+                    else
+                    {
+                        return {input_data};
+                    }
+                }
+            }
+
+            namespace set_12
+            {
+                OutputVector dropout(const Node& node)
+                {
+                    const auto ng_inputs = node.get_ng_inputs();
+                    // seed attribute and ratio input are ignored because traning mode is not
+                    // supported anyway
+                    bool training_mode = false; // default value
+                    if (ng_inputs.size() > 2 && !ngraph::op::is_null(ng_inputs.at(2)))
+                    {
+                        CHECK_VALID_NODE(
+                            node,
+                            ngraph::op::is_constant(ng_inputs.at(2).get_node_shared_ptr()),
+                            "Non-constant training_mode input is not supported.");
+                        training_mode = as_type_ptr<default_opset::Constant>(
+                                            ng_inputs.at(2).get_node_shared_ptr())
+                                            ->cast_vector<bool>()[0];
+                    }
+                    return build_dropout(node, training_mode);
+                }
+            } // namespace set_12
+
+            namespace set_7
+            {
+                OutputVector dropout(const Node& node)
+                {
+                    // "is_test" attribute was removed
+                    // ratio attribute is ignored because traning mode is not supported
+                    const bool training_mode = false;
+
+                    return build_dropout(node, training_mode);
+                }
+            } // namespace set_7
+
+            namespace set_1
+            {
+                OutputVector dropout(const Node& node)
+                {
+                    // legacy consumed_inputs attribute ignored
+                    // ratio attribute is ignored because traning mode is not supported
+                    const bool training_mode = !node.get_attribute_value<int64_t>("is_test", 0);
+
+                    return build_dropout(node, training_mode);
+                }
+            } // namespace set_1
+
+        } // namespace op
+
+    } // namespace onnx_import
+
+} // namespace ngraph

ngraph/frontend/onnx_import/src/op/dropout.hpp

@@ -16,10 +16,6 @@
 
 #pragma once
 
-#include <memory>
-
-#include "core/null_node.hpp"
-#include "ngraph/node.hpp"
 #include "onnx_import/core/node.hpp"
 
 namespace ngraph
@@ -28,15 +24,19 @@ namespace ngraph
     {
         namespace op
         {
+            namespace set_12
+            {
+                OutputVector dropout(const Node& node);
+            } // namespace set_12
+
+            namespace set_7
+            {
+                OutputVector dropout(const Node& node);
+            } // namespace set_7
+
             namespace set_1
             {
-                inline OutputVector dropout(const Node& node)
-                {
-                    // First value is actual output of Dropout,
-                    // the second one is just a placeholder for optional trailing output.
-                    return {node.get_ng_inputs().at(0).get_node_shared_ptr(),
-                            std::make_shared<NullNode>()};
-                }
+                OutputVector dropout(const Node& node);
             } // namespace set_1
 
         } // namespace op

ngraph/frontend/onnx_import/src/op/max_pool.cpp

@@ -17,6 +17,7 @@
 #include <memory>
 
 #include "core/null_node.hpp"
+#include "ngraph/log.hpp"
 #include "ngraph/op/max_pool.hpp"
 #include "op/max_pool.hpp"
 #include "utils/pooling_factory.hpp"
@@ -31,6 +32,11 @@ namespace ngraph
             {
                 OutputVector max_pool(const Node& node)
                 {
+                    if (node.get_outputs_size() > 1)
+                    {
+                        NGRAPH_WARN
+                            << "Indices output is not supported for MaxPooling and was ignored";
+                    }
                     auto max_pool = pooling::PoolingFactory(node).make_max_pool();
                     max_pool.emplace_back(std::make_shared<NullNode>()); // Indices (optional)
                     return max_pool;

ngraph/frontend/onnx_import/src/ops_bridge.cpp

@@ -343,6 +343,8 @@ namespace ngraph
             REGISTER_OPERATOR("Div", 1, div);
             REGISTER_OPERATOR("Div", 7, div);
             REGISTER_OPERATOR("Dropout", 1, dropout);
+            REGISTER_OPERATOR("Dropout", 7, dropout);
+            REGISTER_OPERATOR("Dropout", 12, dropout);
             REGISTER_OPERATOR("Elu", 1, elu);
             REGISTER_OPERATOR("Equal", 1, equal);
             REGISTER_OPERATOR("Erf", 1, erf);

ngraph/python/tests/__init__.py

@@ -107,8 +107,6 @@ def xfail_test(reason="Mark the test as expected to fail", strict=True):
 xfail_issue_38699 = xfail_test(reason="RuntimeError: nGraph does not support the following ONNX operations:"
                                       "ai.onnx.preview.training.Gradient")
 xfail_issue_38701 = xfail_test(reason="RuntimeError: unsupported element type: STRING")
-xfail_issue_38705 = xfail_test(reason="IndexError: deque::_M_range_check: __n (which is 0)"
-                                      ">= this->size() (which is 0)")
 xfail_issue_38706 = xfail_test(reason="RuntimeError: output_3.0 has zero dimension which is not allowed")
 xfail_issue_38707 = xfail_test(reason="RuntimeError: nGraph does not support the following ONNX operations:"
                                       "SoftmaxCrossEntropyLoss")
@@ -152,7 +150,7 @@ def xfail_test(reason="Mark the test as expected to fail", strict=True):
                                       "ai.onnx.preview.training.Adagrad")
 xfail_issue_38736 = xfail_test(reason="RuntimeError: nGraph does not support the following ONNX operations:"
                                       "NegativeLogLikelihoodLoss")
-xfail_issue_45177 = xfail_test(reason="RuntimeError: axes has zero dimension which is not allowed")
+xfail_issue_48052 = xfail_test(reason="Dropout op is not supported in traning mode")
 xfail_issue_45180 = xfail_test(reason="RuntimeError: Unsupported dynamic op: ReduceSum")
 xfail_issue_44839 = xfail_test(reason="Huge computation missmatch")
 xfail_issue_44848 = xfail_test(reason="E   Unsupported dynamic op: Range")
@@ -176,6 +174,7 @@ def xfail_test(reason="Mark the test as expected to fail", strict=True):
 xfail_issue_47330 = xfail_test(reason="RuntimeError: Eltwise node with name `[name]` doesn't support "
                                       "FP64 precision.")
 xfail_issue_47337 = xfail_test(reason="RuntimeError: Unsupported dynamic ops: v1::OneHot")
+xfail_issue_33593 = xfail_test(reason="Current implementation of MaxPool doesn't support indices output")
 
 # Model MSFT issues:
 xfail_issue_37957 = xfail_test(reason="RuntimeError: nGraph does not support the following ONNX operations:"

ngraph/python/tests/runtime.py

@@ -18,10 +18,10 @@
 from typing import Dict, List, Union
 
 import numpy as np
-from openvino.inference_engine import IECore, IENetwork, Blob
+from openvino.inference_engine import IECore, IENetwork, Blob, DataPtr
 
 from ngraph.exceptions import UserInputError
-from ngraph.impl import Function, Node, PartialShape
+from ngraph.impl import Function, Node, PartialShape, Type
 from ngraph.opset1.ops import result
 from ngraph.utils.types import NumericData, get_shape, get_dtype
 
@@ -55,6 +55,18 @@ def _convert_inputs(cnn_network: IENetwork) -> None:
             pass
 
 
+def apply_ng_type(output: DataPtr, ng_type: Type):
+    ng_ie_supported_type_map = {
+        Type.boolean.get_type_name(): "BOOL",
+        Type.f32.get_type_name(): "FP32",
+        Type.i8.get_type_name(): "I8",
+        Type.i32.get_type_name(): "I32",
+        Type.u8.get_type_name(): "U8",
+    }
+    if ng_type.get_type_name() in ng_ie_supported_type_map:
+        output.precision = ng_ie_supported_type_map[ng_type.get_type_name()]
+
+
 class Runtime(object):
     """Represents an nGraph runtime environment."""
 
@@ -103,18 +115,30 @@ def __repr__(self) -> str:
         params_string = ", ".join([param.name for param in self.parameters])
         return "<Computation: {}({})>".format(self.function.get_name(), params_string)
 
-    def __get_ie_output_blob_buffer(self, output_blobs: Dict[str, Blob], ng_result: result) -> np.ndarray:
+    def __get_ie_output_blob_name(self, outputs: Dict, ng_result: result) -> str:
         if len(self.results) == 1:
-            return next(iter(output_blobs.values())).buffer
+            return next(iter(outputs.keys()))
         else:
             prev_layer = ng_result.input(0).get_source_output()
             out_name = prev_layer.get_node().get_friendly_name()
             if prev_layer.get_node().get_output_size() != 1:
                 out_name += "." + str(prev_layer.get_index())
-            return output_blobs[out_name].buffer
+            return out_name
+
+    def __get_ie_output_blob_buffer(self, output_blobs: Dict[str, Blob], ng_result: result) -> np.ndarray:
+        out_name = self.__get_ie_output_blob_name(output_blobs, ng_result)
+        return output_blobs[out_name].buffer
 
     def __call__(self, *input_values: NumericData) -> List[NumericData]:
         """Run computation on input values and return result."""
+        # Input validation
+        if len(input_values) < len(self.parameters):
+            raise UserInputError(
+                "Expected %s params, received not enough %s values.", len(self.parameters), len(input_values)
+            )
+        # ignore not needed input values
+        input_values = input_values[:len(self.parameters)]
+
         input_values = [np.array(input_value) for input_value in input_values]
         input_shapes = [get_shape(input_value) for input_value in input_values]
 
@@ -131,13 +155,13 @@ def __call__(self, *input_values: NumericData) -> List[NumericData]:
         else:
             cnn_network = self.network_cache[str(input_shapes)]
 
+        # set output blobs precission based on nG results
+        for ng_result in self.results:
+            ie_out_name = self.__get_ie_output_blob_name(cnn_network.outputs, ng_result)
+            apply_ng_type(cnn_network.outputs[ie_out_name], ng_result.get_output_element_type(0))
+
         executable_network = self.runtime.backend.load_network(cnn_network, self.runtime.backend_name)
 
-        # Input validation
-        if len(input_values) != len(self.parameters):
-            raise UserInputError(
-                "Expected %s parameters, received %s.", len(self.parameters), len(input_values)
-            )
         for parameter, input in zip(self.parameters, input_values):
             parameter_shape = parameter.get_output_partial_shape(0)
             input_shape = PartialShape(input.shape)

ngraph/python/tests/test_ngraph/test_basic.py

@@ -80,15 +80,15 @@ def test_simple_computation_on_ndarrays(dtype):
 
     value_a = np.array([[1, 2], [3, 4]], dtype=dtype)
     value_b = np.array([[5, 6], [7, 8]], dtype=dtype)
-    value_c = np.array([[9, 10], [11, 12]], dtype=dtype)
+    value_c = np.array([[2, 3], [4, 5]], dtype=dtype)
     result = computation(value_a, value_b, value_c)
-    assert np.allclose(result, np.array([[54, 80], [110, 144]], dtype=dtype))
+    assert np.allclose(result, np.array([[12, 24], [40, 60]], dtype=dtype))
 
-    value_a = np.array([[13, 14], [15, 16]], dtype=dtype)
-    value_b = np.array([[17, 18], [19, 20]], dtype=dtype)
-    value_c = np.array([[21, 22], [23, 24]], dtype=dtype)
+    value_a = np.array([[9, 10], [11, 12]], dtype=dtype)
+    value_b = np.array([[13, 14], [15, 16]], dtype=dtype)
+    value_c = np.array([[5, 4], [3, 2]], dtype=dtype)
     result = computation(value_a, value_b, value_c)
-    assert np.allclose(result, np.array([[630, 704], [782, 864]], dtype=dtype))
+    assert np.allclose(result, np.array([[110, 96], [78, 56]], dtype=dtype))
 
 
 def test_serialization():

ngraph/python/tests/test_ngraph/test_ops_reshape.py

@@ -223,7 +223,6 @@ def test_reshape_v1():
     assert np.allclose(result, expected)
 
 
-@xfail_issue_40957
 def test_shape_of():
     input_tensor = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype=np.float32)