Adding Numpy Style Transpose to the Builder (#271)

* Add numpy_transpose to the builder for numpy-stype transpose operations

* fix docstring

* make sure to throw the error

src/ngraph/CMakeLists.txt

@@ -15,6 +15,7 @@ set (SRC
     autodiff/adjoints.cpp
     builder/autobroadcast.cpp
     builder/reduce_ops.cpp
+    builder/numpy_transpose.cpp
     descriptor/input.cpp
     descriptor/layout/dense_tensor_view_layout.cpp
     descriptor/layout/tensor_view_layout.cpp

src/ngraph/builder/numpy_transpose.cpp

+/*
+ Copyright 2017 Nervana Systems Inc.
+ 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 "ngraph/builder/numpy_transpose.hpp"
+#include <sstream>
+#include "ngraph/except.hpp"
+#include "ngraph/ops/reshape.hpp"
+
+namespace ngraph
+{
+    void numpy_transpose_error(const AxisVector& order, const Shape& in_shape)
+    {
+        std::ostringstream os;
+        os << "The axes order ";
+        os << "[ " << ngraph::join(order) << " ]";
+        os << " is incompatible with the input shape ";
+        os << "[ " << ngraph::join(in_shape) << " ]";
+        os << " during numpy_transpose.";
+        throw ngraph_error(os.str());
+    }
+
+    namespace builder
+    {
+        std::shared_ptr<Node> numpy_transpose(const std::shared_ptr<Node>& node, AxisVector order)
+        {
+            auto in_shape = node->get_shape();
+            // default, reverse the order of the axes
+            if (order.size() == 0)
+            {
+                auto n = in_shape.size();
+                order = AxisVector(n);
+                std::generate(order.begin(), order.end(), [&n]() { return --n; });
+            }
+            else if (order.size() == in_shape.size())
+            {
+                // validate that the axes order is valid, i.e., unique and the right size
+                std::unordered_set<ngraph::AxisVector::value_type> axes;
+                for (auto o : order)
+                {
+                    if (o >= 0 && o < in_shape.size() && !axes.count(o))
+                    {
+                        axes.insert(o);
+                    }
+                    else
+                    {
+                        numpy_transpose_error(order, in_shape);
+                    }
+                }
+            }
+            else
+            {
+                numpy_transpose_error(order, in_shape);
+            }
+
+            // create output shape
+            Shape out_shape;
+            for (size_t i = 0; i < in_shape.size(); ++i)
+                out_shape.push_back(in_shape[order[i]]);
+
+            // do the reshaping with the order
+            return std::make_shared<ngraph::op::Reshape>(node, order, out_shape);
+        }
+
+    } // namespace builder
+} // namespace ngraph

src/ngraph/builder/numpy_transpose.hpp

+/*
+ Copyright 2017 Nervana Systems Inc.
+ 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.
+*/
+
+#pragma once
+
+#include "ngraph/common.hpp"
+#include "ngraph/function.hpp"
+#include "ngraph/node.hpp"
+#include "ngraph/ops/constant.hpp"
+#include "ngraph/ops/parameter.hpp"
+#include "ngraph/ops/reduce.hpp"
+#include "ngraph/types/type.hpp"
+
+namespace ngraph
+{
+    namespace builder
+    {
+        /// \brief Implement's Numpy's multidimensional transpose op. Doubles as DimShuffle.
+        ///
+        /// If `order` is empty, the vector is transposed by reversing it's axes, i.e.
+        ///
+        /// shape [1,2,4] becomes shape [4,2,1]
+        ///
+        /// If `order` is provided, it should be a vector of unique axis positions ranging
+        /// from 0 to N-1, when N is the length of the input shape.  In this case, numpy_transpose acts
+        /// like dimshuffle, so
+        ///
+        /// shape [1,2,4] with order [1,2,0] becomes shape [2,4,1]
+        ///
+        /// |                  | Type                                  | Description                                             |
+        /// | ---------------- | ------------------------------------- | ------------------------------------------------------- |
+        /// | `node`           | \f$E[d_0,\dots,d_{n-1}]~(n \geq 0)\f$ | An input tensor of any shape                            |
+        /// | `order`          | AxisVector (empty default)            | The axes to eliminate through reduction (0 indexed).    |
+        ///
+        /// ## Output
+        ///
+        /// | Type                                                                         | Description                                                                                                      |
+        /// | ---------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------- |
+        /// | \f$E[d_{n-1},\dots,d_0)]\textit{ or }E[d_{order[0]},\dots,d_{order[n-1]}]\f$ | The tensor \f$T\f$, where \f$T\f$ is the input tensor with the axes reordered via Numpy Transpose rules          |
+        std::shared_ptr<Node> numpy_transpose(const std::shared_ptr<Node>& node,
+                                              AxisVector order = {});
+    } // namespace builder
+} // namespace ngraph

src/ngraph/ngraph.hpp

@@ -42,6 +42,7 @@
 ///        recipes, for example auto-broadcast.
 
 #include "ngraph/builder/autobroadcast.hpp"
+#include "ngraph/builder/numpy_transpose.hpp"
 #include "ngraph/builder/reduce_ops.hpp"
 #include "ngraph/common.hpp"
 #include "ngraph/descriptor/buffer.hpp"

test/CMakeLists.txt

@@ -22,10 +22,10 @@ include_directories(
     )
 
 set (SRC
-    builder_autobroadcast.cpp
-    builder_reduce_ops.cpp
     autodiff.cpp
     build_graph.cpp
+    builder.cpp
+    builder_autobroadcast.cpp
     copy.cpp
     eigen.cpp
     element_type.cpp

test/builder_reduce_ops.cpp → test/builder.cpp

@@ -90,20 +90,20 @@ std::shared_ptr<ngraph::runtime::TensorView> make_reduce_result_false(
     return result;
 }
 
-TEST(builder_reduce_ops, l2_norm)
+TEST(builder, l2_norm)
 {
     auto result = make_reduce_result(builder::l2_norm);
     ASSERT_TRUE(
         all_close((vector<float>{5.9160797831f, 7.48331477355f}), result->get_vector<float>()));
 }
 
-TEST(builder_reduce_ops, mean)
+TEST(builder, mean)
 {
     auto result = make_reduce_result(builder::mean);
     ASSERT_TRUE(all_close((vector<float>{3, 4}), result->get_vector<float>()));
 }
 
-TEST(builder_reduce_ops, std_dev)
+TEST(builder, std_dev)
 {
     auto result = make_reduce_result_false(builder::std_dev);
     ASSERT_TRUE(
@@ -112,7 +112,7 @@ TEST(builder_reduce_ops, std_dev)
     ASSERT_TRUE(all_close((vector<float>{2, 2}), result->get_vector<float>()));
 }
 
-TEST(builder_reduce_ops, variance)
+TEST(builder, variance)
 {
     auto result = make_reduce_result_false(builder::variance);
     ASSERT_TRUE(
@@ -120,3 +120,31 @@ TEST(builder_reduce_ops, variance)
     result = make_reduce_result_true(builder::variance);
     ASSERT_TRUE(all_close((vector<float>{4, 4}), result->get_vector<float>()));
 }
+
+TEST(builder, numpy_transpose)
+{
+    // 2D Transpose
+    Shape shape{2, 4};
+    auto param = std::make_shared<op::Parameter>(ngraph::element::Float32::element_type(), shape);
+    auto transposed = std::dynamic_pointer_cast<op::Reshape>(builder::numpy_transpose(param));
+    EXPECT_EQ(Shape({4, 2}), transposed->get_output_shape());
+
+    // Multidimensional Transpose
+    shape = Shape{2, 4, 8};
+    param = std::make_shared<op::Parameter>(ngraph::element::Float32::element_type(), shape);
+    transposed = std::dynamic_pointer_cast<op::Reshape>(builder::numpy_transpose(param));
+    EXPECT_EQ(Shape({8, 4, 2}), transposed->get_output_shape());
+
+    // Dimshuffle
+    shape = Shape{2, 4, 8};
+    param = std::make_shared<op::Parameter>(ngraph::element::Float32::element_type(), shape);
+    transposed = std::dynamic_pointer_cast<op::Reshape>(
+        builder::numpy_transpose(param, AxisVector{2, 0, 1}));
+    EXPECT_EQ(Shape({8, 2, 4}), transposed->get_output_shape());
+
+    // Bad Orders
+    EXPECT_ANY_THROW(
+        std::dynamic_pointer_cast<op::Reshape>(builder::numpy_transpose(param, AxisVector{2})));
+    EXPECT_ANY_THROW(std::dynamic_pointer_cast<op::Reshape>(
+        builder::numpy_transpose(param, AxisVector{2, 2, 1})));
+}