Implement function retrieval APIs; Add documentation for functions (o…

…nnx#1112) * Add for test function build * Add pybind functions for getFunction API * Implement FC function testcase * Lint code using clang-format * Disable warning in covert protobuf::int64 to int * Update doc generation scripts and functions doc * Trimming some comments * Fix doc generation scripts for typecheck * Improve cmake scripts for unittest project dependencies and fix linux build issues * Lint python code * Update API to support functions with multiple versions * Update line ending * Add changelogs for functions * Refactor and lint code * Fix bad linebreaks in markdown files generated * Resolve PR comments; Refine Function type check implementation * Add checker for function version; Bump version in onnx_domain * Update doc_string and generated md files * Add helper function; Update documents for MVN function * Fix a bug in MVN function definition * Experimental: add python test case for MVN * Update line ends * Add attribute mvn testcase * Update generated files on MVN function * Rename buildnode function * Remove node part in generated files; Reformat doc_string for MVN * Fix changelog generation script * Update MVN testcase to caculate expected output on the fly * Rephrase description for MVN function * Formatting docstring for MVN function * Refactor helper function code; Add epsilon before division * Move the helper functions into common folder * Fix typo in function docstring * Refactor model helper function; Add version check; Fix typo * Switch to use FunctionProto class generated in python * Remove unnecessary ignore flag * Minor fix in c++ tests; Generated function docs for ML domain * Refactor code to avoid hardcode domain name * Resolve PR comments
hariharans29 · Jul 19, 2018 · 828d3c3 · 828d3c3
1 parent 5b9f9c1
commit 828d3c3
Show file tree

Hide file tree

Showing 16 changed files with 632 additions and 32 deletions.
diff --git a/cmake/unittest.cmake b/cmake/unittest.cmake
@@ -19,10 +19,10 @@ set(${UT_NAME}_libs
     ${googletest_STATIC_LIBRARIES}
 )
 
-list(APPEND ${UT_NAME}_libs ${PROTOBUF_LIBRARIES})
-list(APPEND ${UT_NAME}_libs onnx_proto)
 add_whole_archive_flag(onnx tmp)
 list(APPEND ${UT_NAME}_libs ${tmp})
+list(APPEND ${UT_NAME}_libs onnx_proto)
+list(APPEND ${UT_NAME}_libs ${PROTOBUF_LIBRARIES})
 
 file(GLOB_RECURSE ${UT_NAME}_src
     "${ONNX_ROOT}/onnx/test/c++/*.cc"

diff --git a/docs/Functions-ml.md b/docs/Functions-ml.md
@@ -0,0 +1,4 @@
+## Functions
+*This file is automatically generated from the
+            [def files](/onnx/defs) via [this script](/onnx/defs/gen_doc.py).
+            Do not modify directly and instead edit function definitions.*
diff --git a/docs/Functions.md b/docs/Functions.md
@@ -0,0 +1,36 @@
+## Functions
+*This file is automatically generated from the
+            [def files](/onnx/defs) via [this script](/onnx/defs/gen_doc.py).
+            Do not modify directly and instead edit function definitions.*
+## ai.onnx (default)
+  * <sub>experimental</sub><a href="#FuncMeanVarianceNormalization">FuncMeanVarianceNormalization</a>
+
+
+
+### <sub>experimental</sub> <a name="FuncMeanVarianceNormalization"></a><a name="funcmeanvariancenormalization">**FuncMeanVarianceNormalization**</a>
+
+  A MeanVarianceNormalization Function: Perform mean variance normalization on the input tensor X using formula: <br/> ``` (X-EX)/sqrt(E(X-EX)^2) ``` <br/><br/><b>INPUT: </b>X(float/float16/double) with shape [N,C,W,H] or N-D shape <br/><br/><b>ATTRIBUTE: </b><br/>&nbsp;&nbsp;&nbsp;&nbsp;<tt>axes: </tt>will be passed to ReducedMean Ops. Use [0,2,3] (without C axis for N-D cases) for calculating means and variances along channels. Two variables with the same C-coordinate are associated with the same mean and variance. Use [0,1,2,3] (with C axis) to calculate global mean and global variance with all variables sharing the same mean/variance.<br/>&nbsp;&nbsp;&nbsp;&nbsp;(The KeepDims attribute in ReducedMean is set to true for calculation)<br/><br/><b>OUTPUT: </b>X_MVN(float/float16/double) with the same shape as input X<br/>
+
+#### Version
+
+This version of the function has been available since version 8 of the default ONNX operator set.
+
+#### Inputs
+
+<dl>
+<dt>X; </dt>
+<br/></dl>
+
+#### Outputs
+
+<dl>
+<dt>X_MVN; </dt>
+<br/></dl>
+
+#### Attributes
+
+<dl>
+<dt>axes;<br/></dt>
+</dl>
+
+
diff --git a/docs/FunctionsChangelog-ml.md b/docs/FunctionsChangelog-ml.md
@@ -0,0 +1,5 @@
+## Function Changelog
+*This file is automatically generated from the
+            [def files](/onnx/defs) via [this script](/onnx/defs/gen_doc.py).
+            Do not modify directly and instead edit function definitions.*
+## ai.onnx.ml
diff --git a/docs/FunctionsChangelog.md b/docs/FunctionsChangelog.md
@@ -0,0 +1,32 @@
+## Function Changelog
+*This file is automatically generated from the
+            [def files](/onnx/defs) via [this script](/onnx/defs/gen_doc.py).
+            Do not modify directly and instead edit function definitions.*
+# ai.onnx (default)
+## Version 8 of domain ai.onnx (default)
+### <a name="FuncMeanVarianceNormalization-8"></a>**FuncMeanVarianceNormalization-8**</a>
+
+  A MeanVarianceNormalization Function: Perform mean variance normalization on the input tensor X using formula: <br/> ``` (X-EX)/sqrt(E(X-EX)^2) ``` <br/><br/><b>INPUT: </b>X(float/float16/double) with shape [N,C,W,H] or N-D shape <br/><br/><b>ATTRIBUTE: </b><br/>&nbsp;&nbsp;&nbsp;&nbsp;<tt>axes: </tt>will be passed to ReducedMean Ops. Use [0,2,3] (without C axis for N-D cases) for calculating means and variances along channels. Two variables with the same C-coordinate are associated with the same mean and variance. Use [0,1,2,3] (with C axis) to calculate global mean and global variance with all variables sharing the same mean/variance.<br/>&nbsp;&nbsp;&nbsp;&nbsp;(The KeepDims attribute in ReducedMean is set to true for calculation)<br/><br/><b>OUTPUT: </b>X_MVN(float/float16/double) with the same shape as input X<br/>
+
+#### Version
+
+This version of the function has been available since version 8 of the default ONNX operator set.
+
+#### Inputs
+
+<dl>
+<dt>X; </dt>
+<br/></dl>
+
+#### Outputs
+
+<dl>
+<dt>X_MVN; </dt>
+<br/></dl>
+
+#### Attributes
+
+<dl>
+<dt>axes;<br/></dt>
+</dl>
+
diff --git a/onnx/backend/test/case/node/mvn.py b/onnx/backend/test/case/node/mvn.py
@@ -0,0 +1,43 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+import numpy as np  # type: ignore
+
+import onnx
+from ..base import Base
+from . import expect
+
+
+class MVN(Base):
+
+    @staticmethod
+    def export():  # type: () -> None
+        node = onnx.helper.make_node(
+            'FuncMeanVarianceNormalization',
+            inputs=['X'],
+            outputs=['Y'],
+            axes=[0, 2, 3]
+        )
+
+        input_data = np.array([[[[0.8439683], [0.5665144], [0.05836735]],
+            [[0.02916367], [0.12964272], [0.5060197]],
+            [[0.79538304], [0.9411346], [0.9546573]]],
+            [[[0.17730942], [0.46192095], [0.26480448]],
+            [[0.6746842], [0.01665257], [0.62473077]],
+            [[0.9240844], [0.9722341], [0.11965699]]],
+            [[[0.41356155], [0.9129373], [0.59330076]],
+            [[0.81929934], [0.7862604], [0.11799799]],
+            [[0.69248444], [0.54119414], [0.07513223]]]], dtype=np.float32)
+
+        # Calculate expected output data
+        data_mean = np.mean(input_data, axis=(0, 2, 3), keepdims=1)
+        data_mean_squared = np.power(data_mean, 2)
+        data_squared = np.power(input_data, 2)
+        data_squared_mean = np.mean(data_squared, axis=(0, 2, 3), keepdims=1)
+        std = np.sqrt(data_squared_mean - data_mean_squared)
+        expected_output = (input_data - data_mean) / (std + 1e-9)
+
+        expect(node, inputs=[input_data], outputs=[expected_output],
+               name='test_mvn')
diff --git a/onnx/common/model_helpers.cc b/onnx/common/model_helpers.cc
@@ -0,0 +1,39 @@
+// Copyright (c) Facebook Inc. and Microsoft Corporation.
+// Licensed under the MIT license.
+
+#include "onnx/common/model_helpers.h"
+#include "onnx/checker.h"
+#include "onnx/defs/schema.h"
+#include "onnx/string_utils.h"
+
+namespace ONNX_NAMESPACE {
+using namespace Common;
+
+Common::Status BuildNode(
+    const std::string& name,
+    const std::string& domain,
+    const std::string& doc_string,
+    const std::string& op_type,
+    std::vector<std::string> const& inputs,
+    std::vector<std::string> const& outputs,
+    NodeProto* node) {
+  if (node == NULL) {
+    return Status(
+        Common::CHECKER,
+        Common::INVALID_ARGUMENT,
+        "node_proto should not be nullptr.");
+  }
+  node->set_name(name);
+  node->set_domain(domain);
+  node->set_doc_string(doc_string);
+  node->set_op_type(op_type);
+  for (auto& input : inputs) {
+    node->add_input(input);
+  }
+  for (auto& output : outputs) {
+    node->add_output(output);
+  }
+
+  return Status::OK();
+}
+} // namespace ONNX_NAMESPACE
diff --git a/onnx/common/model_helpers.h b/onnx/common/model_helpers.h
@@ -0,0 +1,24 @@
+// Copyright (c) Facebook Inc. and Microsoft Corporation.
+// Licensed under the MIT license.
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include "onnx/common/status.h"
+#include "onnx/onnx-operators_pb.h"
+
+namespace ONNX_NAMESPACE {
+
+// Helper function for register nodes in
+// a FunctionProto. Attributes need to be
+// registered separately.
+Common::Status BuildNode(
+    const std::string& name,
+    const std::string& domain,
+    const std::string& doc_string,
+    const std::string& op_type,
+    std::vector<std::string> const& inputs,
+    std::vector<std::string> const& outputs,
+    /*OUT*/ NodeProto* node);
+} // namespace ONNX_NAMESPACE
diff --git a/onnx/cpp2py_export.cc b/onnx/cpp2py_export.cc
@@ -5,6 +5,7 @@
 #include <unordered_map>
 
 #include "onnx/checker.h"
+#include "onnx/defs/function.h"
 #include "onnx/defs/schema.h"
 #include "onnx/optimizer/optimize.h"
 #include "onnx/py_utils.h"
@@ -146,6 +147,34 @@ PYBIND11_MODULE(onnx_cpp2py_export, onnx_cpp2py_export) {
     return OpSchemaRegistry::get_all_schemas_with_history();
   });
 
+  defs.def(
+      "get_all_functions",
+      [](const std::string& domain)
+          -> std::unordered_map<std::string, std::vector<py::bytes>> {
+        std::multimap<std::string, std::unique_ptr<FunctionProto>> temp_ptr_map;
+        std::unordered_map<std::string, std::vector<py::bytes>> temp_map;
+        FunctionBuilderRegistry& function_registry =
+            FunctionBuilderRegistry::OnnxInstance();
+
+        Common::Status status =
+            function_registry.GetFunctions(domain, &temp_ptr_map);
+        if (!status.IsOK()) {
+          throw std::runtime_error(
+              "Failed to retrieve function list for domain '" + domain + "'!");
+        }
+        for (auto iter = temp_ptr_map.begin(); iter != temp_ptr_map.end();
+             ++iter) {
+          std::string bytes;
+          if (!iter->second->SerializeToString(&bytes)) {
+            throw std::runtime_error(
+                "Failed to serilize registered function for '" + iter->first +
+                "'!");
+          }
+          temp_map[iter->first].emplace_back(py::bytes(std::move(bytes)));
+        }
+        return temp_map;
+      });
+
   // Submodule `checker`
   auto checker = onnx_cpp2py_export.def_submodule("checker");
   checker.doc() = "Checker submodule";

diff --git a/onnx/defs/__init__.py b/onnx/defs/__init__.py
@@ -3,12 +3,14 @@
 from __future__ import print_function
 from __future__ import unicode_literals
 
-import onnx
-from onnx import AttributeProto
+from onnx import AttributeProto, FunctionProto
 import onnx.onnx_cpp2py_export.defs as C
 
+from collections import defaultdict
+from typing import List, Dict
 
 ONNX_DOMAIN = ""
+ONNX_ML_DOMAIN = 'ai.onnx.ml'
 
 
 has = C.has_schema
@@ -32,3 +34,14 @@ def _Attribute_default_value(self):  # type: ignore
 
 
 OpSchema.Attribute.default_value = _Attribute_default_value  # type: ignore
+
+
+def get_functions(domain=ONNX_DOMAIN):  # type: ignore
+    function_map = defaultdict(list)  # type: Dict[int, List[FunctionProto]]
+    function_byte_map = C.get_all_functions(domain)  # type: ignore
+    for function_name, raw_functions in function_byte_map.items():
+        for function_bytes in raw_functions:
+            function_proto = FunctionProto()
+            function_proto.ParseFromString(function_bytes)
+            function_map[function_name].append(function_proto)
+    return function_map