From 9e0c6800c53701fc50dfb69a2c8b6de19c52c559 Mon Sep 17 00:00:00 2001 From: Yu Yang Date: Sat, 15 Jul 2017 20:18:54 +0800 Subject: [PATCH] Python Generate OpCreation Methods by OpProto All OpCreation method are generated by `create_op_creation_methods::__bootstrap__` method, and stores in `op_creations` object and its methods. There are three parts to implement this feature. 1. Get all registered `OpProto` from C++ side. It is implemented in `get_all_op_protos` method. 1. Create a function to convert `kwargs` to `OpDesc` base on each op's `OpProto`. The `OpDescCreationMethod` class. 1. Convert `OpProto` to `docstring` by `get_docstring_from_op_proto` method. All three methods are unit tested. The `__bootstrap__` just combines them together and create a method in runtime. For details, please reference the doc string in `create_op_creation_methods.py` and the unit test `test_op_creation_methods.py`. --- paddle/framework/op_registry.h | 24 ++ paddle/framework/operator.cc | 28 +- paddle/framework/operator.h | 8 +- paddle/pybind/pybind.cc | 17 ++ .../framework/create_op_creation_methods.py | 235 +++++++++++++++++ .../tests/test_op_creation_methods.py | 243 +++++++++++++++++- python/paddle/v2/optimizer.py | 2 + 7 files changed, 539 insertions(+), 18 deletions(-) diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h index de20e7af0..3d67541db 100644 --- a/paddle/framework/op_registry.h +++ b/paddle/framework/op_registry.h @@ -1,6 +1,7 @@ #pragma once #include +#include #include #include #include @@ -199,8 +200,12 @@ class OpRegistry { } static OperatorPtr CreateOp(const OpDesc& op_desc) { + //! Create a OpPtr by type. std::string op_type = op_desc.type(); OperatorPtr op(creators().at(op_type)()); + + //! Fill op's data member. Not use constructor because it will be noising + //! for Op developer. op->desc_ = op_desc; op->inputs_.reserve((size_t)op_desc.inputs_size()); std::copy(op_desc.inputs().begin(), op_desc.inputs().end(), @@ -208,10 +213,18 @@ class OpRegistry { op->outputs_.reserve((size_t)op_desc.outputs_size()); std::copy(op_desc.outputs().begin(), op_desc.outputs().end(), std::back_inserter(op->outputs_)); + + //! Fill attrs, and validate attrs. for (auto& attr : op_desc.attrs()) { op->attrs_[attr.name()] = AttrTypeHelper::GetAttrValue(attr); } op_checkers().at(op_type).Check(op->attrs_); + + //! Convert Temporary variable name to an unique variable name. + AssignTempVariable(op.get()); + + //! Other op's custom Init for a complex Op. For simple Op, the Init + //! method do nothing. op->Init(); return op; } @@ -222,6 +235,17 @@ class OpRegistry { }; private: + static void AssignTempVariable(OperatorBase* op) { + static std::atomic gUniqId(0UL); + for (auto& outname : op->outputs_) { + if (outname == OperatorBase::TMP_VAR_NAME()) { + outname += op->Type(); + outname += "@"; + outname += std::to_string(gUniqId.fetch_add(1)); + } + } + } + static std::unordered_map& creators() { static std::unordered_map creators_; return creators_; diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc index d06567082..a467d328e 100644 --- a/paddle/framework/operator.cc +++ b/paddle/framework/operator.cc @@ -19,23 +19,21 @@ namespace framework { std::string OperatorBase::DebugString() const { std::stringstream ss; - ss << "=================\n"; - ss << "type = " << desc_.type() << "\n"; - ss << "inputs = ["; - for (auto& ipt : inputs_) { - ss << ipt << ", "; + ss << "Op(" << Type() << "), inputs:("; + for (size_t i = 0; i < inputs_.size(); ++i) { + ss << inputs_[i]; + if (i != inputs_.size() - 1) { + ss << ", "; + } } - ss << "]\n"; - ss << "outputs = ["; - for (auto& opt : outputs_) { - ss << opt << ", "; + ss << "), outputs:("; + for (size_t i = 0; i < outputs_.size(); ++i) { + ss << outputs_[i]; + if (i != outputs_.size() - 1) { + ss << ", "; + } } - ss << "]\n"; - ss << "attr_keys = ["; - for (auto& attr : attrs_) { - ss << attr.first << ", "; - } - ss << "]\n"; + ss << ")."; return ss.str(); } diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index cf79f379f..cc166048b 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -39,6 +39,13 @@ using OperatorPtr = std::shared_ptr; */ class OperatorBase { public: + /// If a variable is a empty variable, that name will be used. + static std::string EMPTY_VAR_NAME() { return "@EMPTY@"; } + + /// If a variable is a temporary variable, that name will be set in Python, + /// but it will be convert to a unique name in scope after OpCreator. + static std::string TMP_VAR_NAME() { return "@TEMP@"; } + virtual ~OperatorBase() {} template @@ -62,7 +69,6 @@ class OperatorBase { virtual void Run(const ScopePtr& scope, const platform::DeviceContext& dev_ctx) const = 0; - protected: std::string Type() const { return desc_.type(); } public: diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc index c1a025ed0..b5ead21fd 100644 --- a/paddle/pybind/pybind.cc +++ b/paddle/pybind/pybind.cc @@ -63,6 +63,23 @@ All parameter, weight, gradient are variables in Paddle. } return ret_values; }); + m.def_submodule( + "var_names", + "The module will return special predefined variable name in Paddle") + .def("empty", pd::OperatorBase::EMPTY_VAR_NAME) + .def("temp", pd::OperatorBase::TMP_VAR_NAME); + + py::class_(m, "Operator") + .def("__str__", &pd::OperatorBase::DebugString) + .def_static("create", [](const std::string& protobin) { + pd::OpDesc desc; + PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), + "Cannot parse user input to OpDesc"); + PADDLE_ENFORCE(desc.IsInitialized(), + "User OpDesc is not initialized, reason %s", + desc.InitializationErrorString()); + return pd::OpRegistry::CreateOp(desc); + }); return m.ptr(); } diff --git a/python/paddle/v2/framework/create_op_creation_methods.py b/python/paddle/v2/framework/create_op_creation_methods.py index 2fcdfead2..c2a7ae769 100644 --- a/python/paddle/v2/framework/create_op_creation_methods.py +++ b/python/paddle/v2/framework/create_op_creation_methods.py @@ -1,11 +1,246 @@ import paddle.v2.framework.core as core import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2 +import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2 +import paddle.v2.framework.proto.attr_type_pb2 as attr_type_pb2 +import cStringIO def get_all_op_protos(): + """ + Get all registered op proto from Paddle C++ + :return: list of OpProto + """ protostrs = core.get_all_op_protos() ret_values = [] for pbstr in protostrs: op_proto = op_proto_pb2.OpProto.FromString(str(pbstr)) ret_values.append(op_proto) return ret_values + + +class OpDescCreationMethod(object): + """ + A Functor object to convert user input(use key word args) to OpDesc based on + OpProto. + + :param op_proto: The OpProto object. + :type op_proto: op_proto_pb2.OpProto + """ + + def __init__(self, op_proto): + if not isinstance(op_proto, op_proto_pb2.OpProto): + raise TypeError("Argument should be OpProto") + self.__op_proto__ = op_proto + + def __call__(self, *args, **kwargs): + """ + Convert user input to OpDesc. Only key-word args are supported. + :return: OpDesc based on user input + :rtype: op_desc_pb2.OpDesc + """ + if len(args) != 0: + raise ValueError("Only keyword arguments is supported by Paddle") + op_desc = op_desc_pb2.OpDesc() + + # Inputs + ipts, ipt_format, _ = OpDescCreationMethod.extract_input_or_output( + "input", kwargs, self.__op_proto__.inputs) + op_desc.inputs.extend(ipts) + if ipt_format is not None: + op_desc.attrs.extend([ipt_format]) + + # Outputs + outs, out_format, tmp_index = OpDescCreationMethod.extract_input_or_output( + "output", kwargs, self.__op_proto__.outputs) + op_desc.outputs.extend(outs) + if out_format is not None: + op_desc.attrs.extend([out_format]) + if len(tmp_index) != 0: + tmp_index_attr = op_desc.attrs.add() + tmp_index_attr.type = attr_type_pb2.INTS + tmp_index_attr.name = "temporary_index" + tmp_index_attr.ints.extend(tmp_index) + + # Types + op_desc.type = self.__op_proto__.type + + # Attrs + for attr in self.__op_proto__.attrs: + if attr.generated: + continue + user_defined_attr = kwargs.get(attr.name, None) + if user_defined_attr is not None: + new_attr = op_desc.attrs.add() + new_attr.name = attr.name + new_attr.type = attr.type + if attr.type == attr_type_pb2.INT: + new_attr.i = user_defined_attr + elif attr.type == attr_type_pb2.FLOAT: + new_attr.f = user_defined_attr + elif attr.type == attr_type_pb2.STRING: + new_attr.s = user_defined_attr + elif attr.type == attr_type_pb2.INTS: + new_attr.ints.extend(user_defined_attr) + elif attr.type == attr_type_pb2.FLOATS: + new_attr.floats.extend(user_defined_attr) + elif attr.type == attr_type_pb2.STRINGS: + new_attr.strings.extend(user_defined_attr) + else: + raise NotImplementedError("Not support attribute type " + + attr.type) + + return op_desc + + @staticmethod + def extract_input_or_output(in_out, kwargs, meta): + """ + Extract input variable names or output variable names from key-word + arguments, which base on VarProtos. + + :param in_out: "input" or "output" + :param kwargs: key-word arguments that user inputted. + :param meta: a list of VarProto + :return: The three object will be return. The variable names. The + input_format or output_format attribute(None if the input or output is + not multiple). The temporary variable index list. + """ + multiple = OpDescCreationMethod.any_is_true((m.multiple for m in meta)) + tmp_index = [] + retv = [] + if multiple: + var_format = op_desc_pb2.AttrDesc() + var_format.type = attr_type_pb2.INTS + var_format.name = "%s_format" % in_out + var_format.ints.append(0) + + for var in meta: + var_name = var.name + + if var.temporary: + var_name = [core.var_names.temp()] + tmp_index.append(len(retv)) + else: + var_name = kwargs.get(var_name, []) + if not isinstance(var_name, list): + var_name = [var_name] + retv.extend(var_name) + var_format.ints.append(len(var_name) + var_format.ints[-1]) + return retv, var_format, tmp_index + else: + for var in meta: + if var.temporary: + retv.append(kwargs.get(var.name, core.var_names.temp())) + tmp_index.append(len(retv)) + else: + retv.append(kwargs.get(var.name, core.var_names.empty())) + return retv, None, tmp_index + + @staticmethod + def any_is_true(generator): + """ + Reduce a bool array to one. If any of them is True, then return True. + """ + for flag in generator: + if flag: + return True + return False + + +def get_docstring_from_op_proto(op_proto): + """ + Generate docstring from a OpProto + :param op_proto: a OpProto instance. + :type op_proto: op_proto_pb2.OpProto + :return: docstring + """ + if not isinstance(op_proto, op_proto_pb2.OpProto): + raise TypeError("Input must be OpProto") + f = cStringIO.StringIO() + f.write(op_proto.comment) + f.write("\n") + + def __append_param__(name, comment, type): + # Maybe replace the following line with template engine is better. + f.write(":param ") + f.write(name) + f.write(": ") + f.write(comment) + f.write("\n") + f.write(":type ") + f.write(name) + f.write(": ") + f.write(type) + f.write("\n") + + for ipt in op_proto.inputs: + __append_param__(ipt.name, ipt.comment, "list | basestr" + if ipt.multiple else "basestr") + + temp_var_prefix = \ + "This is a temporary variable. It does not have to set by user. " + for opt in op_proto.outputs: + __append_param__(opt.name, opt.comment if not opt.temporary else + temp_var_prefix + opt.comment, "list | basestr" + if opt.multiple else "basestr") + + for attr in op_proto.attrs: + attr_type = None + if attr.type == attr_type_pb2.INT: + attr_type = "int" + elif attr.type == attr_type_pb2.FLOAT: + attr_type = "float" + elif attr.type == attr_type_pb2.STRING: + attr_type = "basestr" + elif attr.type == attr_type_pb2.INTS: + attr_type = "list of int" + elif attr.type == attr_type_pb2.FLOATS: + attr_type = "list of float" + elif attr.type == attr_type_pb2.STRINGS: + attr_type = "list of basestr" + + if attr_type is None: + raise RuntimeError("Not supported attribute type " + attr.type) + + __append_param__(attr.name, attr.comment, attr_type) + + return f.getvalue() + + +def create_op_creation_method(op_proto): + """ + Generate op creation method for an OpProto + """ + method = OpDescCreationMethod(op_proto) + + def __impl__(*args, **kwargs): + opdesc = method(*args, **kwargs) + return core.Operator.create(opdesc.SerializeToString()) + + __impl__.__doc__ = get_docstring_from_op_proto(op_proto) + return __impl__ + + +class OpCreationsHolder(object): + """ + A object will holds all op creation methods. + + Use `op_creations.xxx_op` to access them. + """ + pass + + +op_creations = OpCreationsHolder() + + +def __bootstrap__(): + """ + Bootstrap function for this module. It will dynamic create all op creation + methods in runtime. + """ + for op_proto in get_all_op_protos(): + func = create_op_creation_method(op_proto) + func.__name__ = str(op_proto.type) + setattr(op_creations, func.__name__, func) + + +__bootstrap__() diff --git a/python/paddle/v2/framework/tests/test_op_creation_methods.py b/python/paddle/v2/framework/tests/test_op_creation_methods.py index b205e2cab..41db7c0d5 100644 --- a/python/paddle/v2/framework/tests/test_op_creation_methods.py +++ b/python/paddle/v2/framework/tests/test_op_creation_methods.py @@ -1,9 +1,13 @@ import unittest import paddle.v2.framework.create_op_creation_methods as creation +import paddle.v2.framework.core as core +import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2 +import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2 +import paddle.v2.framework.proto.attr_type_pb2 as attr_type_pb2 -class TestOpCreationsMethods(unittest.TestCase): - def test_all_protos(self): +class TestGetAllProtos(unittest.TestCase): + def test_all(self): all_protos = creation.get_all_op_protos() self.assertNotEqual(0, len(all_protos)) @@ -11,5 +15,240 @@ class TestOpCreationsMethods(unittest.TestCase): self.assertTrue(each.IsInitialized()) +class TestOpDescCreationMethod(unittest.TestCase): + def test_plain_input_output(self): + op = op_proto_pb2.OpProto() + op.type = "test" + ipt = op.inputs.add() + ipt.name = "X" + ipt.comment = "not matter" + + ipt = op.inputs.add() + ipt.name = "Y" + ipt.comment = "not matter" + + opt = op.outputs.add() + opt.name = "Z" + opt.comment = "not matter" + + op.comment = "not matter" + + self.assertTrue(op.IsInitialized()) + + method = creation.OpDescCreationMethod(op) + output = method(X="a", Y="b", Z="c") + + expected = op_desc_pb2.OpDesc() + expected.type = "test" + expected.inputs.extend(["a", "b"]) + expected.outputs.append("c") + self.assertEqual(expected, output) + + def test_multiple_input_plain_output(self): + op = op_proto_pb2.OpProto() + op.type = "fc" + ipt = op.inputs.add() + ipt.name = "X" + ipt.comment = "" + ipt.multiple = True + + ipt = op.inputs.add() + ipt.name = "W" + ipt.comment = "" + ipt.multiple = True + + ipt = op.inputs.add() + ipt.name = "b" + ipt.comment = "" + + out = op.outputs.add() + out.name = "Y" + out.comment = "" + + op.comment = "" + self.assertTrue(op.IsInitialized()) + method = creation.OpDescCreationMethod(op) + + generated1 = method(X="x", W="w", b="b", Y="y") + expected1 = op_desc_pb2.OpDesc() + expected1.inputs.extend(['x', 'w', 'b']) + expected1.outputs.extend(['y']) + expected1.type = 'fc' + attr = expected1.attrs.add() + attr.name = 'input_format' + attr.type = attr_type_pb2.INTS + attr.ints.extend([0, 1, 2, 3]) + self.assertEqual(expected1, generated1) + + generated2 = method( + X=['x1', 'x2', 'x3'], b='b', W=['w1', 'w2', 'w3'], Y='y') + expected2 = op_desc_pb2.OpDesc() + expected2.inputs.extend(['x1', 'x2', 'x3', 'w1', 'w2', 'w3', 'b']) + expected2.outputs.extend(['y']) + expected2.type = 'fc' + attr = expected2.attrs.add() + attr.name = 'input_format' + attr.type = attr_type_pb2.INTS + attr.ints.extend([0, 3, 6, 7]) + self.assertEqual(expected2, generated2) + + def test_attrs(self): + op = op_proto_pb2.OpProto() + op.type = "test" + ipt = op.inputs.add() + ipt.name = 'X' + ipt.comment = "" + + def __add_attr__(name, type): + attr = op.attrs.add() + attr.name = name + attr.comment = "" + attr.type = type + + __add_attr__("int_attr", attr_type_pb2.INT) + __add_attr__("float_attr", attr_type_pb2.FLOAT) + __add_attr__("string_attr", attr_type_pb2.STRING) + __add_attr__("ints_attr", attr_type_pb2.INTS) + __add_attr__("floats_attr", attr_type_pb2.FLOATS) + __add_attr__("strings_attr", attr_type_pb2.STRINGS) + + op.comment = "" + self.assertTrue(op.IsInitialized()) + + method = creation.OpDescCreationMethod(op) + + generated = method( + X="a", + int_attr=10, + float_attr=3.2, + string_attr="test_str", + ints_attr=[0, 1, 2, 3, 4], + floats_attr=[0.2, 3.2, 4.5], + strings_attr=["a", "b", "c"]) + + expected = op_desc_pb2.OpDesc() + expected.type = "test" + expected.inputs.extend(['a']) + attr = expected.attrs.add() + attr.name = "int_attr" + attr.type = attr_type_pb2.INT + attr.i = 10 + + attr = expected.attrs.add() + attr.name = "float_attr" + attr.type = attr_type_pb2.FLOAT + attr.f = 3.2 + + attr = expected.attrs.add() + attr.name = "string_attr" + attr.type = attr_type_pb2.STRING + attr.s = "test_str" + + attr = expected.attrs.add() + attr.name = "ints_attr" + attr.type = attr_type_pb2.INTS + attr.ints.extend([0, 1, 2, 3, 4]) + + attr = expected.attrs.add() + attr.name = 'floats_attr' + attr.type = attr_type_pb2.FLOATS + attr.floats.extend([0.2, 3.2, 4.5]) + + attr = expected.attrs.add() + attr.name = 'strings_attr' + attr.type = attr_type_pb2.STRINGS + attr.strings.extend(['a', 'b', 'c']) + + self.assertEqual(expected, generated) + + def test_input_temporary_output(self): + op = op_proto_pb2.OpProto() + op.type = "test" + out = op.outputs.add() + out.name = "OUT" + out.comment = "" + + out = op.outputs.add() + out.name = "TMP" + out.comment = "" + out.temporary = True + + out = op.outputs.add() + out.name = "OUT2" + out.comment = "" + op.comment = "" + + method = creation.OpDescCreationMethod(op) + generated = method(OUT="a", OUT2="b") + desc = op_desc_pb2.OpDesc() + desc.outputs.extend(["a", core.var_names.temp(), "b"]) + desc.type = "test" + attr = desc.attrs.add() + attr.name = "temporary_index" + attr.type = attr_type_pb2.INTS + attr.ints.append(2) + self.assertEqual(generated, desc) + + +class TestOpCreationDocStr(unittest.TestCase): + def test_all(self): + op = op_proto_pb2.OpProto() + op.type = "test" + op.comment = """Test Op. + +This op is used for unit test, not a real op. +""" + a = op.inputs.add() + a.name = "a" + a.comment = "Input a for test op" + a.multiple = True + + b = op.inputs.add() + b.name = "b" + b.comment = "Input b for test op" + self.assertTrue(op.IsInitialized()) + + o1 = op.outputs.add() + o1.name = "output" + o1.comment = "The output of test op" + + o2 = op.outputs.add() + o2.name = "temp output" + o2.comment = "The temporary output of test op" + o2.temporary = True + + test_str = op.attrs.add() + test_str.name = "str_attr" + test_str.type = attr_type_pb2.STRING + test_str.comment = "A string attribute for test op" + + actual = creation.get_docstring_from_op_proto(op) + expected_docstring = '''Test Op. + +This op is used for unit test, not a real op. + +:param a: Input a for test op +:type a: list | basestr +:param b: Input b for test op +:type b: basestr +:param output: The output of test op +:type output: basestr +:param temp output: This is a temporary variable. It does not have to set by user. The temporary output of test op +:type temp output: basestr +:param str_attr: A string attribute for test op +:type str_attr: basestr +''' + self.assertEqual(expected_docstring, actual) + + +class TestOpCreations(unittest.TestCase): + def test_all(self): + add_op = creation.op_creations.add_two(X="a", Y="b", Out="z") + self.assertIsNotNone(add_op) + # Invoke C++ DebugString() + self.assertEqual('Op(add_two), inputs:(a, b), outputs:(z).', + str(add_op)) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/optimizer.py b/python/paddle/v2/optimizer.py index b6ee51cfe..173a30a41 100644 --- a/python/paddle/v2/optimizer.py +++ b/python/paddle/v2/optimizer.py @@ -25,6 +25,8 @@ class Optimizer(object): self.__opt_conf_proto__ = config_parser_utils.parse_optimizer_config( __impl__) + if swig_api is None: + raise RuntimeError("paddle.v2 currently need swig_paddle") self.__opt_conf__ = swig_api.OptimizationConfig.createFromProto( self.__opt_conf_proto__) -- GitLab