tensorflow--tensorflow/tensorflow/python/grappler/tf_optimizer_wrapper.cc

/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.

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 <stdexcept>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>

#include "pybind11/pybind11.h"  // from @pybind11
#include "pybind11_protobuf/native_proto_caster.h"  // from @pybind11_protobuf
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/common_runtime/device_factory.h"
#include "tensorflow/core/framework/device_attributes.pb.h"
#include "tensorflow/core/framework/device_base.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/framework/graph_def_util.h"
#include "tensorflow/core/grappler/clusters/cluster.h"
#include "tensorflow/core/grappler/clusters/utils.h"
#include "tensorflow/core/grappler/grappler_item.h"
#include "tensorflow/core/grappler/grappler_item_builder.h"
#include "tensorflow/core/grappler/optimizers/meta_optimizer.h"
#include "tensorflow/core/protobuf/config.pb.h"
#include "tensorflow/core/protobuf/device_properties.pb.h"
#include "tensorflow/core/protobuf/meta_graph.pb.h"
#include "tensorflow/core/public/session_options.h"
#include "tensorflow/python/lib/core/pybind11_status.h"

namespace py = pybind11;

void DetectDevices(
    std::unordered_map<std::string, tensorflow::DeviceProperties>* device_map) {
  tensorflow::SessionOptions options;
  std::vector<std::unique_ptr<tensorflow::Device>> devices;
  if (!tensorflow::DeviceFactory::AddDevices(options, "", &devices).ok()) {
    return;
  }

  for (const std::unique_ptr<tensorflow::Device>& device : devices) {
    tensorflow::DeviceProperties& prop = (*device_map)[device->name()];
    prop = tensorflow::grappler::GetDeviceInfo(device->parsed_name());

    // Overwrite the memory limit since users might have requested to use only a
    // fraction of the available device memory.
    const tensorflow::DeviceAttributes& attr = device->attributes();
    prop.set_memory_size(attr.memory_limit());
  }
}

namespace {
tensorflow::GraphDef TF_OptimizeGraph(
    tensorflow::grappler::Cluster* cluster,
    const std::string& serialized_config_proto,
    const tensorflow::MetaGraphDef& metagraph, bool verbose,
    const std::string& graph_id, bool strip_default_attributes) {
  tensorflow::GraphDef out_graph;
  {
    py::gil_scoped_release gil_release;
    tensorflow::ConfigProto config_proto;
    if (!config_proto.ParseFromString(serialized_config_proto)) {
      throw std::invalid_argument(
          "The ConfigProto could not be parsed as a valid protocol "
          "buffer");
    }

    tensorflow::grappler::ItemConfig item_config;
    // This disables graph optimizations in the older graph optimizer,
    // which tend to overlap / be redundant with those in Grappler.
    item_config.apply_optimizations = false;
    item_config.ignore_user_placement = false;
    std::unique_ptr<tensorflow::grappler::GrapplerItem> grappler_item =
        tensorflow::grappler::GrapplerItemFromMetaGraphDef(graph_id, metagraph,
                                                           item_config);
    if (!grappler_item) {
      throw std::invalid_argument(
          "Failed to import metagraph, check error log for more info.");
    }

    tensorflow::DeviceBase* cpu_device = nullptr;
    tensorflow::grappler::MetaOptimizer optimizer(cpu_device, config_proto);

    tsl::MaybeRaiseRegisteredFromStatusWithGIL(
        optimizer.Optimize(cluster, *grappler_item, &out_graph));
    if (strip_default_attributes) {
      tensorflow::StripDefaultAttributes(*tensorflow::OpRegistry::Global(),
                                         out_graph.mutable_node());
    }
    if (verbose) {
      optimizer.PrintResult();
    }
  }
  return out_graph;
}
}  // namespace

// Here two bindings of `TF_OptimizeGraph` are introduced, one serializes the
// `MetaGraphDef` to string when passing it to C++, another uses pybind's
// protobuf interface, and enables its native proto caster feature
// (https://github.com/pybind/pybind11_protobuf#c-native-vs-python-native-types)
// to prevent serialization. This is necessary for grappler to support models
// larger than 2GiB.
// At the moment, the open source python API defaults to the serialized
// implementation.
PYBIND11_MODULE(_pywrap_tf_optimizer, m) {
  pybind11_protobuf::ImportNativeProtoCasters();
  m.def("TF_OptimizeGraphSerialized",
        [](tensorflow::grappler::Cluster* cluster,
           const std::string& serialized_config_proto,
           const std::string& serialized_metagraph, bool verbose,
           const std::string& graph_id,
           bool strip_default_attributes) -> py::bytes {
          std::string out_graph_bytes;

          {
            tensorflow::MetaGraphDef metagraph;
            if (!metagraph.ParseFromString(serialized_metagraph)) {
              throw std::invalid_argument(
                  "The MetaGraphDef could not be parsed as a valid protocol "
                  "buffer");
            }
            const tensorflow::GraphDef out_graph =
                TF_OptimizeGraph(cluster, serialized_config_proto, metagraph,
                                 verbose, graph_id, strip_default_attributes);
            out_graph_bytes = out_graph.SerializeAsString();
          }
          return py::bytes(std::move(out_graph_bytes));
        });

  m.def("TF_OptimizeGraph", &TF_OptimizeGraph);
}