TensorFlow实现自定义Op

『写在前面』

以CTC Beam search decoder为例，简单整理一下TensorFlow实现自定义Op的操作流程。

基本的流程

1. 定义Op接口

#include "tensorflow/core/framework/op.h"

REGISTER_OP("Custom")    
    .Input("custom_input: int32")
    .Output("custom_output: int32");

2. 为Op实现Compute操作(CPU)或实现kernel(GPU)

#include "tensorflow/core/framework/op_kernel.h"

using namespace tensorflow;

class CustomOp : public OpKernel{
    public:
    explicit CustomOp(OpKernelConstruction* context) : OpKernel(context) {}
    void Compute(OpKernelContext* context) override {
    // 获取输入 tensor.
    const Tensor& input_tensor = context->input(0);
    auto input = input_tensor.flat<int32>();
    // 创建一个输出 tensor.
    Tensor* output_tensor = NULL;
    OP_REQUIRES_OK(context, context->allocate_output(0, input_tensor.shape(),
                                                     &output_tensor));
    auto output = output_tensor->template flat<int32>();
    //进行具体的运算，操作input和output
    //……
  }
};

3. 将实现的kernel注册到TensorFlow系统中

REGISTER_KERNEL_BUILDER(Name("Custom").Device(DEVICE_CPU), CustomOp);

 

CTCBeamSearchDecoder自定义

 

该Op对应TensorFlow中的源码部分 

• Op接口的定义：

tensorflow-master/tensorflow/core/ops/ctc_ops.cc

• CTCBeamSearchDecoder本身的定义：

tensorflow-master/tensorflow/core/util/ctc/ctc_beam_search.cc

• Op-Class的封装与Op注册：

tensorflow-master/tensorflow/core/kernels/ctc_decoder_ops.cc

 

基于源码修改的Op

#include <algorithm>
#include <vector>
#include <cmath>

#include "tensorflow/core/util/ctc/ctc_beam_search.h"

#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/shape_inference.h"
#include "tensorflow/core/kernels/bounds_check.h"

namespace tf = tensorflow;
using tf::shape_inference::DimensionHandle;
using tf::shape_inference::InferenceContext;
using tf::shape_inference::ShapeHandle;

using namespace tensorflow;

REGISTER_OP("CTCBeamSearchDecoderWithParam")
    .Input("inputs: float")
    .Input("sequence_length: int32")
    .Attr("beam_width: int >= 1")
    .Attr("top_paths: int >= 1")
    .Attr("merge_repeated: bool = true")
    //新添加了两个参数
    .Attr("label_selection_size: int >= 0 = 0") 
    .Attr("label_selection_margin: float") 
    .Output("decoded_indices: top_paths * int64")
    .Output("decoded_values: top_paths * int64")
    .Output("decoded_shape: top_paths * int64")
    .Output("log_probability: float")
    .SetShapeFn([](InferenceContext* c) {
      ShapeHandle inputs;
      ShapeHandle sequence_length;

      TF_RETURN_IF_ERROR(c->WithRank(c->input(0), 3, &inputs));
      TF_RETURN_IF_ERROR(c->WithRank(c->input(1), 1, &sequence_length));

      // Get batch size from inputs and sequence_length.
      DimensionHandle batch_size;
      TF_RETURN_IF_ERROR(
          c->Merge(c->Dim(inputs, 1), c->Dim(sequence_length, 0), &batch_size));

      int32 top_paths;
      TF_RETURN_IF_ERROR(c->GetAttr("top_paths", &top_paths));

      // Outputs.
      int out_idx = 0;
      for (int i = 0; i < top_paths; ++i) {  // decoded_indices
        c->set_output(out_idx++, c->Matrix(InferenceContext::kUnknownDim, 2));
      }
      for (int i = 0; i < top_paths; ++i) {  // decoded_values
        c->set_output(out_idx++, c->Vector(InferenceContext::kUnknownDim));
      }
      ShapeHandle shape_v = c->Vector(2);
      for (int i = 0; i < top_paths; ++i) {  // decoded_shape
        c->set_output(out_idx++, shape_v);
      }
      c->set_output(out_idx++, c->Matrix(batch_size, top_paths));
      return Status::OK();
    });

typedef Eigen::ThreadPoolDevice CPUDevice;

inline float RowMax(const TTypes<float>::UnalignedConstMatrix& m, int r,
                    int* c) {
  *c = 0;
  CHECK_LT(0, m.dimension(1));
  float p = m(r, 0);
  for (int i = 1; i < m.dimension(1); ++i) {
    if (m(r, i) > p) {
      p = m(r, i);
      *c = i;
    }
  }
  return p;
}

class CTCDecodeHelper {
 public:
  CTCDecodeHelper() : top_paths_(1) {}

  inline int GetTopPaths() const { return top_paths_; }
  void SetTopPaths(int tp) { top_paths_ = tp; }

  Status ValidateInputsGenerateOutputs(
      OpKernelContext* ctx, const Tensor** inputs, const Tensor** seq_len,
      Tensor** log_prob, OpOutputList* decoded_indices,
      OpOutputList* decoded_values, OpOutputList* decoded_shape) const {
    Status status = ctx->input("inputs", inputs);
    if (!status.ok()) return status;
    status = ctx->input("sequence_length", seq_len);
    if (!status.ok()) return status;

    const TensorShape& inputs_shape = (*inputs)->shape();

    if (inputs_shape.dims() != 3) {
      return errors::InvalidArgument("inputs is not a 3-Tensor");
    }

    const int64 max_time = inputs_shape.dim_size(0);
    const int64 batch_size = inputs_shape.dim_size(1);

    if (max_time == 0) {
      return errors::InvalidArgument("max_time is 0");
    }
    if (!TensorShapeUtils::IsVector((*seq_len)->shape())) {
      return errors::InvalidArgument("sequence_length is not a vector");
    }

    if (!(batch_size == (*seq_len)->dim_size(0))) {
      return errors::FailedPrecondition(
          "len(sequence_length) != batch_size.  ", "len(sequence_length):  ",
          (*seq_len)->dim_size(0), " batch_size: ", batch_size);
    }

    auto seq_len_t = (*seq_len)->vec<int32>();

    for (int b = 0; b < batch_size; ++b) {
      if (!(seq_len_t(b) <= max_time)) {
        return errors::FailedPrecondition("sequence_length(", b, ") <= ",
                                          max_time);
      }
    }

    Status s = ctx->allocate_output(
        "log_probability", TensorShape({batch_size, top_paths_}), log_prob);
    if (!s.ok()) return s;

    s = ctx->output_list("decoded_indices", decoded_indices);
    if (!s.ok()) return s;
    s = ctx->output_list("decoded_values", decoded_values);
    if (!s.ok()) return s;
    s = ctx->output_list("decoded_shape", decoded_shape);
    if (!s.ok()) return s;

    return Status::OK();
  }

  // sequences[b][p][ix] stores decoded value "ix" of path "p" for batch "b".
  Status StoreAllDecodedSequences(
      const std::vector<std::vector<std::vector<int> > >& sequences,
      OpOutputList* decoded_indices, OpOutputList* decoded_values,
      OpOutputList* decoded_shape) const {
    // Calculate the total number of entries for each path
    const int64 batch_size = sequences.size();
    std::vector<int64> num_entries(top_paths_, 0);

    // Calculate num_entries per path
    for (const auto& batch_s : sequences) {
      CHECK_EQ(batch_s.size(), top_paths_);
      for (int p = 0; p < top_paths_; ++p) {
        num_entries[p] += batch_s[p].size();
      }
    }

    for (int p = 0; p < top_paths_; ++p) {
      Tensor* p_indices = nullptr;
      Tensor* p_values = nullptr;
      Tensor* p_shape = nullptr;

      const int64 p_num = num_entries[p];

      Status s =
          decoded_indices->allocate(p, TensorShape({p_num, 2}), &p_indices);
      if (!s.ok()) return s;
      s = decoded_values->allocate(p, TensorShape({p_num}), &p_values);
      if (!s.ok()) return s;
      s = decoded_shape->allocate(p, TensorShape({2}), &p_shape);
      if (!s.ok()) return s;

      auto indices_t = p_indices->matrix<int64>();
      auto values_t = p_values->vec<int64>();
      auto shape_t = p_shape->vec<int64>();

      int64 max_decoded = 0;
      int64 offset = 0;

      for (int64 b = 0; b < batch_size; ++b) {
        auto& p_batch = sequences[b][p];
        int64 num_decoded = p_batch.size();
        max_decoded = std::max(max_decoded, num_decoded);
        std::copy_n(p_batch.begin(), num_decoded, &values_t(offset));
        for (int64 t = 0; t < num_decoded; ++t, ++offset) {
          indices_t(offset, 0) = b;
          indices_t(offset, 1) = t;
        }
      }

      shape_t(0) = batch_size;
      shape_t(1) = max_decoded;
    }
    return Status::OK();
  }

 private:
  int top_paths_;
  TF_DISALLOW_COPY_AND_ASSIGN(CTCDecodeHelper);
};

// CTC beam search
class CTCBeamSearchDecoderWithParamOp : public OpKernel {
 public:
  explicit CTCBeamSearchDecoderWithParamOp(OpKernelConstruction* ctx) : OpKernel(ctx) {
    OP_REQUIRES_OK(ctx, ctx->GetAttr("merge_repeated", &merge_repeated_));
    OP_REQUIRES_OK(ctx, ctx->GetAttr("beam_width", &beam_width_));
    //从参数列表中读取新添的两个参数
    OP_REQUIRES_OK(ctx, ctx->GetAttr("label_selection_size", &label_selection_size));
    OP_REQUIRES_OK(ctx, ctx->GetAttr("label_selection_margin", &label_selection_margin));
    int top_paths;
    OP_REQUIRES_OK(ctx, ctx->GetAttr("top_paths", &top_paths));
    decode_helper_.SetTopPaths(top_paths);
  }

  void Compute(OpKernelContext* ctx) override {
    const Tensor* inputs;
    const Tensor* seq_len;
    Tensor* log_prob = nullptr;
    OpOutputList decoded_indices;
    OpOutputList decoded_values;
    OpOutputList decoded_shape;
    OP_REQUIRES_OK(ctx, decode_helper_.ValidateInputsGenerateOutputs(
                            ctx, &inputs, &seq_len, &log_prob, &decoded_indices,
                            &decoded_values, &decoded_shape));

    auto inputs_t = inputs->tensor<float, 3>();
    auto seq_len_t = seq_len->vec<int32>();
    auto log_prob_t = log_prob->matrix<float>();

    const TensorShape& inputs_shape = inputs->shape();

    const int64 max_time = inputs_shape.dim_size(0);
    const int64 batch_size = inputs_shape.dim_size(1);
    const int64 num_classes_raw = inputs_shape.dim_size(2);
    OP_REQUIRES(
        ctx, FastBoundsCheck(num_classes_raw, std::numeric_limits<int>::max()),
        errors::InvalidArgument("num_classes cannot exceed max int"));
    const int num_classes = static_cast<const int>(num_classes_raw);

    log_prob_t.setZero();

    std::vector<TTypes<float>::UnalignedConstMatrix> input_list_t;

    for (std::size_t t = 0; t < max_time; ++t) {
      input_list_t.emplace_back(inputs_t.data() + t * batch_size * num_classes,
                                batch_size, num_classes);
    }

    ctc::CTCBeamSearchDecoder<> beam_search(num_classes, beam_width_,
                                            &beam_scorer_, 1 /* batch_size */,
                                            merge_repeated_);
    //使用传入的两个参数进行Set
    beam_search.SetLabelSelectionParameters(label_selection_size, label_selection_margin);
    Tensor input_chip(DT_FLOAT, TensorShape({num_classes}));
    auto input_chip_t = input_chip.flat<float>();

    std::vector<std::vector<std::vector<int> > > best_paths(batch_size);
    std::vector<float> log_probs;

    // Assumption: the blank index is num_classes - 1
    for (int b = 0; b < batch_size; ++b) {
      auto& best_paths_b = best_paths[b];
      best_paths_b.resize(decode_helper_.GetTopPaths());
      for (int t = 0; t < seq_len_t(b); ++t) {
        input_chip_t = input_list_t[t].chip(b, 0);
        auto input_bi =
            Eigen::Map<const Eigen::ArrayXf>(input_chip_t.data(), num_classes);
        beam_search.Step(input_bi);
      }
      OP_REQUIRES_OK(
          ctx, beam_search.TopPaths(decode_helper_.GetTopPaths(), &best_paths_b,
                                    &log_probs, merge_repeated_));

      beam_search.Reset();

      for (int bp = 0; bp < decode_helper_.GetTopPaths(); ++bp) {
        log_prob_t(b, bp) = log_probs[bp];
      }
    }

    OP_REQUIRES_OK(ctx, decode_helper_.StoreAllDecodedSequences(
                            best_paths, &decoded_indices, &decoded_values,
                            &decoded_shape));
  }

 private:
  CTCDecodeHelper decode_helper_;
  ctc::CTCBeamSearchDecoder<>::DefaultBeamScorer beam_scorer_;
  bool merge_repeated_;
  int beam_width_;
  //新添两个数据成员，用于存储新加的参数
  int label_selection_size;
  float label_selection_margin;
  TF_DISALLOW_COPY_AND_ASSIGN(CTCBeamSearchDecoderWithParamOp);
};

REGISTER_KERNEL_BUILDER(Name("CTCBeamSearchDecoderWithParam").Device(DEVICE_CPU),
                        CTCBeamSearchDecoderWithParamOp);

将自定义的Op编译成.so文件

• 在tensorflow-master目录下新建一个文件夹custom_op
• cd custom_op
• 新建一个BUILD文件，并在其中添加如下代码：

cc_library(
    name = "ctc_decoder_with_param",
    srcs = [
            "new_beamsearch.cc"
           ] +
           glob(["boost_locale/**/*.hpp"]),
    includes = ["boost_locale"],
    copts = ["-std=c++11"],
    deps = ["//tensorflow/core:core",
            "//tensorflow/core/util/ctc",
            "//third_party/eigen3",
    ],
)

• 编译过程:

1. cd 到 tensorflow-master 目录下
2. bazel build -c opt --copt=-O3 //tensorflow:libtensorflow_cc.so //custom_op:ctc_decoder_with_param
3. bazel-bin/custom_op 目录下生成 libctc_decoder_with_param.so

在训练（预测）程序中使用自定义的Op

在程序中定义如下的方法：

decode_param_op_module = tf.load_op_library('libctc_decoder_with_param.so')
def decode_with_param(inputs, sequence_length, beam_width=100,
                   top_paths=1, merge_repeated=True):
    decoded_ixs, decoded_vals, decoded_shapes, log_probabilities = (
        decode_param_op_module.ctc_beam_search_decoder_with_param(
            inputs, sequence_length, beam_width=beam_width,
            top_paths=top_paths, merge_repeated=merge_repeated,
            label_selection_size=40, label_selection_margin=0.99))
    return (
        [tf.SparseTensor(ix, val, shape) for (ix, val, shape)
         in zip(decoded_ixs, decoded_vals, decoded_shapes)],
        log_probabilities)

然后就可以像使用tf.nn.ctc_beam_search_decoder一样使用该Op了。