tensor_inspector.h

Go to the documentation of this file.

/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you 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.
 */

#ifndef MXNET_COMMON_TENSOR_INSPECTOR_H_
#define MXNET_COMMON_TENSOR_INSPECTOR_H_

#include <algorithm>
#include <cmath>
#include <string>
#include <vector>
#include <fstream>
#include "../../3rdparty/mshadow/mshadow/base.h"

namespace mxnet {

struct InspectorManager {
  static InspectorManager* get() {
    static std::mutex mtx;
    static std::unique_ptr<InspectorManager> im = nullptr;
    if (!im) {
      std::unique_lock<std::mutex> lk(mtx);
      if (!im)
        im = std::make_unique<InspectorManager>();
    }
    return im.get();
  }
  /* !\brief mutex used to lock interactive_print() and check_value() */
  std::mutex mutex_;
  /* !\brief skip all interactive prints */
  bool interactive_print_skip_all_ = false;
  /* !\brief skip all value checks */
  bool check_value_skip_all_ = false;
  /* !\brief visit count for interactive print tags */
  std::unordered_map<std::string, int> interactive_print_tag_counter_;
  /* !\brief visit count for check value tags */
  std::unordered_map<std::string, int> check_value_tag_counter_;
  /* !\brief visit count for dump value tags */
  std::unordered_map<std::string, int> dump_to_file_tag_counter_;
};

enum CheckerType {
  NegativeChecker,  // check if is negative
  PositiveChecker,  // check if is positive
  ZeroChecker,  // check if is zero
  NaNChecker,  // check if is NaN, will always return false if DType is not a float type
  InfChecker,  // check if is infinity, will always return false if DType is not a float type
  PositiveInfChecker,  // check if is positive infinity,
                       // will always return false if DType is not a float type
  NegativeInfChecker,  // check if is nagative infinity,
                       // will always return false if DType is not a float type
  FiniteChecker,  // check if is finite, will always return false if DType is not a float type
  NormalChecker,  // check if is neither infinity nor NaN
  AbnormalChecker,  // chekck if is infinity or nan
};

class TensorInspector {
 private:
  template<typename DType, typename StreamType>
  void tensor_info_to_string(StreamType* os) {
    const int dimension = tb_.ndim();
    *os << "<" << infer_type_string(typeid(DType)) << " Tensor ";
    *os << tb_.shape_[0];
    for (int i = 1; i < dimension; ++i) {
      *os << 'x' << tb_.shape_[i];
    }
    *os << ">" << std::endl;
  }

  template<typename DType, typename StreamType>
  void tensor_info_to_string(StreamType* os, const std::vector<index_t>& shape) {
    const int dimension = shape.size();
    *os << "<" << infer_type_string(typeid(DType)) << " Tensor ";
    *os << shape[0];
    for (int i = 1; i < dimension; ++i) {
      *os << 'x' << shape[i];
    }
    *os << ">" << std::endl;
  }

  template<typename DType, typename StreamType>
  void to_string_helper(StreamType* os) {
#if MXNET_USE_CUDA
    if (tb_.dev_mask() == gpu::kDevMask) {
      TensorInspector(test::CAccessAsCPU(ctx_, tb_, false)(), ctx_)
          .to_string_helper<DType>(os);
      return;
    }
#endif  // MXNET_USE_CUDA
    const int dimension = tb_.ndim();
    std::vector<index_t> offsets;
    index_t multiple = 1;
    for (int i = dimension - 1; i >= 0; --i) {
      multiple *= tb_.shape_[i];
      offsets.push_back(multiple);
    }
    *os << std::string(dimension, '[');
    *os << tb_.dptr<DType>()[0];
    for (index_t i = 1; i < static_cast<index_t>(tb_.shape_.Size()); ++i) {
      int n = 0;
      for (auto off : offsets) {
        n += (i % off == 0);
      }
      if (n) {
        *os << std::string(n, ']') << ", " <<  std::string(n, '[');
      } else  {
        *os << ", ";
      }
      *os << tb_.dptr<DType>()[i];
    }
    *os << std::string(dimension, ']') << std::endl;
    tensor_info_to_string<DType>(os);
  }

  template<typename DType, typename StreamType>
  void to_string_helper(StreamType* os, const DType* dptr) {
#if MXNET_USE_CUDA
    if (tb_.dev_mask() == gpu::kDevMask) {
      TensorInspector(test::CAccessAsCPU(ctx_, tb_, false)(), ctx_)
          .to_string_helper<DType>(os, dptr);
      return;
    }
#endif  // MXNET_USE_CUDA
    *os << *dptr << std::endl;
    *os << "<" << typeid(*dptr).name() << ">" << std::endl;
  }

  template<typename DType, typename StreamType>
  void to_string_helper(StreamType* os, const std::vector<index_t>& sub_shape, index_t offset) {
#if MXNET_USE_CUDA
    if (tb_.dev_mask() == gpu::kDevMask) {
      TensorInspector(test::CAccessAsCPU(ctx_, tb_, false)(), ctx_)
          .to_string_helper<DType>(os, sub_shape, offset);
      return;
    }
#endif  // MXNET_USE_CUDA
    DType* dptr = tb_.dptr<DType>() + offset;
    if (sub_shape.size() == 0) {
      to_string_helper<DType>(os, dptr);
      return;
    }
    const int dimension = sub_shape.size();
    std::vector<index_t> offsets;
    index_t multiple = 1;
    for (int i = dimension - 1; i >= 0; --i) {
      multiple *= sub_shape[i];
      offsets.push_back(multiple);
    }
    std::stringstream ss;
    *os << std::string(dimension, '[');
    *os << dptr[0];
    for (index_t i = 1; i < multiple; ++i) {
      int n = 0;
      for (auto off : offsets) {
        n += (i % off == 0);
      }
      if (n) {
        *os << std::string(n, ']') << ", " <<  std::string(n, '[');
      } else  {
        *os << ", ";
      }
      *os << dptr[i];
    }
    *os << std::string(dimension, ']') << std::endl;
    tensor_info_to_string<DType>(os, sub_shape);
  }

  void print_locator(const std::vector<index_t>& pos, std::vector<index_t>* sub_shape,
      index_t* offset) {
    const int dimension = tb_.ndim();
    const int sub_dim = dimension - pos.size();
    sub_shape->resize(sub_dim);
    index_t multiple = 1;
    for (size_t i = pos.size(), j = 0; i < static_cast<size_t>(dimension); ++i, ++j) {
      (*sub_shape)[j] = tb_.shape_[i];
      multiple *= tb_.shape_[i];
    }
    index_t sum = 0;
    index_t m = 1;
    for (index_t i = pos.size() - 1; i >= 0; --i) {
      sum += pos[i] * m;
      m *= tb_.shape_[i];
    }
    *offset = sum * multiple;
  }

  bool parse_position(std::vector<index_t>* pos, const std::string& str) {
    const int dimension = tb_.ndim();
    std::istringstream ss(str);
    index_t n;
    while (ss >> n) {
      pos->push_back(n);
      if (ss.peek() == ',') {
        ss.ignore();
      }
    }
    if (pos->size() > static_cast<size_t>(dimension)) {
      return false;
    }
    for (size_t i = 0; i < pos->size(); ++i) {
      if ((*pos)[i] > (tb_.shape_[i] - 1) || (*pos)[i] < 0) {
        return false;
      }
    }
    return !pos->empty();
  }

  template<typename DType>
  void interactive_print_helper(std::string tag) {
#if MXNET_USE_CUDA
    if (tb_.dev_mask() == gpu::kDevMask) {
      TensorInspector(test::CAccessAsCPU(ctx_, tb_, false)(), ctx_)
          .interactive_print_helper<DType>(tag);
      return;
    }
#endif  // MXNET_USE_CUDA
    std::lock_guard<std::mutex> lock(InspectorManager::get()->mutex_);
    InspectorManager::get()->interactive_print_tag_counter_[tag] += 1;
    while (!InspectorManager::get()->interactive_print_skip_all_) {
      std::cout << "----------Interactive Print----------" << std::endl;
      if (tag != "") {
        std::cout << "Tag: " << tag << "  Visit: " <<
            InspectorManager::get()->interactive_print_tag_counter_[tag] <<  std::endl;
      }
      tensor_info_to_string<DType>(&std::cout);
      std::cout << "To print a part of the tensor, " <<
          "please specify a position, seperated by \",\"" << std::endl;
      std::cout << "\"e\" for the entire tensor, " <<
          "\"d\" to dump value to file, " <<
          "\"b\" to break, " <<
          "\"s\" to skip all: ";
      std::string str;
      std::cin >> str;
      if (str == "b") {
        break;
      } else if (str == "e") {
        to_string_helper<DType>(&std::cout);
        continue;
      } else if (str == "s") {
        InspectorManager::get()->interactive_print_skip_all_ = true;
        break;
      } else if (str == "d") {
        while (true) {
          std::cout << "Please enter a tag: ";
          std::cin >> str;
          if (str.find(' ') != std::string::npos) {
            std::cout << "Invalid tag name. No space allowed.";
            continue;
          }
          dump_to_file_helper<DType>(str);
          break;
        }
        continue;
      }
      std::vector<index_t> pos;
      if (parse_position(&pos, str)) {
        std::vector<index_t> sub_shape;
        index_t offset;
        print_locator(pos, &sub_shape, &offset);
        to_string_helper<DType>(&std::cout, sub_shape, offset);
      } else {
        std::cout << "invalid command/indices" << std::endl;
      }
    }
  }

  template<typename DType>
  std::function<bool(DType)> get_checker(CheckerType ct) {
    switch (ct) {
      case NegativeChecker:
        return [] (DType x) {
              return x < 0;
            };
      case PositiveChecker:
        return [] (DType x) {
              return x > 0;
            };
      case ZeroChecker:
        return [] (DType x) {
              return x == 0;
            };
      case NaNChecker:
        if (std::is_same<DType, float>::value || std::is_same<DType, double>::value ||
            std::is_same<DType, mshadow::half::half_t>::value) {
          return [] (DType x) {
                return x != x;
              };
        } else {
          LOG(WARNING) << "NaNChecker only applies to float types. " <<
              "Lambda will always return false.";
        }
        break;
      case InfChecker:
        if (std::is_same<DType, float>::value || std::is_same<DType, double>::value ||
            std::is_same<DType, mshadow::half::half_t>::value) {
          return [] (DType x) {
                return x == (DType)1.0 / 0.0f || x == -(DType)1.0 / 0.0f;
              };
        } else {
          LOG(WARNING) << "InfChecker only applies to float types. " <<
              "Lambda will always return false.";
        }
        break;
      case PositiveInfChecker:
        if (std::is_same<DType, float>::value || std::is_same<DType, double>::value ||
            std::is_same<DType, mshadow::half::half_t>::value) {
          return [] (DType x) {
                return x == (DType)1.0 /  0.0f;
              };
        } else {
          LOG(WARNING) << "PositiveInfChecker only applies to float types. " <<
              "Lambda will always return false.";
        }
        break;
      case NegativeInfChecker:
        if (std::is_same<DType, float>::value || std::is_same<DType, double>::value ||
            std::is_same<DType, mshadow::half::half_t>::value) {
          return [] (DType x) {
                return x == -(DType)1.0 /  0.0f;
              };
        } else {
          LOG(WARNING) << "NegativeInfChecker only applies to float types. " <<
              "Lambda will always return false.";
        }
        break;
      case FiniteChecker:
        if (std::is_same<DType, float>::value || std::is_same<DType, double>::value ||
            std::is_same<DType, mshadow::half::half_t>::value) {
          return [] (DType x) {
                return x != (DType)1.0 /  0.0f && x != -(DType)1.0 /  0.0f;
              };
        } else {
          LOG(WARNING) << "FiniteChecker only applies to float types. " <<
              "Lambda will always return false.";
        }
        break;
      case NormalChecker:
        if (std::is_same<DType, float>::value || std::is_same<DType, double>::value ||
            std::is_same<DType, mshadow::half::half_t>::value) {
          return [] (DType x) {
                return x != (DType)1.0 /  0.0f && x != -(DType)1.0 /  0.0f &&
                    x == x;
              };
        } else {
          LOG(WARNING) << "NormalChecker only applies to float types. " <<
              "Lambda will always return false.";
        }
        break;
      case AbnormalChecker:
        if (std::is_same<DType, float>::value || std::is_same<DType, double>::value ||
            std::is_same<DType, mshadow::half::half_t>::value) {
          return [] (DType x) {
                return x == (DType)1.0 /  0.0f || x == -(DType)1.0 /  0.0f ||
                    x != x;
              };
        } else {
          LOG(WARNING) << "AbnormalChecker only applies to float types. " <<
              "Lambda will always return false.";
        }
        break;
      default:
        return [] (DType x) {
              return false;
            };
    }
    return [] (DType x) {return false;};
  }

  std::vector<index_t> index_to_coordinates(index_t idx) {
    const int dimension = tb_.ndim();
    std::vector<index_t> ret;
    for (int i = dimension - 1; i >= 0; --i) {
      ret.push_back(idx % tb_.shape_[i]);
      idx /= tb_.shape_[i];
    }
    std::reverse(ret.begin(), ret.end());
    return ret;
  }

  template<typename DType>
  void check_value_helper(std::vector<std::vector<index_t>>* ret,
      const std::function<bool(DType)>& checker, bool interactive, std::string tag) {
#if MXNET_USE_CUDA
    if (tb_.dev_mask() == gpu::kDevMask) {
      return TensorInspector(test::CAccessAsCPU(ctx_, tb_, false)(), ctx_)
          .check_value_helper<DType>(ret, checker, interactive, tag);
    }
#endif  // MXNET_USE_CUDA
    index_t count = 0;
    std::stringstream ss;
    ss << "[";
    bool first_pass = true;
    for (index_t i = 0; i <static_cast<index_t>(tb_.shape_.Size()); ++i) {
      if (checker(tb_.dptr<DType>()[i])) {
        ++count;
        if (!first_pass) {
          ss  << ", ";
        }
        first_pass = false;
        std::vector<index_t> coords = index_to_coordinates(i);
        ss << "(" << coords[0];
        for (size_t i = 1; i < coords.size(); ++i) {
          ss << ", " << coords[i];
        }
        ss << ")";
        ret->push_back(coords);
      }
    }
    ss << "]" << std::endl;
    if (interactive) {
      std::lock_guard<std::mutex> lock(InspectorManager::get()->mutex_);
       InspectorManager::get()->check_value_tag_counter_[tag] += 1;
      while (!InspectorManager::get()->check_value_skip_all_) {
        std::cout << "----------Value Check----------" << std::endl;
        tensor_info_to_string<DType>(&std::cout);
        if (tag != "") {
          std::cout << "Tag: " << tag << "  Visit: " <<
              InspectorManager::get()->check_value_tag_counter_[tag] <<  std::endl;
        }
        std::cout << count << " value(s) found." << std::endl;
        std::cout << "To print a part of the tensor," <<
            " please specify a position, seperated by \",\"" << std::endl;
        std::cout << "\"e\" for the entire tensor, " <<
            "\"p\" to print the coordinates of the values found, " <<
            "\"b\" to break, " <<
            "\"s\" to skip all: ";
        std::string str;
        std::cin >> str;
        if (str == "b") {
          break;
        } else if (str == "e") {
          to_string_helper<DType>(&std::cout);
          continue;
        } else if (str == "p") {
          std::cout << ss.str() << std::endl;
          continue;
        } else if (str == "s") {
          InspectorManager::get()->check_value_skip_all_ = true;
          break;
        }
        std::vector<index_t> pos;
        if (parse_position(&pos, str)) {
          std::vector<index_t> sub_shape;
          index_t offset;
          print_locator(pos, &sub_shape, &offset);
          to_string_helper<DType>(&std::cout, sub_shape, offset);
        } else {
          std::cout << "invalid command/indices" << std::endl;
        }
      }
    }
  }

  inline char infer_type(const std::type_info& ti) {
    if (ti == typeid(float)) return 'f';
    else if (ti == typeid(double)) return 'f';
    else if (ti == typeid(mshadow::half::half_t) ) return 'f';
    else if (ti == typeid(uint8_t)) return 'u';
    else if (ti == typeid(int32_t)) return 'i';
    else if (ti == typeid(int64_t)) return 'i';
    else
      return '?';
  }

  inline std::string infer_type_string(const std::type_info& ti) {
    if (ti == typeid(float)) return "float";
    else if (ti == typeid(double)) return "double";
    else if (ti == typeid(mshadow::half::half_t) ) return "mshasow::half::half_t";
    else if (ti == typeid(uint8_t)) return "uint8_t";
    else if (ti == typeid(int32_t)) return "int32_t";
    else if (ti == typeid(int64_t)) return "int64_t";
    else
      return "unknown tyoe";
  }

  inline char endian_test() {
    int x = 1;
    return (reinterpret_cast<char*>(&x)[0]) ? '<' : '>';
  }

  template<typename DType>
  std::string get_header() {
    const int dimension = tb_.ndim();
    std::string dict;
    dict += "{'descr':'";
    dict += endian_test();
    dict += infer_type(typeid(DType));
    dict += std::to_string(sizeof(DType));
    dict += "','fortran_order':False,'shape':(";
    dict += std::to_string(tb_.shape_[0]);
    for (int i = 1; i < dimension; ++i) {
      dict += ',';
      dict += std::to_string(tb_.shape_[i]);
    }
    if (dimension == 1) {
       dict += ",";
    }
    dict += ")} ";
    int padding_size = 64 - ((10 + dict.size()) % 64);
    dict += std::string(padding_size, ' ');
    dict.back() = '\n';
    std::string header;
    header += static_cast<char>(0x93);
    header += "NUMPY";
    header += static_cast<char>(0x01);
    header += static_cast<char>(0x00);
    header += static_cast<char>((uint16_t)dict.size() & 0x00ff);
    header += static_cast<char>(((uint16_t)dict.size() >> 8) & 0x00ff);
    header += dict;
    return header;
  }

  template<typename DType>
  void write_npy(const std::string& header, const std::string& filename) {
    std::ofstream file;
    file.exceptions(std::ofstream::failbit | std::ofstream::badbit);
    try {
      file.open(filename, std::ios::out | std::ios::binary);
      file.write(header.c_str(), header.size());
      file.write(reinterpret_cast<char*>(tb_.dptr<DType>()), sizeof(DType) * tb_.shape_.Size());
      file.close();
      std::cout << "Tensor dumped to file: " << filename << std::endl;
    } catch (std::ofstream::failure e) {
      std::cerr << "Exception opening/writing/closing file " << filename << std::endl;
    }
  }

  template<typename DType>
  void dump_to_file_helper(const std::string& tag) {
#if MXNET_USE_CUDA
    if (tb_.dev_mask() == gpu::kDevMask) {
      TensorInspector(test::CAccessAsCPU(ctx_, tb_, false)(), ctx_)
          .dump_to_file_helper<DType>(tag);
      return;
    }
#endif  // MXNET_USE_CUDA
    std::string header = get_header<DType>();
    InspectorManager::get()->dump_to_file_tag_counter_[tag] += 1;
    const int visit = InspectorManager::get()->dump_to_file_tag_counter_[tag];
    std::string filename = tag + "_" + std::to_string(visit) + ".npy";
    write_npy<DType>(header, filename);
  }

  inline void validate_shape() {
    const int dimension = tb_.ndim();
    CHECK(dimension > 0) << "Tensor Inspector does not support empty tensors " <<
        "or tensors of unknow shape.";
    for (int i = 0; i < dimension; ++i) {
      CHECK(tb_.shape_[i] != 0) << "Invalid tensor shape: shape_[" << i << "] is 0";
    }
  }

  /* !\brief the tensor blob */
  const TBlob tb_;
  /* !\brief the run context of the tensor */
  const RunContext& ctx_;

 public:
  template<typename Device, int dimension, typename DType>
  TensorInspector(const mshadow::Tensor<Device, dimension, DType>& ts, const RunContext& ctx):
      tb_(ts), ctx_(ctx) {
    validate_shape();
  }

  TensorInspector(const TBlob& tb, const RunContext& ctx):
      tb_(tb), ctx_(ctx) {
    validate_shape();
  }

  TensorInspector(const NDArray& arr, const RunContext& ctx):
      tb_(arr.data()), ctx_(ctx) {
    validate_shape();
  }

  void print_string() {
    std::cout << to_string() << std::endl;
  }

  std::string to_string() {
    std::stringstream ss;
    MSHADOW_TYPE_SWITCH(tb_.type_flag_, DType, {
      to_string_helper<DType>(&ss);
    });
    return ss.str();
  }

  void interactive_print(std::string tag = "") {
    MSHADOW_TYPE_SWITCH(tb_.type_flag_, DType, {
      interactive_print_helper<DType>(tag);
    });
  }

  template<typename ValueChecker>
  std::vector<std::vector<index_t>> check_value(const ValueChecker& checker,
      bool interactive = false, std::string tag = "") {
    std::vector<std::vector<index_t>> ret;
    MSHADOW_TYPE_SWITCH(tb_.type_flag_, DType, {
      check_value_helper<DType>(&ret, checker, ret, interactive, tag);
    });
    return ret;
  }

  std::vector<std::vector<index_t>> check_value(CheckerType ct, bool interactive = false,
      std::string tag = "") {
    std::vector<std::vector<index_t>> ret;
    MSHADOW_TYPE_SWITCH(tb_.type_flag_, DType, {
      check_value_helper<DType>(&ret, get_checker<DType>(ct), interactive, tag);
    });
    return ret;
  }

  void dump_to_file(std::string tag) {
    MSHADOW_TYPE_SWITCH(tb_.type_flag_, DType, {
      dump_to_file_helper<DType>(tag);
    });
  }
};

}  // namespace mxnet

#endif  // MXNET_COMMON_TENSOR_INSPECTOR_H_