!12484 MD C++ api decouple ABI compile macro

From: @luoyang42
Reviewed-by: 
Signed-off-by:

include/api/dual_abi_helper.h

@@ -16,11 +16,147 @@
 #ifndef MINDSPORE_INCLUDE_API_DUAL_ABI_HELPER_H_
 #define MINDSPORE_INCLUDE_API_DUAL_ABI_HELPER_H_
 
+#include <algorithm>
+#include <map>
+#include <memory>
+#include <optional>
 #include <string>
+#include <set>
+#include <unordered_map>
+#include <utility>
 #include <vector>
 
 namespace mindspore {
 inline std::vector<char> StringToChar(const std::string &s) { return std::vector<char>(s.begin(), s.end()); }
+
 inline std::string CharToString(const std::vector<char> &c) { return std::string(c.begin(), c.end()); }
+
+inline std::optional<std::vector<char>> OptionalStringToChar(const std::optional<std::string> &s) {
+  std::optional<std::vector<char>> ret = std::vector<char>(s->begin(), s->end());
+  return (s == std::nullopt) ? std::nullopt : ret;
+}
+
+inline std::optional<std::string> OptionalCharToString(const std::optional<std::vector<char>> &c) {
+  std::optional<std::string> ret = std::string(c->begin(), c->end());
+  return (c == std::nullopt) ? std::nullopt : ret;
+}
+
+inline std::pair<std::vector<char>, int32_t> PairStringToChar(const std::pair<std::string, int32_t> &s) {
+  return std::pair<std::vector<char>, int32_t>(std::vector<char>(s.first.begin(), s.first.end()), s.second);
+}
+
+inline std::pair<std::string, int32_t> PairCharToString(const std::pair<std::vector<char>, int32_t> &c) {
+  return std::pair<std::string, int32_t>(std::string(c.first.begin(), c.first.end()), c.second);
+}
+
+inline std::vector<std::vector<char>> VectorStringToChar(const std::vector<std::string> &s) {
+  std::vector<std::vector<char>> ret;
+  std::transform(s.begin(), s.end(), std::back_inserter(ret),
+                 [](auto str) { return std::vector<char>(str.begin(), str.end()); });
+  return ret;
+}
+
+inline std::vector<std::string> VectorCharToString(const std::vector<std::vector<char>> &c) {
+  std::vector<std::string> ret;
+  std::transform(c.begin(), c.end(), std::back_inserter(ret),
+                 [](auto ch) { return std::string(ch.begin(), ch.end()); });
+  return ret;
+}
+
+inline std::set<std::vector<char>> SetStringToChar(const std::set<std::string> &s) {
+  std::set<std::vector<char>> ret;
+  std::transform(s.begin(), s.end(), std::inserter(ret, ret.begin()),
+                 [](auto str) { return std::vector<char>(str.begin(), str.end()); });
+  return ret;
+}
+
+inline std::set<std::string> SetCharToString(const std::set<std::vector<char>> &c) {
+  std::set<std::string> ret;
+  std::transform(c.begin(), c.end(), std::inserter(ret, ret.begin()),
+                 [](auto ch) { return std::string(ch.begin(), ch.end()); });
+  return ret;
+}
+
+inline std::map<std::vector<char>, int32_t> MapStringToChar(const std::map<std::string, int32_t> &s) {
+  std::map<std::vector<char>, int32_t> ret;
+  std::transform(s.begin(), s.end(), std::inserter(ret, ret.begin()), [](auto str) {
+    return std::pair<std::vector<char>, int32_t>(std::vector<char>(str.first.begin(), str.first.end()), str.second);
+  });
+  return ret;
+}
+
+inline std::map<std::string, int32_t> MapCharToString(const std::map<std::vector<char>, int32_t> &c) {
+  std::map<std::string, int32_t> ret;
+  std::transform(c.begin(), c.end(), std::inserter(ret, ret.begin()), [](auto ch) {
+    return std::pair<std::string, int32_t>(std::string(ch.first.begin(), ch.first.end()), ch.second);
+  });
+  return ret;
+}
+
+inline std::map<std::vector<char>, std::vector<char>> UnorderedMapStringToChar(
+  const std::unordered_map<std::string, std::string> &s) {
+  std::map<std::vector<char>, std::vector<char>> ret;
+  std::transform(s.begin(), s.end(), std::inserter(ret, ret.begin()), [](auto str) {
+    return std::pair<std::vector<char>, std::vector<char>>(std::vector<char>(str.first.begin(), str.first.end()),
+                                                           std::vector<char>(str.second.begin(), str.second.end()));
+  });
+  return ret;
+}
+
+inline std::unordered_map<std::string, std::string> UnorderedMapCharToString(
+  const std::map<std::vector<char>, std::vector<char>> &c) {
+  std::unordered_map<std::string, std::string> ret;
+  std::transform(c.begin(), c.end(), std::inserter(ret, ret.begin()), [](auto ch) {
+    return std::pair<std::string, std::string>(std::string(ch.first.begin(), ch.first.end()),
+                                               std::string(ch.second.begin(), ch.second.end()));
+  });
+  return ret;
+}
+
+inline std::vector<std::pair<std::vector<char>, std::vector<int32_t>>> ClassIndexStringToChar(
+  const std::vector<std::pair<std::string, std::vector<int32_t>>> &s) {
+  std::vector<std::pair<std::vector<char>, std::vector<int32_t>>> ret;
+  std::transform(s.begin(), s.end(), std::back_inserter(ret), [](auto str) {
+    return std::pair<std::vector<char>, std::vector<int32_t>>(std::vector<char>(str.first.begin(), str.first.end()),
+                                                              str.second);
+  });
+  return ret;
+}
+
+inline std::vector<std::pair<std::string, std::vector<int32_t>>> ClassIndexCharToString(
+  const std::vector<std::pair<std::vector<char>, std::vector<int32_t>>> &c) {
+  std::vector<std::pair<std::string, std::vector<int32_t>>> ret;
+  std::transform(c.begin(), c.end(), std::back_inserter(ret), [](auto ch) {
+    return std::pair<std::string, std::vector<int32_t>>(std::string(ch.first.begin(), ch.first.end()), ch.second);
+  });
+  return ret;
+}
+
+template <class T>
+inline std::map<std::vector<char>, T> PadInfoStringToChar(const std::map<std::string, T> &s_pad_info) {
+  std::map<std::vector<char>, T> ret;
+  std::transform(s_pad_info.begin(), s_pad_info.end(), std::inserter(ret, ret.begin()), [](auto str) {
+    return std::pair<std::vector<char>, T>(std::vector<char>(str.first.begin(), str.first.end()), str.second);
+  });
+  return ret;
+}
+
+template <class T>
+inline std::map<std::string, T> PadInfoCharToString(const std::map<std::vector<char>, T> &c_pad_info) {
+  std::map<std::string, T> ret;
+  std::transform(c_pad_info.begin(), c_pad_info.end(), std::inserter(ret, ret.begin()), [](auto ch) {
+    return std::pair<std::string, T>(std::string(ch.first.begin(), ch.first.end()), ch.second);
+  });
+  return ret;
+}
+
+template <class T>
+inline void TensorMapCharToString(const std::map<std::vector<char>, T> *c, std::unordered_map<std::string, T> *s) {
+  for (auto ch : *c) {
+    auto key = std::string(ch.first.begin(), ch.first.end());
+    auto val = ch.second;
+    s->insert(std::pair<std::string, T>(key, val));
+  }
+}
 }  // namespace mindspore
 #endif  // MINDSPORE_INCLUDE_API_DUAL_ABI_HELPER_H_

mindspore/ccsrc/minddata/dataset/api/config.cc

@@ -94,8 +94,8 @@ bool set_callback_timeback(int32_t timeout) {
 int32_t get_callback_timeout() { return _config->callback_timeout(); }
 
 // Function to load configurations from a file
-bool load(std::string file) {
+bool load(const std::vector<char> &file) {
-  Status rc = _config->LoadFile(file);
+  Status rc = _config->LoadFile(CharToString(file));
   if (rc.IsError()) {
     MS_LOG(ERROR) << rc << file;
     return false;

mindspore/ccsrc/minddata/dataset/api/iterator.cc

@@ -26,7 +26,7 @@ Iterator::Iterator() : consumer_(nullptr) {}
 Iterator::~Iterator() { Stop(); }
 
 // Get the next row from the data pipeline.
-Status Iterator::GetNextRow(MSTensorMap *row) {
+Status Iterator::GetNextRowCharIF(MSTensorMapChar *row) {
   // Clean data buffer
   row->clear();
   std::unordered_map<std::string, std::shared_ptr<dataset::Tensor>> md_map;
@@ -38,7 +38,8 @@ Status Iterator::GetNextRow(MSTensorMap *row) {
   }
   for (auto de_tensor : md_map) {
     CHECK_FAIL_RETURN_UNEXPECTED(de_tensor.second->HasData(), "Apply transform failed, output tensor has no data");
-    row->insert(std::make_pair(de_tensor.first, mindspore::MSTensor(std::make_shared<DETensor>(de_tensor.second))));
+    std::vector<char> col_name(de_tensor.first.begin(), de_tensor.first.end());
+    row->insert(std::make_pair(col_name, mindspore::MSTensor(std::make_shared<DETensor>(de_tensor.second))));
   }
 
   return Status::OK();

mindspore/ccsrc/minddata/dataset/api/text.cc

@@ -31,38 +31,66 @@ namespace text {
 
 #ifndef _WIN32
 // BasicTokenizer
+struct BasicTokenizer::Data {
+  Data(bool lower_case, bool keep_whitespace, const NormalizeForm normalize_form, bool preserve_unused_token,
+       bool with_offsets)
+      : lower_case_(lower_case),
+        keep_whitespace_(keep_whitespace),
+        normalize_form_(normalize_form),
+        preserve_unused_token_(preserve_unused_token),
+        with_offsets_(with_offsets) {}
+  bool lower_case_;
+  bool keep_whitespace_;
+  NormalizeForm normalize_form_;
+  bool preserve_unused_token_;
+  bool with_offsets_;
+};
+
 BasicTokenizer::BasicTokenizer(bool lower_case, bool keep_whitespace, const NormalizeForm normalize_form,
                                bool preserve_unused_token, bool with_offsets)
-    : lower_case_(lower_case),
+    : data_(std::make_shared<Data>(lower_case, keep_whitespace, normalize_form, preserve_unused_token, with_offsets)) {}
-      keep_whitespace_(keep_whitespace),
-      normalize_form_(normalize_form),
-      preserve_unused_token_(preserve_unused_token),
-      with_offsets_(with_offsets) {}
 
 std::shared_ptr<TensorOperation> BasicTokenizer::Parse() {
-  return std::make_shared<BasicTokenizerOperation>(lower_case_, keep_whitespace_, normalize_form_,
+  return std::make_shared<BasicTokenizerOperation>(data_->lower_case_, data_->keep_whitespace_, data_->normalize_form_,
-                                                   preserve_unused_token_, with_offsets_);
+                                                   data_->preserve_unused_token_, data_->with_offsets_);
 }
 
 // BertTokenizer
-BertTokenizer::BertTokenizer(const std::shared_ptr<Vocab> &vocab, const std::string &suffix_indicator,
+struct BertTokenizer::Data {
-                             int32_t max_bytes_per_token, const std::string &unknown_token, bool lower_case,
+  Data(const std::shared_ptr<Vocab> &vocab, const std::vector<char> &suffix_indicator, int32_t max_bytes_per_token,
+       const std::vector<char> &unknown_token, bool lower_case, bool keep_whitespace,
+       const NormalizeForm normalize_form, bool preserve_unused_token, bool with_offsets)
+      : vocab_(vocab),
+        suffix_indicator_(CharToString(suffix_indicator)),
+        max_bytes_per_token_(max_bytes_per_token),
+        unknown_token_(CharToString(unknown_token)),
+        lower_case_(lower_case),
+        keep_whitespace_(keep_whitespace),
+        normalize_form_(normalize_form),
+        preserve_unused_token_(preserve_unused_token),
+        with_offsets_(with_offsets) {}
+  std::shared_ptr<Vocab> vocab_;
+  std::string suffix_indicator_;
+  int32_t max_bytes_per_token_;
+  std::string unknown_token_;
+  bool lower_case_;
+  bool keep_whitespace_;
+  NormalizeForm normalize_form_;
+  bool preserve_unused_token_;
+  bool with_offsets_;
+};
+
+BertTokenizer::BertTokenizer(const std::shared_ptr<Vocab> &vocab, const std::vector<char> &suffix_indicator,
+                             int32_t max_bytes_per_token, const std::vector<char> &unknown_token, bool lower_case,
                              bool keep_whitespace, const NormalizeForm normalize_form, bool preserve_unused_token,
                              bool with_offsets)
-    : vocab_(vocab),
+    : data_(std::make_shared<Data>(vocab, suffix_indicator, max_bytes_per_token, unknown_token, lower_case,
-      suffix_indicator_(suffix_indicator),
+                                   keep_whitespace, normalize_form, preserve_unused_token, with_offsets)) {}
-      max_bytes_per_token_(max_bytes_per_token),
-      unknown_token_(unknown_token),
-      lower_case_(lower_case),
-      keep_whitespace_(keep_whitespace),
-      normalize_form_(normalize_form),
-      preserve_unused_token_(preserve_unused_token),
-      with_offsets_(with_offsets) {}
 
 std::shared_ptr<TensorOperation> BertTokenizer::Parse() {
-  return std::make_shared<BertTokenizerOperation>(vocab_, suffix_indicator_, max_bytes_per_token_, unknown_token_,
+  return std::make_shared<BertTokenizerOperation>(
-                                                  lower_case_, keep_whitespace_, normalize_form_,
+    data_->vocab_, data_->suffix_indicator_, data_->max_bytes_per_token_, data_->unknown_token_, data_->lower_case_,
-                                                  preserve_unused_token_, with_offsets_);
+    data_->keep_whitespace_, data_->normalize_form_, data_->preserve_unused_token_, data_->with_offsets_);
 }
 
 // CaseFold
@@ -72,14 +100,28 @@ std::shared_ptr<TensorOperation> CaseFold::Parse() { return std::make_shared<Cas
 #endif
 
 // JiebaTokenizer
-JiebaTokenizer::JiebaTokenizer(const std::string &hmm_path, const std::string &mp_path, const JiebaMode &mode,
+struct JiebaTokenizer::Data {
-                               bool with_offsets)
+  Data(const std::vector<char> &hmm_path, const std::vector<char> &mp_path, const JiebaMode &mode, bool with_offsets)
-    : hmm_path_(hmm_path), mp_path_(mp_path), mode_(mode), with_offsets_(with_offsets) {}
+      : hmm_path_(CharToString(hmm_path)),
+        mp_path_(CharToString(mp_path)),
+        mode_(mode),
+        with_offsets_(with_offsets),
+        words_list_({}) {}
+  std::string hmm_path_;
+  std::string mp_path_;
+  JiebaMode mode_;
+  bool with_offsets_;
+  std::vector<std::pair<std::string, int64_t>> words_list_;
+};
+
+JiebaTokenizer::JiebaTokenizer(const std::vector<char> &hmm_path, const std::vector<char> &mp_path,
+                               const JiebaMode &mode, bool with_offsets)
+    : data_(std::make_shared<Data>(hmm_path, mp_path, mode, with_offsets)) {}
 
 std::shared_ptr<TensorOperation> JiebaTokenizer::Parse() {
   std::shared_ptr<JiebaTokenizerOperation> jieba_tokenizer =
-    std::make_shared<JiebaTokenizerOperation>(hmm_path_, mp_path_, mode_, with_offsets_);
+    std::make_shared<JiebaTokenizerOperation>(data_->hmm_path_, data_->mp_path_, data_->mode_, data_->with_offsets_);
-  for (auto &word : words_list_) {
+  for (auto &word : data_->words_list_) {
     Status rc = jieba_tokenizer->AddWord(word.first, word.second);
     if (rc.IsError()) {
       MS_LOG(ERROR) << rc;
@@ -100,109 +142,199 @@ Status JiebaTokenizer::AddWord(const std::string &word, int64_t freq) {
     MS_LOG(ERROR) << err_msg;
     RETURN_STATUS_SYNTAX_ERROR(err_msg);
   }
-  words_list_.emplace_back(word, freq);
+  data_->words_list_.emplace_back(word, freq);
   return Status::OK();
 }
 
 // Lookup
-Lookup::Lookup(const std::shared_ptr<Vocab> &vocab, const std::optional<std::string> &unknown_token,
+struct Lookup::Data {
-               const std::string &data_type)
+  Data(const std::shared_ptr<Vocab> &vocab, const std::optional<std::vector<char>> &unknown_token,
-    : vocab_(vocab), unknown_token_(unknown_token), data_type_(data_type) {}
+       const std::vector<char> &data_type)
+      : vocab_(vocab), unknown_token_(OptionalCharToString(unknown_token)), data_type_(CharToString(data_type)) {}
+  std::shared_ptr<Vocab> vocab_;
+  std::optional<std::string> unknown_token_;
+  std::string data_type_;
+};
+
+Lookup::Lookup(const std::shared_ptr<Vocab> &vocab, const std::optional<std::vector<char>> &unknown_token,
+               const std::vector<char> &data_type)
+    : data_(std::make_shared<Data>(vocab, unknown_token, data_type)) {}
 
 std::shared_ptr<TensorOperation> Lookup::Parse() {
-  return std::make_shared<LookupOperation>(vocab_, unknown_token_, data_type_);
+  return std::make_shared<LookupOperation>(data_->vocab_, data_->unknown_token_, data_->data_type_);
 }
 
 // Ngram
-Ngram::Ngram(const std::vector<int32_t> &ngrams, const std::pair<std::string, int32_t> &left_pad,
+struct Ngram::Data {
-             const std::pair<std::string, int32_t> &right_pad, const std::string &separator)
+  Data(const std::vector<int32_t> &ngrams, const std::pair<std::vector<char>, int32_t> &left_pad,
-    : ngrams_(ngrams), left_pad_(left_pad), right_pad_(right_pad), separator_(separator) {}
+       const std::pair<std::vector<char>, int32_t> &right_pad, const std::vector<char> &separator)
+      : ngrams_(ngrams),
+        left_pad_(PairCharToString(left_pad)),
+        right_pad_(PairCharToString(right_pad)),
+        separator_(CharToString(separator)) {}
+  std::vector<int32_t> ngrams_;
+  std::pair<std::string, int32_t> left_pad_;
+  std::pair<std::string, int32_t> right_pad_;
+  std::string separator_;
+};
+
+Ngram::Ngram(const std::vector<int32_t> &ngrams, const std::pair<std::vector<char>, int32_t> &left_pad,
+             const std::pair<std::vector<char>, int32_t> &right_pad, const std::vector<char> &separator)
+    : data_(std::make_shared<Data>(ngrams, left_pad, right_pad, separator)) {}
 
 std::shared_ptr<TensorOperation> Ngram::Parse() {
-  return std::make_shared<NgramOperation>(ngrams_, left_pad_, right_pad_, separator_);
+  return std::make_shared<NgramOperation>(data_->ngrams_, data_->left_pad_, data_->right_pad_, data_->separator_);
 }
 
 #ifndef _WIN32
 // NormalizeUTF8
-NormalizeUTF8::NormalizeUTF8(NormalizeForm normalize_form) : normalize_form_(normalize_form) {}
+struct NormalizeUTF8::Data {
+  explicit Data(NormalizeForm normalize_form) : normalize_form_(normalize_form) {}
+  NormalizeForm normalize_form_;
+};
+
+NormalizeUTF8::NormalizeUTF8(NormalizeForm normalize_form) : data_(std::make_shared<Data>(normalize_form)) {}
 
 std::shared_ptr<TensorOperation> NormalizeUTF8::Parse() {
-  return std::make_shared<NormalizeUTF8Operation>(normalize_form_);
+  return std::make_shared<NormalizeUTF8Operation>(data_->normalize_form_);
 }
 
 // RegexReplace
-RegexReplace::RegexReplace(std::string pattern, std::string replace, bool replace_all)
+struct RegexReplace::Data {
-    : pattern_(pattern), replace_(replace), replace_all_(replace_all) {}
+  Data(const std::vector<char> &pattern, const std::vector<char> &replace, bool replace_all)
+      : pattern_(CharToString(pattern)), replace_(CharToString(replace)), replace_all_(replace_all) {}
+  std::string pattern_;
+  std::string replace_;
+  bool replace_all_;
+};
+
+RegexReplace::RegexReplace(const std::vector<char> &pattern, const std::vector<char> &replace, bool replace_all)
+    : data_(std::make_shared<Data>(pattern, replace, replace_all)) {}
 
 std::shared_ptr<TensorOperation> RegexReplace::Parse() {
-  return std::make_shared<RegexReplaceOperation>(pattern_, replace_, replace_all_);
+  return std::make_shared<RegexReplaceOperation>(data_->pattern_, data_->replace_, data_->replace_all_);
 }
 
 // RegexTokenizer
-RegexTokenizer::RegexTokenizer(std::string delim_pattern, std::string keep_delim_pattern, bool with_offsets)
+struct RegexTokenizer::Data {
-    : delim_pattern_(delim_pattern), keep_delim_pattern_(keep_delim_pattern), with_offsets_(with_offsets) {}
+  Data(const std::vector<char> &delim_pattern, const std::vector<char> &keep_delim_pattern, bool with_offsets)
+      : delim_pattern_(CharToString(delim_pattern)),
+        keep_delim_pattern_(CharToString(keep_delim_pattern)),
+        with_offsets_(with_offsets) {}
+  std::string delim_pattern_;
+  std::string keep_delim_pattern_;
+  bool with_offsets_;
+};
+
+RegexTokenizer::RegexTokenizer(const std::vector<char> &delim_pattern, const std::vector<char> &keep_delim_pattern,
+                               bool with_offsets)
+    : data_(std::make_shared<Data>(delim_pattern, keep_delim_pattern, with_offsets)) {}
 
 std::shared_ptr<TensorOperation> RegexTokenizer::Parse() {
-  return std::make_shared<RegexTokenizerOperation>(delim_pattern_, keep_delim_pattern_, with_offsets_);
+  return std::make_shared<RegexTokenizerOperation>(data_->delim_pattern_, data_->keep_delim_pattern_,
+                                                   data_->with_offsets_);
 }
 #endif
 
 // SentencePieceTokenizer
+struct SentencePieceTokenizer::Data {
+  Data(const std::shared_ptr<SentencePieceVocab> &vocab, SPieceTokenizerOutType out_type)
+      : vocab_(vocab), out_type_(out_type) {}
+  Data(const std::vector<char> &vocab_path, SPieceTokenizerOutType out_type)
+      : vocab_path_(CharToString(vocab_path)), out_type_(out_type) {}
+  std::shared_ptr<SentencePieceVocab> vocab_;
+  std::string vocab_path_;
+  SPieceTokenizerLoadType load_type_;
+  SPieceTokenizerOutType out_type_;
+};
+
 SentencePieceTokenizer::SentencePieceTokenizer(const std::shared_ptr<SentencePieceVocab> &vocab,
                                                SPieceTokenizerOutType out_type)
-    : vocab_(vocab), out_type_(out_type) {}
+    : data_(std::make_shared<Data>(vocab, out_type)) {}
 
-SentencePieceTokenizer::SentencePieceTokenizer(const std::string &vocab_path, SPieceTokenizerOutType out_type)
+SentencePieceTokenizer::SentencePieceTokenizer(const std::vector<char> &vocab_path, SPieceTokenizerOutType out_type)
-    : vocab_path_(vocab_path), out_type_(out_type) {}
+    : data_(std::make_shared<Data>(vocab_path, out_type)) {}
 
 std::shared_ptr<TensorOperation> SentencePieceTokenizer::Parse() {
-  if (vocab_ != nullptr) {
+  if (data_->vocab_ != nullptr) {
-    return std::make_shared<SentencePieceTokenizerOperation>(vocab_, out_type_);
+    return std::make_shared<SentencePieceTokenizerOperation>(data_->vocab_, data_->out_type_);
   } else {
-    return std::make_shared<SentencePieceTokenizerOperation>(vocab_path_, out_type_);
+    return std::make_shared<SentencePieceTokenizerOperation>(data_->vocab_path_, data_->out_type_);
   }
 }
 
 // SlidingWindow
-SlidingWindow::SlidingWindow(const int32_t width, const int32_t axis) : width_(width), axis_(axis) {}
+struct SlidingWindow::Data {
+  Data(const int32_t width, const int32_t axis) : width_(width), axis_(axis) {}
+  int32_t width_;
+  int32_t axis_;
+};
+
+SlidingWindow::SlidingWindow(const int32_t width, const int32_t axis) : data_(std::make_shared<Data>(width, axis)) {}
 
 std::shared_ptr<TensorOperation> SlidingWindow::Parse() {
-  return std::make_shared<SlidingWindowOperation>(width_, axis_);
+  return std::make_shared<SlidingWindowOperation>(data_->width_, data_->axis_);
 }
 
 // ToNumber
-ToNumber::ToNumber(const std::string &data_type) : data_type_(data_type) {}
+struct ToNumber::Data {
+  explicit Data(const std::vector<char> &data_type) : data_type_(CharToString(data_type)) {}
+  std::string data_type_;
+};
 
-std::shared_ptr<TensorOperation> ToNumber::Parse() { return std::make_shared<ToNumberOperation>(data_type_); }
+ToNumber::ToNumber(const std::vector<char> &data_type) : data_(std::make_shared<Data>(data_type)) {}
+
+std::shared_ptr<TensorOperation> ToNumber::Parse() { return std::make_shared<ToNumberOperation>(data_->data_type_); }
 
 // TruncateSequencePair
-TruncateSequencePair::TruncateSequencePair(int32_t max_length) : max_length_(max_length) {}
+struct TruncateSequencePair::Data {
+  explicit Data(int32_t max_length) : max_length_(max_length) {}
+  int32_t max_length_;
+};
+
+TruncateSequencePair::TruncateSequencePair(int32_t max_length) : data_(std::make_shared<Data>(max_length)) {}
 
 std::shared_ptr<TensorOperation> TruncateSequencePair::Parse() {
-  return std::make_shared<TruncateSequencePairOperation>(max_length_);
+  return std::make_shared<TruncateSequencePairOperation>(data_->max_length_);
 }
 
 // UnicodeCharTokenizer
-UnicodeCharTokenizer::UnicodeCharTokenizer(bool with_offsets) : with_offsets_(with_offsets) {}
+struct UnicodeCharTokenizer::Data {
+  explicit Data(bool with_offsets) : with_offsets_(with_offsets) {}
+  bool with_offsets_;
+};
+
+UnicodeCharTokenizer::UnicodeCharTokenizer(bool with_offsets) : data_(std::make_shared<Data>(with_offsets)) {}
 
 std::shared_ptr<TensorOperation> UnicodeCharTokenizer::Parse() {
-  return std::make_shared<UnicodeCharTokenizerOperation>(with_offsets_);
+  return std::make_shared<UnicodeCharTokenizerOperation>(data_->with_offsets_);
 }
 
 #ifndef _WIN32
 // UnicodeScriptTokenizer
+struct UnicodeScriptTokenizer::Data {
+  Data(bool keep_whitespace, bool with_offsets) : keep_whitespace_(keep_whitespace), with_offsets_(with_offsets) {}
+  bool keep_whitespace_;
+  bool with_offsets_;
+};
+
 UnicodeScriptTokenizer::UnicodeScriptTokenizer(bool keep_whitespace, bool with_offsets)
-    : keep_whitespace_(keep_whitespace), with_offsets_(with_offsets) {}
+    : data_(std::make_shared<Data>(keep_whitespace, with_offsets)) {}
 
 std::shared_ptr<TensorOperation> UnicodeScriptTokenizer::Parse() {
-  return std::make_shared<UnicodeScriptTokenizerOperation>(keep_whitespace_, with_offsets_);
+  return std::make_shared<UnicodeScriptTokenizerOperation>(data_->keep_whitespace_, data_->with_offsets_);
 }
 
 // WhitespaceTokenizer
-WhitespaceTokenizer::WhitespaceTokenizer(bool with_offsets) : with_offsets_(with_offsets) {}
+struct WhitespaceTokenizer::Data {
+  explicit Data(bool with_offsets) : with_offsets_(with_offsets) {}
+  bool with_offsets_;
+};
+
+WhitespaceTokenizer::WhitespaceTokenizer(bool with_offsets) : data_(std::make_shared<Data>(with_offsets)) {}
 
 std::shared_ptr<TensorOperation> WhitespaceTokenizer::Parse() {
-  return std::make_shared<WhitespaceTokenizerOperation>(with_offsets_);
+  return std::make_shared<WhitespaceTokenizerOperation>(data_->with_offsets_);
 }
 #endif
 }  // namespace text

mindspore/ccsrc/minddata/dataset/api/transforms.cc

@@ -30,25 +30,30 @@ namespace transforms {
 // (In alphabetical order)
 
 // Constructor to Compose.
-Compose::Compose(const std::vector<TensorTransform *> &transforms) {
+struct Compose::Data {
+  std::vector<std::shared_ptr<TensorOperation>> transforms_;
+};
+
+Compose::Compose(const std::vector<TensorTransform *> &transforms) : data_(std::make_shared<Data>()) {
   (void)std::transform(
-    transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+    transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
     [](TensorTransform *op) -> std::shared_ptr<TensorOperation> { return op != nullptr ? op->Parse() : nullptr; });
 }
 
-Compose::Compose(const std::vector<std::shared_ptr<TensorTransform>> &transforms) {
+Compose::Compose(const std::vector<std::shared_ptr<TensorTransform>> &transforms) : data_(std::make_shared<Data>()) {
-  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
                        [](std::shared_ptr<TensorTransform> op) -> std::shared_ptr<TensorOperation> {
                          return op != nullptr ? op->Parse() : nullptr;
                        });
 }
 
-Compose::Compose(const std::vector<std::reference_wrapper<TensorTransform>> &transforms) {
+Compose::Compose(const std::vector<std::reference_wrapper<TensorTransform>> &transforms)
-  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+    : data_(std::make_shared<Data>()) {
+  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
                        [](TensorTransform &op) -> std::shared_ptr<TensorOperation> { return op.Parse(); });
 }
 
-std::shared_ptr<TensorOperation> Compose::Parse() { return std::make_shared<ComposeOperation>(transforms_); }
+std::shared_ptr<TensorOperation> Compose::Parse() { return std::make_shared<ComposeOperation>(data_->transforms_); }
 
 // Constructor to Duplicate
 Duplicate::Duplicate() {}
@@ -56,59 +61,87 @@ Duplicate::Duplicate() {}
 std::shared_ptr<TensorOperation> Duplicate::Parse() { return std::make_shared<DuplicateOperation>(); }
 
 // Constructor to OneHot
-OneHot::OneHot(int32_t num_classes) : num_classes_(num_classes) {}
+struct OneHot::Data {
+  explicit Data(int32_t num_classes) : num_classes_(num_classes) {}
+  float num_classes_;
+};
 
-std::shared_ptr<TensorOperation> OneHot::Parse() { return std::make_shared<OneHotOperation>(num_classes_); }
+OneHot::OneHot(int32_t num_classes) : data_(std::make_shared<Data>(num_classes)) {}
+
+std::shared_ptr<TensorOperation> OneHot::Parse() { return std::make_shared<OneHotOperation>(data_->num_classes_); }
 
 // Constructor to RandomApply.
-RandomApply::RandomApply(const std::vector<TensorTransform *> &transforms, double prob) : prob_(prob) {
+struct RandomApply::Data {
+  std::vector<std::shared_ptr<TensorOperation>> transforms_;
+  double prob_;
+};
+
+RandomApply::RandomApply(const std::vector<TensorTransform *> &transforms, double prob)
+    : data_(std::make_shared<Data>()) {
   (void)std::transform(
-    transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+    transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
     [](TensorTransform *op) -> std::shared_ptr<TensorOperation> { return op != nullptr ? op->Parse() : nullptr; });
+  data_->prob_ = prob;
 }
 
-RandomApply::RandomApply(const std::vector<std::shared_ptr<TensorTransform>> &transforms, double prob) : prob_(prob) {
+RandomApply::RandomApply(const std::vector<std::shared_ptr<TensorTransform>> &transforms, double prob)
-  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+    : data_(std::make_shared<Data>()) {
+  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
                        [](std::shared_ptr<TensorTransform> op) -> std::shared_ptr<TensorOperation> {
                          return op != nullptr ? op->Parse() : nullptr;
                        });
+  data_->prob_ = prob;
 }
 
 RandomApply::RandomApply(const std::vector<std::reference_wrapper<TensorTransform>> &transforms, double prob)
-    : prob_(prob) {
+    : data_(std::make_shared<Data>()) {
-  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
                        [](TensorTransform &op) -> std::shared_ptr<TensorOperation> { return op.Parse(); });
+  data_->prob_ = prob;
 }
 
 std::shared_ptr<TensorOperation> RandomApply::Parse() {
-  return std::make_shared<RandomApplyOperation>(transforms_, prob_);
+  return std::make_shared<RandomApplyOperation>(data_->transforms_, data_->prob_);
 }
 
 // Constructor to RandomChoice.
-RandomChoice::RandomChoice(const std::vector<TensorTransform *> &transforms) {
+struct RandomChoice::Data {
+  std::vector<std::shared_ptr<TensorOperation>> transforms_;
+};
+
+RandomChoice::RandomChoice(const std::vector<TensorTransform *> &transforms) : data_(std::make_shared<Data>()) {
   (void)std::transform(
-    transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+    transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
     [](TensorTransform *op) -> std::shared_ptr<TensorOperation> { return op != nullptr ? op->Parse() : nullptr; });
 }
 
-RandomChoice::RandomChoice(const std::vector<std::shared_ptr<TensorTransform>> &transforms) {
+RandomChoice::RandomChoice(const std::vector<std::shared_ptr<TensorTransform>> &transforms)
-  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+    : data_(std::make_shared<Data>()) {
+  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
                        [](std::shared_ptr<TensorTransform> op) -> std::shared_ptr<TensorOperation> {
                          return op != nullptr ? op->Parse() : nullptr;
                        });
 }
 
-RandomChoice::RandomChoice(const std::vector<std::reference_wrapper<TensorTransform>> &transforms) {
+RandomChoice::RandomChoice(const std::vector<std::reference_wrapper<TensorTransform>> &transforms)
-  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+    : data_(std::make_shared<Data>()) {
+  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
                        [](TensorTransform &op) -> std::shared_ptr<TensorOperation> { return op.Parse(); });
 }
 
-std::shared_ptr<TensorOperation> RandomChoice::Parse() { return std::make_shared<RandomChoiceOperation>(transforms_); }
+std::shared_ptr<TensorOperation> RandomChoice::Parse() {
+  return std::make_shared<RandomChoiceOperation>(data_->transforms_);
+}
 
 // Constructor to TypeCast
-TypeCast::TypeCast(std::string data_type) : data_type_(data_type) {}
+struct TypeCast::Data {
+  explicit Data(const std::vector<char> &data_type) : data_type_(CharToString(data_type)) {}
+  std::string data_type_;
+};
 
-std::shared_ptr<TensorOperation> TypeCast::Parse() { return std::make_shared<TypeCastOperation>(data_type_); }
+TypeCast::TypeCast(const std::vector<char> &data_type) : data_(std::make_shared<Data>(data_type)) {}
+
+std::shared_ptr<TensorOperation> TypeCast::Parse() { return std::make_shared<TypeCastOperation>(data_->data_type_); }
 
 // Constructor to Unique
 Unique::Unique() {}

mindspore/ccsrc/minddata/dataset/api/vision.cc

@@ -42,85 +42,153 @@ namespace vision {
 // CONSTRUCTORS FOR API CLASSES TO CREATE VISION TENSOR TRANSFORM OPERATIONS
 // (In alphabetical order)
 
+// Affine Transform Operation.
+struct Affine::Data {
+  Data(float_t degrees, const std::vector<float> &translation, float scale, const std::vector<float> &shear,
+       InterpolationMode interpolation, const std::vector<uint8_t> &fill_value)
+      : degrees_(degrees),
+        translation_(translation),
+        scale_(scale),
+        shear_(shear),
+        interpolation_(interpolation),
+        fill_value_(fill_value) {}
+  float degrees_;
+  std::vector<float> translation_;
+  float scale_;
+  std::vector<float> shear_;
+  InterpolationMode interpolation_;
+  std::vector<uint8_t> fill_value_;
+};
+
 Affine::Affine(float_t degrees, const std::vector<float> &translation, float scale, const std::vector<float> &shear,
                InterpolationMode interpolation, const std::vector<uint8_t> &fill_value)
-    : degrees_(degrees),
+    : data_(std::make_shared<Data>(degrees, translation, scale, shear, interpolation, fill_value)) {}
-      translation_(translation),
-      scale_(scale),
-      shear_(shear),
-      interpolation_(interpolation),
-      fill_value_(fill_value) {}
 
 std::shared_ptr<TensorOperation> Affine::Parse() {
-  return std::make_shared<AffineOperation>(degrees_, translation_, scale_, shear_, interpolation_, fill_value_);
+  return std::make_shared<AffineOperation>(data_->degrees_, data_->translation_, data_->scale_, data_->shear_,
+                                           data_->interpolation_, data_->fill_value_);
 }
 
 // AutoContrast Transform Operation.
-AutoContrast::AutoContrast(float cutoff, std::vector<uint32_t> ignore) : cutoff_(cutoff), ignore_(ignore) {}
+struct AutoContrast::Data {
+  Data(float cutoff, const std::vector<uint32_t> &ignore) : cutoff_(cutoff), ignore_(ignore) {}
+  float cutoff_;
+  std::vector<uint32_t> ignore_;
+};
+
+AutoContrast::AutoContrast(float cutoff, std::vector<uint32_t> ignore)
+    : data_(std::make_shared<Data>(cutoff, ignore)) {}
 
 std::shared_ptr<TensorOperation> AutoContrast::Parse() {
-  return std::make_shared<AutoContrastOperation>(cutoff_, ignore_);
+  return std::make_shared<AutoContrastOperation>(data_->cutoff_, data_->ignore_);
 }
 
 // BoundingBoxAugment Transform Operation.
-BoundingBoxAugment::BoundingBoxAugment(TensorTransform *transform, float ratio) : ratio_(ratio) {
+struct BoundingBoxAugment::Data {
-  transform_ = transform ? transform->Parse() : nullptr;
+  std::shared_ptr<TensorOperation> transform_;
+  float ratio_;
+};
+
+BoundingBoxAugment::BoundingBoxAugment(TensorTransform *transform, float ratio) : data_(std::make_shared<Data>()) {
+  data_->transform_ = transform ? transform->Parse() : nullptr;
+  data_->ratio_ = ratio;
 }
 
-BoundingBoxAugment::BoundingBoxAugment(const std::shared_ptr<TensorTransform> &transform, float ratio) : ratio_(ratio) {
+BoundingBoxAugment::BoundingBoxAugment(const std::shared_ptr<TensorTransform> &transform, float ratio)
-  transform_ = transform ? transform->Parse() : nullptr;
+    : data_(std::make_shared<Data>()) {
+  data_->transform_ = transform ? transform->Parse() : nullptr;
+  data_->ratio_ = ratio;
 }
 
 BoundingBoxAugment::BoundingBoxAugment(const std::reference_wrapper<TensorTransform> transform, float ratio)
-    : ratio_(ratio) {
+    : data_(std::make_shared<Data>()) {
-  transform_ = transform.get().Parse();
+  data_->transform_ = transform.get().Parse();
+  data_->ratio_ = ratio;
 }
 
 std::shared_ptr<TensorOperation> BoundingBoxAugment::Parse() {
-  return std::make_shared<BoundingBoxAugmentOperation>(transform_, ratio_);
+  return std::make_shared<BoundingBoxAugmentOperation>(data_->transform_, data_->ratio_);
 }
 #endif  // not ENABLE_ANDROID
 
 // CenterCrop Transform Operation.
-CenterCrop::CenterCrop(std::vector<int32_t> size) : size_(size) {}
+struct CenterCrop::Data {
+  explicit Data(const std::vector<int32_t> &size) : size_(size) {}
+  std::vector<int32_t> size_;
+};
 
-std::shared_ptr<TensorOperation> CenterCrop::Parse() { return std::make_shared<CenterCropOperation>(size_); }
+CenterCrop::CenterCrop(std::vector<int32_t> size) : data_(std::make_shared<Data>(size)) {}
+
+std::shared_ptr<TensorOperation> CenterCrop::Parse() { return std::make_shared<CenterCropOperation>(data_->size_); }
 
 std::shared_ptr<TensorOperation> CenterCrop::Parse(const MapTargetDevice &env) {
   if (env == MapTargetDevice::kAscend310) {
 #ifdef ENABLE_ACL
     std::vector<uint32_t> usize_;
-    usize_.reserve(size_.size());
+    usize_.reserve(data_->size_.size());
-    std::transform(size_.begin(), size_.end(), std::back_inserter(usize_), [](int32_t i) { return (uint32_t)i; });
+    std::transform(data_->size_.begin(), data_->size_.end(), std::back_inserter(usize_),
+                   [](int32_t i) { return (uint32_t)i; });
     return std::make_shared<DvppCropJpegOperation>(usize_);
 #endif  // ENABLE_ACL
   }
-  return std::make_shared<CenterCropOperation>(size_);
+  return std::make_shared<CenterCropOperation>(data_->size_);
 }
 
 // Crop Transform Operation.
-Crop::Crop(std::vector<int32_t> coordinates, std::vector<int32_t> size) : coordinates_(coordinates), size_(size) {}
+struct Crop::Data {
+  Data(const std::vector<int32_t> &coordinates, const std::vector<int32_t> &size)
+      : coordinates_(coordinates), size_(size) {}
+  std::vector<int32_t> coordinates_;
+  std::vector<int32_t> size_;
+};
 
-std::shared_ptr<TensorOperation> Crop::Parse() { return std::make_shared<CropOperation>(coordinates_, size_); }
+Crop::Crop(std::vector<int32_t> coordinates, std::vector<int32_t> size)
+    : data_(std::make_shared<Data>(coordinates, size)) {}
+
+std::shared_ptr<TensorOperation> Crop::Parse() {
+  return std::make_shared<CropOperation>(data_->coordinates_, data_->size_);
+}
 
 #ifndef ENABLE_ANDROID
 // CutMixBatch Transform Operation.
+struct CutMixBatch::Data {
+  Data(ImageBatchFormat image_batch_format, float alpha, float prob)
+      : image_batch_format_(image_batch_format), alpha_(alpha), prob_(prob) {}
+  float alpha_;
+  float prob_;
+  ImageBatchFormat image_batch_format_;
+};
+
 CutMixBatch::CutMixBatch(ImageBatchFormat image_batch_format, float alpha, float prob)
-    : image_batch_format_(image_batch_format), alpha_(alpha), prob_(prob) {}
+    : data_(std::make_shared<Data>(image_batch_format, alpha, prob)) {}
 
 std::shared_ptr<TensorOperation> CutMixBatch::Parse() {
-  return std::make_shared<CutMixBatchOperation>(image_batch_format_, alpha_, prob_);
+  return std::make_shared<CutMixBatchOperation>(data_->image_batch_format_, data_->alpha_, data_->prob_);
 }
 
 // CutOutOp.
-CutOut::CutOut(int32_t length, int32_t num_patches) : length_(length), num_patches_(num_patches) {}
+struct CutOut::Data {
+  Data(int32_t length, int32_t num_patches) : length_(length), num_patches_(num_patches) {}
+  int32_t length_;
+  int32_t num_patches_;
+};
 
-std::shared_ptr<TensorOperation> CutOut::Parse() { return std::make_shared<CutOutOperation>(length_, num_patches_); }
+CutOut::CutOut(int32_t length, int32_t num_patches) : data_(std::make_shared<Data>(length, num_patches)) {}
+
+std::shared_ptr<TensorOperation> CutOut::Parse() {
+  return std::make_shared<CutOutOperation>(data_->length_, data_->num_patches_);
+}
 #endif  // not ENABLE_ANDROID
 
 // Decode Transform Operation.
-Decode::Decode(bool rgb) : rgb_(rgb) {}
+struct Decode::Data {
-std::shared_ptr<TensorOperation> Decode::Parse() { return std::make_shared<DecodeOperation>(rgb_); }
+  explicit Data(bool rgb) : rgb_(rgb) {}
+  bool rgb_;
+};
+
+Decode::Decode(bool rgb) : data_(std::make_shared<Data>(rgb)) {}
+
+std::shared_ptr<TensorOperation> Decode::Parse() { return std::make_shared<DecodeOperation>(data_->rgb_); }
 
 std::shared_ptr<TensorOperation> Decode::Parse(const MapTargetDevice &env) {
   if (env == MapTargetDevice::kAscend310) {
@@ -128,31 +196,42 @@ std::shared_ptr<TensorOperation> Decode::Parse(const MapTargetDevice &env) {
     return std::make_shared<DvppDecodeJpegOperation>();
 #endif  // ENABLE_ACL
   }
-  return std::make_shared<DecodeOperation>(rgb_);
+  return std::make_shared<DecodeOperation>(data_->rgb_);
 }
 
 #ifdef ENABLE_ACL
 // DvppDecodeResize Transform Operation.
-DvppDecodeResizeJpeg::DvppDecodeResizeJpeg(std::vector<uint32_t> resize) : resize_(resize) {}
+struct DvppDecodeResizeJpeg::Data {
+  explicit Data(const std::vector<uint32_t> &resize) : resize_(resize) {}
+  std::vector<uint32_t> resize_;
+};
+
+DvppDecodeResizeJpeg::DvppDecodeResizeJpeg(std::vector<uint32_t> resize) : data_(std::make_shared<Data>(resize)) {}
 
 std::shared_ptr<TensorOperation> DvppDecodeResizeJpeg::Parse() {
-  return std::make_shared<DvppDecodeResizeOperation>(resize_);
+  return std::make_shared<DvppDecodeResizeOperation>(data_->resize_);
 }
 
 std::shared_ptr<TensorOperation> DvppDecodeResizeJpeg::Parse(const MapTargetDevice &env) {
-  return std::make_shared<DvppDecodeResizeOperation>(resize_);
+  return std::make_shared<DvppDecodeResizeOperation>(data_->resize_);
 }
 
 // DvppDecodeResizeCrop Transform Operation.
+struct DvppDecodeResizeCropJpeg::Data {
+  Data(const std::vector<uint32_t> &crop, const std::vector<uint32_t> &resize) : crop_(crop), resize_(resize) {}
+  std::vector<uint32_t> crop_;
+  std::vector<uint32_t> resize_;
+};
+
 DvppDecodeResizeCropJpeg::DvppDecodeResizeCropJpeg(std::vector<uint32_t> crop, std::vector<uint32_t> resize)
-    : crop_(crop), resize_(resize) {}
+    : data_(std::make_shared<Data>(crop, resize)) {}
 
 std::shared_ptr<TensorOperation> DvppDecodeResizeCropJpeg::Parse() {
-  return std::make_shared<DvppDecodeResizeCropOperation>(crop_, resize_);
+  return std::make_shared<DvppDecodeResizeCropOperation>(data_->crop_, data_->resize_);
 }
 
 std::shared_ptr<TensorOperation> DvppDecodeResizeCropJpeg::Parse(const MapTargetDevice &env) {
-  return std::make_shared<DvppDecodeResizeCropOperation>(crop_, resize_);
+  return std::make_shared<DvppDecodeResizeCropOperation>(data_->crop_, data_->resize_);
 }
 
 // DvppDecodePng Transform Operation.
@@ -181,174 +260,339 @@ Invert::Invert() {}
 std::shared_ptr<TensorOperation> Invert::Parse() { return std::make_shared<InvertOperation>(); }
 
 // MixUpBatch Transform Operation.
-MixUpBatch::MixUpBatch(float alpha) : alpha_(alpha) {}
+struct MixUpBatch::Data {
+  explicit Data(float alpha) : alpha_(alpha) {}
+  float alpha_;
+};
 
-std::shared_ptr<TensorOperation> MixUpBatch::Parse() { return std::make_shared<MixUpBatchOperation>(alpha_); }
+MixUpBatch::MixUpBatch(float alpha) : data_(std::make_shared<Data>(alpha)) {}
+
+std::shared_ptr<TensorOperation> MixUpBatch::Parse() { return std::make_shared<MixUpBatchOperation>(data_->alpha_); }
 #endif  // not ENABLE_ANDROID
 
 // Normalize Transform Operation.
-Normalize::Normalize(std::vector<float> mean, std::vector<float> std) : mean_(mean), std_(std) {}
+struct Normalize::Data {
+  Data(const std::vector<float> &mean, const std::vector<float> &std) : mean_(mean), std_(std) {}
+  std::vector<float> mean_;
+  std::vector<float> std_;
+};
 
-std::shared_ptr<TensorOperation> Normalize::Parse() { return std::make_shared<NormalizeOperation>(mean_, std_); }
+Normalize::Normalize(std::vector<float> mean, std::vector<float> std) : data_(std::make_shared<Data>(mean, std)) {}
+
+std::shared_ptr<TensorOperation> Normalize::Parse() {
+  return std::make_shared<NormalizeOperation>(data_->mean_, data_->std_);
+}
 
 std::shared_ptr<TensorOperation> Normalize::Parse(const MapTargetDevice &env) {
   if (env == MapTargetDevice::kAscend310) {
 #ifdef ENABLE_ACL
-    return std::make_shared<DvppNormalizeOperation>(mean_, std_);
+    return std::make_shared<DvppNormalizeOperation>(data_->mean_, data_->std_);
 #endif
   }
-  return std::make_shared<NormalizeOperation>(mean_, std_);
+  return std::make_shared<NormalizeOperation>(data_->mean_, data_->std_);
 }
 
 #ifndef ENABLE_ANDROID
 // NormalizePad Transform Operation.
-NormalizePad::NormalizePad(const std::vector<float> &mean, const std::vector<float> &std, const std::string &dtype)
+struct NormalizePad::Data {
-    : mean_(mean), std_(std), dtype_(dtype) {}
+  Data(const std::vector<float> &mean, const std::vector<float> &std, const std::string &dtype)
+      : mean_(mean), std_(std), dtype_(dtype) {}
+  std::vector<float> mean_;
+  std::vector<float> std_;
+  std::string dtype_;
+};
+
+NormalizePad::NormalizePad(const std::vector<float> &mean, const std::vector<float> &std,
+                           const std::vector<char> &dtype)
+    : data_(std::make_shared<Data>(mean, std, CharToString(dtype))) {}
 
 std::shared_ptr<TensorOperation> NormalizePad::Parse() {
-  return std::make_shared<NormalizePadOperation>(mean_, std_, dtype_);
+  return std::make_shared<NormalizePadOperation>(data_->mean_, data_->std_, data_->dtype_);
 }
 
 // Pad Transform Operation.
+struct Pad::Data {
+  Data(const std::vector<int32_t> &padding, const std::vector<uint8_t> &fill_value, BorderType padding_mode)
+      : padding_(padding), fill_value_(fill_value), padding_mode_(padding_mode) {}
+  std::vector<int32_t> padding_;
+  std::vector<uint8_t> fill_value_;
+  BorderType padding_mode_;
+};
+
 Pad::Pad(std::vector<int32_t> padding, std::vector<uint8_t> fill_value, BorderType padding_mode)
-    : padding_(padding), fill_value_(fill_value), padding_mode_(padding_mode) {}
+    : data_(std::make_shared<Data>(padding, fill_value, padding_mode)) {}
 
 std::shared_ptr<TensorOperation> Pad::Parse() {
-  return std::make_shared<PadOperation>(padding_, fill_value_, padding_mode_);
+  return std::make_shared<PadOperation>(data_->padding_, data_->fill_value_, data_->padding_mode_);
 }
 
 // RandomAffine Transform Operation.
+struct RandomAffine::Data {
+  Data(const std::vector<float_t> &degrees, const std::vector<float_t> &translate_range,
+       const std::vector<float_t> &scale_range, const std::vector<float_t> &shear_ranges,
+       InterpolationMode interpolation, const std::vector<uint8_t> &fill_value)
+      : degrees_(degrees),
+        translate_range_(translate_range),
+        scale_range_(scale_range),
+        shear_ranges_(shear_ranges),
+        interpolation_(interpolation),
+        fill_value_(fill_value) {}
+  std::vector<float_t> degrees_;          // min_degree, max_degree
+  std::vector<float_t> translate_range_;  // maximum x translation percentage, maximum y translation percentage
+  std::vector<float_t> scale_range_;      // min_scale, max_scale
+  std::vector<float_t> shear_ranges_;     // min_x_shear, max_x_shear, min_y_shear, max_y_shear
+  InterpolationMode interpolation_;
+  std::vector<uint8_t> fill_value_;
+};
+
 RandomAffine::RandomAffine(const std::vector<float_t> &degrees, const std::vector<float_t> &translate_range,
                            const std::vector<float_t> &scale_range, const std::vector<float_t> &shear_ranges,
                            InterpolationMode interpolation, const std::vector<uint8_t> &fill_value)
-    : degrees_(degrees),
+    : data_(std::make_shared<Data>(degrees, translate_range, scale_range, shear_ranges, interpolation, fill_value)) {}
-      translate_range_(translate_range),
-      scale_range_(scale_range),
-      shear_ranges_(shear_ranges),
-      interpolation_(interpolation),
-      fill_value_(fill_value) {}
 
 std::shared_ptr<TensorOperation> RandomAffine::Parse() {
-  return std::make_shared<RandomAffineOperation>(degrees_, translate_range_, scale_range_, shear_ranges_,
+  return std::make_shared<RandomAffineOperation>(data_->degrees_, data_->translate_range_, data_->scale_range_,
-                                                 interpolation_, fill_value_);
+                                                 data_->shear_ranges_, data_->interpolation_, data_->fill_value_);
 }
 
 // RandomColor Transform Operation.
-RandomColor::RandomColor(float t_lb, float t_ub) : t_lb_(t_lb), t_ub_(t_ub) {}
+struct RandomColor::Data {
+  Data(float t_lb, float t_ub) : t_lb_(t_lb), t_ub_(t_ub) {}
+  float t_lb_;
+  float t_ub_;
+};
 
-std::shared_ptr<TensorOperation> RandomColor::Parse() { return std::make_shared<RandomColorOperation>(t_lb_, t_ub_); }
+RandomColor::RandomColor(float t_lb, float t_ub) : data_(std::make_shared<Data>(t_lb, t_ub)) {}
+
+std::shared_ptr<TensorOperation> RandomColor::Parse() {
+  return std::make_shared<RandomColorOperation>(data_->t_lb_, data_->t_ub_);
+}
 
 // RandomColorAdjust Transform Operation.
+struct RandomColorAdjust::Data {
+  Data(const std::vector<float> &brightness, const std::vector<float> &contrast, const std::vector<float> &saturation,
+       const std::vector<float> &hue)
+      : brightness_(brightness), contrast_(contrast), saturation_(saturation), hue_(hue) {}
+  std::vector<float> brightness_;
+  std::vector<float> contrast_;
+  std::vector<float> saturation_;
+  std::vector<float> hue_;
+};
+
 RandomColorAdjust::RandomColorAdjust(std::vector<float> brightness, std::vector<float> contrast,
                                      std::vector<float> saturation, std::vector<float> hue)
-    : brightness_(brightness), contrast_(contrast), saturation_(saturation), hue_(hue) {}
+    : data_(std::make_shared<Data>(brightness, contrast, saturation, hue)) {}
+
 std::shared_ptr<TensorOperation> RandomColorAdjust::Parse() {
-  return std::make_shared<RandomColorAdjustOperation>(brightness_, contrast_, saturation_, hue_);
+  return std::make_shared<RandomColorAdjustOperation>(data_->brightness_, data_->contrast_, data_->saturation_,
+                                                      data_->hue_);
 }
 
 // RandomCrop Transform Operation.
+struct RandomCrop::Data {
+  Data(const std::vector<int32_t> &size, const std::vector<int32_t> &padding, bool pad_if_needed,
+       const std::vector<uint8_t> &fill_value, BorderType padding_mode)
+      : size_(size),
+        padding_(padding),
+        pad_if_needed_(pad_if_needed),
+        fill_value_(fill_value),
+        padding_mode_(padding_mode) {}
+  std::vector<int32_t> size_;
+  std::vector<int32_t> padding_;
+  bool pad_if_needed_;
+  std::vector<uint8_t> fill_value_;
+  BorderType padding_mode_;
+};
+
 RandomCrop::RandomCrop(std::vector<int32_t> size, std::vector<int32_t> padding, bool pad_if_needed,
                        std::vector<uint8_t> fill_value, BorderType padding_mode)
-    : size_(size),
+    : data_(std::make_shared<Data>(size, padding, pad_if_needed, fill_value, padding_mode)) {}
-      padding_(padding),
-      pad_if_needed_(pad_if_needed),
-      fill_value_(fill_value),
-      padding_mode_(padding_mode) {}
 
 std::shared_ptr<TensorOperation> RandomCrop::Parse() {
-  return std::make_shared<RandomCropOperation>(size_, padding_, pad_if_needed_, fill_value_, padding_mode_);
+  return std::make_shared<RandomCropOperation>(data_->size_, data_->padding_, data_->pad_if_needed_, data_->fill_value_,
+                                               data_->padding_mode_);
 }
 
 // RandomCropDecodeResize Transform Operation.
+struct RandomCropDecodeResize::Data {
+  Data(const std::vector<int32_t> &size, const std::vector<float> &scale, const std::vector<float> &ratio,
+       InterpolationMode interpolation, int32_t max_attempts)
+      : size_(size), scale_(scale), ratio_(ratio), interpolation_(interpolation), max_attempts_(max_attempts) {}
+  std::vector<int32_t> size_;
+  std::vector<float> scale_;
+  std::vector<float> ratio_;
+  InterpolationMode interpolation_;
+  int32_t max_attempts_;
+};
+
 RandomCropDecodeResize::RandomCropDecodeResize(std::vector<int32_t> size, std::vector<float> scale,
                                                std::vector<float> ratio, InterpolationMode interpolation,
                                                int32_t max_attempts)
-    : size_(size), scale_(scale), ratio_(ratio), interpolation_(interpolation), max_attempts_(max_attempts) {}
+    : data_(std::make_shared<Data>(size, scale, ratio, interpolation, max_attempts)) {}
 
 std::shared_ptr<TensorOperation> RandomCropDecodeResize::Parse() {
-  return std::make_shared<RandomCropDecodeResizeOperation>(size_, scale_, ratio_, interpolation_, max_attempts_);
+  return std::make_shared<RandomCropDecodeResizeOperation>(data_->size_, data_->scale_, data_->ratio_,
+                                                           data_->interpolation_, data_->max_attempts_);
 }
 
 // RandomCropWithBBox Transform Operation.
+struct RandomCropWithBBox::Data {
+  Data(const std::vector<int32_t> &size, const std::vector<int32_t> &padding, bool pad_if_needed,
+       const std::vector<uint8_t> &fill_value, BorderType padding_mode)
+      : size_(size),
+        padding_(padding),
+        pad_if_needed_(pad_if_needed),
+        fill_value_(fill_value),
+        padding_mode_(padding_mode) {}
+  std::vector<int32_t> size_;
+  std::vector<int32_t> padding_;
+  bool pad_if_needed_;
+  std::vector<uint8_t> fill_value_;
+  BorderType padding_mode_;
+};
+
 RandomCropWithBBox::RandomCropWithBBox(std::vector<int32_t> size, std::vector<int32_t> padding, bool pad_if_needed,
                                        std::vector<uint8_t> fill_value, BorderType padding_mode)
-    : size_(size),
+    : data_(std::make_shared<Data>(size, padding, pad_if_needed, fill_value, padding_mode)) {}
-      padding_(padding),
-      pad_if_needed_(pad_if_needed),
-      fill_value_(fill_value),
-      padding_mode_(padding_mode) {}
 
 std::shared_ptr<TensorOperation> RandomCropWithBBox::Parse() {
-  return std::make_shared<RandomCropWithBBoxOperation>(size_, padding_, pad_if_needed_, fill_value_, padding_mode_);
+  return std::make_shared<RandomCropWithBBoxOperation>(data_->size_, data_->padding_, data_->pad_if_needed_,
+                                                       data_->fill_value_, data_->padding_mode_);
 }
 
 // RandomHorizontalFlip.
-RandomHorizontalFlip::RandomHorizontalFlip(float prob) : probability_(prob) {}
+struct RandomHorizontalFlip::Data {
+  explicit Data(float prob) : probability_(prob) {}
+  float probability_;
+};
+
+RandomHorizontalFlip::RandomHorizontalFlip(float prob) : data_(std::make_shared<Data>(prob)) {}
 
 std::shared_ptr<TensorOperation> RandomHorizontalFlip::Parse() {
-  return std::make_shared<RandomHorizontalFlipOperation>(probability_);
+  return std::make_shared<RandomHorizontalFlipOperation>(data_->probability_);
 }
 
 // RandomHorizontalFlipWithBBox
-RandomHorizontalFlipWithBBox::RandomHorizontalFlipWithBBox(float prob) : probability_(prob) {}
+struct RandomHorizontalFlipWithBBox::Data {
+  explicit Data(float prob) : probability_(prob) {}
+  float probability_;
+};
+
+RandomHorizontalFlipWithBBox::RandomHorizontalFlipWithBBox(float prob) : data_(std::make_shared<Data>(prob)) {}
 
 std::shared_ptr<TensorOperation> RandomHorizontalFlipWithBBox::Parse() {
-  return std::make_shared<RandomHorizontalFlipWithBBoxOperation>(probability_);
+  return std::make_shared<RandomHorizontalFlipWithBBoxOperation>(data_->probability_);
 }
 
 // RandomPosterize Transform Operation.
-RandomPosterize::RandomPosterize(const std::vector<uint8_t> &bit_range) : bit_range_(bit_range) {}
+struct RandomPosterize::Data {
+  explicit Data(const std::vector<uint8_t> &bit_range) : bit_range_(bit_range) {}
+  std::vector<uint8_t> bit_range_;
+};
+
+RandomPosterize::RandomPosterize(const std::vector<uint8_t> &bit_range) : data_(std::make_shared<Data>(bit_range)) {}
 
 std::shared_ptr<TensorOperation> RandomPosterize::Parse() {
-  return std::make_shared<RandomPosterizeOperation>(bit_range_);
+  return std::make_shared<RandomPosterizeOperation>(data_->bit_range_);
 }
 
 // RandomResize Transform Operation.
-RandomResize::RandomResize(std::vector<int32_t> size) : size_(size) {}
+struct RandomResize::Data {
+  explicit Data(const std::vector<int32_t> &size) : size_(size) {}
+  std::vector<int32_t> size_;
+};
 
-std::shared_ptr<TensorOperation> RandomResize::Parse() { return std::make_shared<RandomResizeOperation>(size_); }
+RandomResize::RandomResize(std::vector<int32_t> size) : data_(std::make_shared<Data>(size)) {}
+
+std::shared_ptr<TensorOperation> RandomResize::Parse() { return std::make_shared<RandomResizeOperation>(data_->size_); }
 
 // RandomResizeWithBBox Transform Operation.
-RandomResizeWithBBox::RandomResizeWithBBox(std::vector<int32_t> size) : size_(size) {}
+struct RandomResizeWithBBox::Data {
+  explicit Data(const std::vector<int32_t> &size) : size_(size) {}
+  std::vector<int32_t> size_;
+};
+
+RandomResizeWithBBox::RandomResizeWithBBox(std::vector<int32_t> size) : data_(std::make_shared<Data>(size)) {}
 
 std::shared_ptr<TensorOperation> RandomResizeWithBBox::Parse() {
-  return std::make_shared<RandomResizeWithBBoxOperation>(size_);
+  return std::make_shared<RandomResizeWithBBoxOperation>(data_->size_);
 }
 
 // RandomResizedCrop Transform Operation.
+struct RandomResizedCrop::Data {
+  Data(const std::vector<int32_t> &size, const std::vector<float> &scale, const std::vector<float> &ratio,
+       InterpolationMode interpolation, int32_t max_attempts)
+      : size_(size), scale_(scale), ratio_(ratio), interpolation_(interpolation), max_attempts_(max_attempts) {}
+  std::vector<int32_t> size_;
+  std::vector<float> scale_;
+  std::vector<float> ratio_;
+  InterpolationMode interpolation_;
+  int32_t max_attempts_;
+};
+
 RandomResizedCrop::RandomResizedCrop(std::vector<int32_t> size, std::vector<float> scale, std::vector<float> ratio,
                                      InterpolationMode interpolation, int32_t max_attempts)
-    : size_(size), scale_(scale), ratio_(ratio), interpolation_(interpolation), max_attempts_(max_attempts) {}
+    : data_(std::make_shared<Data>(size, scale, ratio, interpolation, max_attempts)) {}
 
 std::shared_ptr<TensorOperation> RandomResizedCrop::Parse() {
-  return std::make_shared<RandomResizedCropOperation>(size_, scale_, ratio_, interpolation_, max_attempts_);
+  return std::make_shared<RandomResizedCropOperation>(data_->size_, data_->scale_, data_->ratio_, data_->interpolation_,
+                                                      data_->max_attempts_);
 }
 
 // RandomResizedCrop Transform Operation.
+struct RandomResizedCropWithBBox::Data {
+  Data(const std::vector<int32_t> &size, const std::vector<float> &scale, const std::vector<float> &ratio,
+       InterpolationMode interpolation, int32_t max_attempts)
+      : size_(size), scale_(scale), ratio_(ratio), interpolation_(interpolation), max_attempts_(max_attempts) {}
+  std::vector<int32_t> size_;
+  std::vector<float> scale_;
+  std::vector<float> ratio_;
+  InterpolationMode interpolation_;
+  int32_t max_attempts_;
+};
+
 RandomResizedCropWithBBox::RandomResizedCropWithBBox(std::vector<int32_t> size, std::vector<float> scale,
                                                      std::vector<float> ratio, InterpolationMode interpolation,
                                                      int32_t max_attempts)
-    : size_(size), scale_(scale), ratio_(ratio), interpolation_(interpolation), max_attempts_(max_attempts) {}
+    : data_(std::make_shared<Data>(size, scale, ratio, interpolation, max_attempts)) {}
 
 std::shared_ptr<TensorOperation> RandomResizedCropWithBBox::Parse() {
-  return std::make_shared<RandomResizedCropWithBBoxOperation>(size_, scale_, ratio_, interpolation_, max_attempts_);
+  return std::make_shared<RandomResizedCropWithBBoxOperation>(data_->size_, data_->scale_, data_->ratio_,
+                                                              data_->interpolation_, data_->max_attempts_);
 }
 
 // RandomRotation Transform Operation.
+struct RandomRotation::Data {
+  Data(const std::vector<float> &degrees, InterpolationMode interpolation_mode, bool expand,
+       const std::vector<float> &center, const std::vector<uint8_t> &fill_value)
+      : degrees_(degrees),
+        interpolation_mode_(interpolation_mode),
+        expand_(expand),
+        center_(center),
+        fill_value_(fill_value) {}
+  std::vector<float> degrees_;
+  InterpolationMode interpolation_mode_;
+  std::vector<float> center_;
+  bool expand_;
+  std::vector<uint8_t> fill_value_;
+};
+
 RandomRotation::RandomRotation(std::vector<float> degrees, InterpolationMode interpolation_mode, bool expand,
                                std::vector<float> center, std::vector<uint8_t> fill_value)
-    : degrees_(degrees),
+    : data_(std::make_shared<Data>(degrees, interpolation_mode, expand, center, fill_value)) {}
-      interpolation_mode_(interpolation_mode),
-      expand_(expand),
-      center_(center),
-      fill_value_(fill_value) {}
 
 std::shared_ptr<TensorOperation> RandomRotation::Parse() {
-  return std::make_shared<RandomRotationOperation>(degrees_, interpolation_mode_, expand_, center_, fill_value_);
+  return std::make_shared<RandomRotationOperation>(data_->degrees_, data_->interpolation_mode_, data_->expand_,
+                                                   data_->center_, data_->fill_value_);
 }
 
 // RandomSelectSubpolicy Transform Operation.
-RandomSelectSubpolicy::RandomSelectSubpolicy(std::vector<std::vector<std::pair<TensorTransform *, double>>> policy) {
+struct RandomSelectSubpolicy::Data {
+  std::vector<std::vector<std::pair<std::shared_ptr<TensorOperation>, double>>> policy_;
+};
+
+RandomSelectSubpolicy::RandomSelectSubpolicy(std::vector<std::vector<std::pair<TensorTransform *, double>>> policy)
+    : data_(std::make_shared<Data>()) {
   for (int32_t i = 0; i < policy.size(); i++) {
     std::vector<std::pair<std::shared_ptr<TensorOperation>, double>> subpolicy;
 
@@ -358,12 +602,13 @@ RandomSelectSubpolicy::RandomSelectSubpolicy(std::vector<std::vector<std::pair<T
       double prob = policy[i][j].second;
       subpolicy.emplace_back(std::move(std::make_pair(operation, prob)));
     }
-    policy_.emplace_back(subpolicy);
+    data_->policy_.emplace_back(subpolicy);
   }
 }
 
 RandomSelectSubpolicy::RandomSelectSubpolicy(
-  std::vector<std::vector<std::pair<std::shared_ptr<TensorTransform>, double>>> policy) {
+  std::vector<std::vector<std::pair<std::shared_ptr<TensorTransform>, double>>> policy)
+    : data_(std::make_shared<Data>()) {
   for (int32_t i = 0; i < policy.size(); i++) {
     std::vector<std::pair<std::shared_ptr<TensorOperation>, double>> subpolicy;
 
@@ -373,12 +618,13 @@ RandomSelectSubpolicy::RandomSelectSubpolicy(
       double prob = policy[i][j].second;
       subpolicy.emplace_back(std::move(std::make_pair(operation, prob)));
     }
-    policy_.emplace_back(subpolicy);
+    data_->policy_.emplace_back(subpolicy);
   }
 }
 
 RandomSelectSubpolicy::RandomSelectSubpolicy(
-  std::vector<std::vector<std::pair<std::reference_wrapper<TensorTransform>, double>>> policy) {
+  std::vector<std::vector<std::pair<std::reference_wrapper<TensorTransform>, double>>> policy)
+    : data_(std::make_shared<Data>()) {
   for (int32_t i = 0; i < policy.size(); i++) {
     std::vector<std::pair<std::shared_ptr<TensorOperation>, double>> subpolicy;
 
@@ -388,64 +634,102 @@ RandomSelectSubpolicy::RandomSelectSubpolicy(
       double prob = policy[i][j].second;
       subpolicy.emplace_back(std::move(std::make_pair(operation, prob)));
     }
-    policy_.emplace_back(subpolicy);
+    data_->policy_.emplace_back(subpolicy);
   }
 }
 
 std::shared_ptr<TensorOperation> RandomSelectSubpolicy::Parse() {
-  return std::make_shared<RandomSelectSubpolicyOperation>(policy_);
+  return std::make_shared<RandomSelectSubpolicyOperation>(data_->policy_);
 }
 
 // RandomSharpness Transform Operation.
-RandomSharpness::RandomSharpness(std::vector<float> degrees) : degrees_(degrees) {}
+struct RandomSharpness::Data {
+  explicit Data(const std::vector<float> &degrees) : degrees_(degrees) {}
+  std::vector<float> degrees_;
+};
+
+RandomSharpness::RandomSharpness(std::vector<float> degrees) : data_(std::make_shared<Data>(degrees)) {}
 
 std::shared_ptr<TensorOperation> RandomSharpness::Parse() {
-  return std::make_shared<RandomSharpnessOperation>(degrees_);
+  return std::make_shared<RandomSharpnessOperation>(data_->degrees_);
 }
 
 // RandomSolarize Transform Operation.
-RandomSolarize::RandomSolarize(std::vector<uint8_t> threshold) : threshold_(threshold) {}
+struct RandomSolarize::Data {
+  explicit Data(const std::vector<uint8_t> &threshold) : threshold_(threshold) {}
+  std::vector<uint8_t> threshold_;
+};
+
+RandomSolarize::RandomSolarize(std::vector<uint8_t> threshold) : data_(std::make_shared<Data>(threshold)) {}
 
 std::shared_ptr<TensorOperation> RandomSolarize::Parse() {
-  return std::make_shared<RandomSolarizeOperation>(threshold_);
+  return std::make_shared<RandomSolarizeOperation>(data_->threshold_);
 }
 
 // RandomVerticalFlip Transform Operation.
-RandomVerticalFlip::RandomVerticalFlip(float prob) : probability_(prob) {}
+struct RandomVerticalFlip::Data {
+  explicit Data(float prob) : probability_(prob) {}
+  float probability_;
+};
+
+RandomVerticalFlip::RandomVerticalFlip(float prob) : data_(std::make_shared<Data>(prob)) {}
 
 std::shared_ptr<TensorOperation> RandomVerticalFlip::Parse() {
-  return std::make_shared<RandomVerticalFlipOperation>(probability_);
+  return std::make_shared<RandomVerticalFlipOperation>(data_->probability_);
 }
 
 // RandomVerticalFlipWithBBox Transform Operation.
-RandomVerticalFlipWithBBox::RandomVerticalFlipWithBBox(float prob) : probability_(prob) {}
+struct RandomVerticalFlipWithBBox::Data {
+  explicit Data(float prob) : probability_(prob) {}
+  float probability_;
+};
+
+RandomVerticalFlipWithBBox::RandomVerticalFlipWithBBox(float prob) : data_(std::make_shared<Data>(prob)) {}
 
 std::shared_ptr<TensorOperation> RandomVerticalFlipWithBBox::Parse() {
-  return std::make_shared<RandomVerticalFlipWithBBoxOperation>(probability_);
+  return std::make_shared<RandomVerticalFlipWithBBoxOperation>(data_->probability_);
 }
 
 // Rescale Transform Operation.
-Rescale::Rescale(float rescale, float shift) : rescale_(rescale), shift_(shift) {}
+struct Rescale::Data {
+  Data(float rescale, float shift) : rescale_(rescale), shift_(shift) {}
+  float rescale_;
+  float shift_;
+};
 
-std::shared_ptr<TensorOperation> Rescale::Parse() { return std::make_shared<RescaleOperation>(rescale_, shift_); }
+Rescale::Rescale(float rescale, float shift) : data_(std::make_shared<Data>(rescale, shift)) {}
+
+std::shared_ptr<TensorOperation> Rescale::Parse() {
+  return std::make_shared<RescaleOperation>(data_->rescale_, data_->shift_);
+}
 #endif  // not ENABLE_ANDROID
 
 // Resize Transform Operation.
-Resize::Resize(std::vector<int32_t> size, InterpolationMode interpolation)
+struct Resize::Data {
-    : size_(size), interpolation_(interpolation) {}
+  Data(const std::vector<int32_t> &size, InterpolationMode interpolation)
+      : size_(size), interpolation_(interpolation) {}
+  std::vector<int32_t> size_;
+  InterpolationMode interpolation_;
+};
 
-std::shared_ptr<TensorOperation> Resize::Parse() { return std::make_shared<ResizeOperation>(size_, interpolation_); }
+Resize::Resize(std::vector<int32_t> size, InterpolationMode interpolation)
+    : data_(std::make_shared<Data>(size, interpolation)) {}
+
+std::shared_ptr<TensorOperation> Resize::Parse() {
+  return std::make_shared<ResizeOperation>(data_->size_, data_->interpolation_);
+}
 
 std::shared_ptr<TensorOperation> Resize::Parse(const MapTargetDevice &env) {
   if (env == MapTargetDevice::kAscend310) {
 #ifdef ENABLE_ACL
     std::vector<uint32_t> usize_;
-    usize_.reserve(size_.size());
+    usize_.reserve(data_->size_.size());
-    std::transform(size_.begin(), size_.end(), std::back_inserter(usize_), [](int32_t i) { return (uint32_t)i; });
+    std::transform(data_->size_.begin(), data_->size_.end(), std::back_inserter(usize_),
+                   [](int32_t i) { return (uint32_t)i; });
     return std::make_shared<DvppResizeJpegOperation>(usize_);
 #endif  // ENABLE_ACL
   }
-  return std::make_shared<ResizeOperation>(size_, interpolation_);
+  return std::make_shared<ResizeOperation>(data_->size_, data_->interpolation_);
 }
 
 #ifdef ENABLE_ANDROID
@@ -457,11 +741,18 @@ std::shared_ptr<TensorOperation> Rotate::Parse() { return std::make_shared<Rotat
 
 #ifndef ENABLE_ANDROID
 // ResizeWithBBox Transform Operation.
+struct ResizeWithBBox::Data {
+  Data(const std::vector<int32_t> &size, InterpolationMode interpolation)
+      : size_(size), interpolation_(interpolation) {}
+  std::vector<int32_t> size_;
+  InterpolationMode interpolation_;
+};
+
 ResizeWithBBox::ResizeWithBBox(std::vector<int32_t> size, InterpolationMode interpolation)
-    : size_(size), interpolation_(interpolation) {}
+    : data_(std::make_shared<Data>(size, interpolation)) {}
 
 std::shared_ptr<TensorOperation> ResizeWithBBox::Parse() {
-  return std::make_shared<ResizeWithBBoxOperation>(size_, interpolation_);
+  return std::make_shared<ResizeWithBBoxOperation>(data_->size_, data_->interpolation_);
 }
 
 // RgbaToBgr Transform Operation.
@@ -475,19 +766,36 @@ RGBA2RGB::RGBA2RGB() {}
 std::shared_ptr<TensorOperation> RGBA2RGB::Parse() { return std::make_shared<RgbaToRgbOperation>(); }
 
 // SoftDvppDecodeRandomCropResizeJpeg Transform Operation.
+struct SoftDvppDecodeRandomCropResizeJpeg::Data {
+  Data(const std::vector<int32_t> &size, const std::vector<float> &scale, const std::vector<float> &ratio,
+       int32_t max_attempts)
+      : size_(size), scale_(scale), ratio_(ratio), max_attempts_(max_attempts) {}
+  std::vector<int32_t> size_;
+  std::vector<float> scale_;
+  std::vector<float> ratio_;
+  int32_t max_attempts_;
+};
+
 SoftDvppDecodeRandomCropResizeJpeg::SoftDvppDecodeRandomCropResizeJpeg(std::vector<int32_t> size,
                                                                        std::vector<float> scale,
                                                                        std::vector<float> ratio, int32_t max_attempts)
-    : size_(size), scale_(scale), ratio_(ratio), max_attempts_(max_attempts) {}
+    : data_(std::make_shared<Data>(size, scale, ratio, max_attempts)) {}
+
 std::shared_ptr<TensorOperation> SoftDvppDecodeRandomCropResizeJpeg::Parse() {
-  return std::make_shared<SoftDvppDecodeRandomCropResizeJpegOperation>(size_, scale_, ratio_, max_attempts_);
+  return std::make_shared<SoftDvppDecodeRandomCropResizeJpegOperation>(data_->size_, data_->scale_, data_->ratio_,
+                                                                       data_->max_attempts_);
 }
 
 // SoftDvppDecodeResizeJpeg Transform Operation.
-SoftDvppDecodeResizeJpeg::SoftDvppDecodeResizeJpeg(std::vector<int32_t> size) : size_(size) {}
+struct SoftDvppDecodeResizeJpeg::Data {
+  explicit Data(const std::vector<int32_t> &size) : size_(size) {}
+  std::vector<int32_t> size_;
+};
+
+SoftDvppDecodeResizeJpeg::SoftDvppDecodeResizeJpeg(std::vector<int32_t> size) : data_(std::make_shared<Data>(size)) {}
 
 std::shared_ptr<TensorOperation> SoftDvppDecodeResizeJpeg::Parse() {
-  return std::make_shared<SoftDvppDecodeResizeJpegOperation>(size_);
+  return std::make_shared<SoftDvppDecodeResizeJpegOperation>(data_->size_);
 }
 
 // SwapRedBlue Transform Operation.
@@ -496,27 +804,36 @@ SwapRedBlue::SwapRedBlue() {}
 std::shared_ptr<TensorOperation> SwapRedBlue::Parse() { return std::make_shared<SwapRedBlueOperation>(); }
 
 // UniformAug Transform Operation.
-UniformAugment::UniformAugment(const std::vector<TensorTransform *> &transforms, int32_t num_ops) : num_ops_(num_ops) {
+struct UniformAugment::Data {
+  std::vector<std::shared_ptr<TensorOperation>> transforms_;
+  int32_t num_ops_;
+};
+
+UniformAugment::UniformAugment(const std::vector<TensorTransform *> &transforms, int32_t num_ops)
+    : data_(std::make_shared<Data>()) {
   (void)std::transform(
-    transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+    transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
     [](TensorTransform *op) -> std::shared_ptr<TensorOperation> { return op ? op->Parse() : nullptr; });
+  data_->num_ops_ = num_ops;
 }
 
 UniformAugment::UniformAugment(const std::vector<std::shared_ptr<TensorTransform>> &transforms, int32_t num_ops)
-    : num_ops_(num_ops) {
+    : data_(std::make_shared<Data>()) {
   (void)std::transform(
-    transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+    transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
     [](std::shared_ptr<TensorTransform> op) -> std::shared_ptr<TensorOperation> { return op ? op->Parse() : nullptr; });
+  data_->num_ops_ = num_ops;
 }
 
 UniformAugment::UniformAugment(const std::vector<std::reference_wrapper<TensorTransform>> &transforms, int32_t num_ops)
-    : num_ops_(num_ops) {
+    : data_(std::make_shared<Data>()) {
-  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(transforms_),
+  (void)std::transform(transforms.begin(), transforms.end(), std::back_inserter(data_->transforms_),
                        [](TensorTransform &op) -> std::shared_ptr<TensorOperation> { return op.Parse(); });
+  data_->num_ops_ = num_ops;
 }
 
 std::shared_ptr<TensorOperation> UniformAugment::Parse() {
-  return std::make_shared<UniformAugOperation>(transforms_, num_ops_);
+  return std::make_shared<UniformAugOperation>(data_->transforms_, data_->num_ops_);
 }
 #endif  // not ENABLE_ANDROID
 

mindspore/ccsrc/minddata/dataset/engine/ir/cache/dataset_cache_impl.h

@@ -20,6 +20,8 @@
 #include <string>
 #include <optional>
 #include <utility>
+#include <vector>
+#include "include/api/dual_abi_helper.h"
 #include "minddata/dataset/engine/cache/cache_client.h"
 #include "minddata/dataset/engine/datasetops/cache_op.h"
 #include "minddata/dataset/engine/ir/cache/dataset_cache.h"
@@ -39,13 +41,13 @@ class DatasetCacheImpl : public DatasetCache {
   /// \param port optional port (default=50052).
   /// \param num_connections optional number of connections (default=12).
   /// \param prefetch_sz optional prefetch size (default=20).
-  DatasetCacheImpl(session_id_type id, uint64_t mem_sz, bool spill, std::optional<std::string> hostname,
+  DatasetCacheImpl(session_id_type id, uint64_t mem_sz, bool spill, std::optional<std::vector<char>> hostname,
                    std::optional<int32_t> port, std::optional<int32_t> num_connections,
                    std::optional<int32_t> prefetch_sz)
       : session_id_(id),
         cache_mem_sz_(mem_sz),
         spill_(spill),
-        hostname_(std::move(hostname)),
+        hostname_(OptionalCharToString(hostname)),
         port_(std::move(port)),
         num_connections_(std::move(num_connections)),
         prefetch_sz_(std::move(prefetch_sz)) {}

mindspore/ccsrc/minddata/dataset/include/config.h

@@ -19,6 +19,8 @@
 
 #include <cstdint>
 #include <string>
+#include <vector>
+#include "include/api/dual_abi_helper.h"
 
 namespace mindspore {
 namespace dataset {
@@ -70,7 +72,12 @@ int32_t get_callback_timeout();
 
 /// \brief Function to load configuration from a file.
 /// \param[in] file path of the configuration file to be loaded.
-bool load(std::string file);
+/// \note This api exists because std::string will constrained by ABI compile macro but char don't.
+bool load(const std::vector<char> &file);
+
+/// \brief Function to load configuration from a file.
+/// \param[in] file path of the configuration file to be loaded.
+inline bool load(std::string file) { return load(StringToChar(file)); }
 
 }  // namespace config
 }  // namespace dataset

mindspore/ccsrc/minddata/dataset/include/iterator.h

@@ -17,10 +17,12 @@
 #ifndef MINDSPORE_CCSRC_MINDDATA_DATASET_INCLUDE_ITERATOR_H_
 #define MINDSPORE_CCSRC_MINDDATA_DATASET_INCLUDE_ITERATOR_H_
 
+#include <map>
 #include <memory>
 #include <string>
 #include <unordered_map>
 #include <vector>
+#include "include/api/dual_abi_helper.h"
 #include "include/api/status.h"
 #include "include/api/types.h"
 
@@ -39,6 +41,7 @@ class IteratorConsumer;
 class Dataset;
 
 using MSTensorMap = std::unordered_map<std::string, mindspore::MSTensor>;
+using MSTensorMapChar = std::map<std::vector<char>, mindspore::MSTensor>;
 using MSTensorVec = std::vector<mindspore::MSTensor>;
 
 // Abstract class for iterating over the dataset.
@@ -60,7 +63,18 @@ class Iterator {
   /// \note Type of return data is a map(with column name).
   /// \param[out] row - the output tensor row.
   /// \return - a Status error code, returns OK if no error encountered.
-  Status GetNextRow(MSTensorMap *row);
+  Status GetNextRow(MSTensorMap *row) {
+    MSTensorMapChar row_;
+    row_.clear();
+    row->clear();
+    Status s = GetNextRowCharIF(&row_);
+    TensorMapCharToString(&row_, row);
+    return s;
+  }
+
+  // Char interface(CharIF) of GetNextRow
+  // This api exists because std::string will constrained by ABI compile macro but char don't.
+  Status GetNextRowCharIF(MSTensorMapChar *row);
 
   /// \brief Function to get the next row from the data pipeline.
   /// \note Type of return data is a vector(without column name).

mindspore/ccsrc/minddata/dataset/include/text.h

@@ -23,6 +23,7 @@
 #include <utility>
 #include <vector>
 
+#include "include/api/dual_abi_helper.h"
 #include "include/api/status.h"
 #include "minddata/dataset/include/constants.h"
 #include "minddata/dataset/include/transforms.h"
@@ -64,11 +65,8 @@ class BasicTokenizer : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  bool lower_case_;
+  struct Data;
-  bool keep_whitespace_;
+  std::shared_ptr<Data> data_;
-  NormalizeForm normalize_form_;
-  bool preserve_unused_token_;
-  bool with_offsets_;
 };
 
 /// \brief Tokenizer used for Bert text process.
@@ -94,7 +92,14 @@ class BertTokenizer : public TensorTransform {
                          int32_t max_bytes_per_token = 100, const std::string &unknown_token = "[UNK]",
                          bool lower_case = false, bool keep_whitespace = false,
                          const NormalizeForm normalize_form = NormalizeForm::kNone, bool preserve_unused_token = true,
-                         bool with_offsets = false);
+                         bool with_offsets = false)
+      : BertTokenizer(vocab, StringToChar(suffix_indicator), max_bytes_per_token, StringToChar(unknown_token),
+                      lower_case, keep_whitespace, normalize_form, preserve_unused_token, with_offsets) {}
+
+  explicit BertTokenizer(const std::shared_ptr<Vocab> &vocab, const std::vector<char> &suffix_indicator,
+                         int32_t max_bytes_per_token, const std::vector<char> &unknown_token, bool lower_case,
+                         bool keep_whitespace, const NormalizeForm normalize_form, bool preserve_unused_token,
+                         bool with_offsets);
 
   /// \brief Destructor
   ~BertTokenizer() = default;
@@ -104,15 +109,8 @@ class BertTokenizer : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::shared_ptr<Vocab> vocab_;
+  struct Data;
-  std::string suffix_indicator_;
+  std::shared_ptr<Data> data_;
-  int32_t max_bytes_per_token_;
-  std::string unknown_token_;
-  bool lower_case_;
-  bool keep_whitespace_;
-  NormalizeForm normalize_form_;
-  bool preserve_unused_token_;
-  bool with_offsets_;
 };
 
 /// \brief Apply case fold operation on UTF-8 string tensor.
@@ -146,7 +144,11 @@ class JiebaTokenizer : public TensorTransform {
   ///   - JiebaMode.kMIX, tokenize with a mix of MPSegment and HMMSegment algorithm.
   /// \param[in] with_offsets If or not output offsets of tokens (default=false).
   explicit JiebaTokenizer(const std::string &hmm_path, const std::string &mp_path,
-                          const JiebaMode &mode = JiebaMode::kMix, bool with_offsets = false);
+                          const JiebaMode &mode = JiebaMode::kMix, bool with_offsets = false)
+      : JiebaTokenizer(StringToChar(hmm_path), StringToChar(mp_path), mode, with_offsets) {}
+
+  explicit JiebaTokenizer(const std::vector<char> &hmm_path, const std::vector<char> &mp_path, const JiebaMode &mode,
+                          bool with_offsets);
 
   /// \brief Destructor
   ~JiebaTokenizer() = default;
@@ -158,11 +160,8 @@ class JiebaTokenizer : public TensorTransform {
   Status AddWord(const std::string &word, int64_t freq = 0);
 
  private:
-  std::string hmm_path_;
+  struct Data;
-  std::string mp_path_;
+  std::shared_ptr<Data> data_;
-  JiebaMode mode_;
-  bool with_offsets_;
-  std::vector<std::pair<std::string, int64_t>> words_list_;
 };
 
 /// \brief Look up a word into an id according to the input vocabulary table.
@@ -175,7 +174,11 @@ class Lookup : public TensorTransform {
   ///    specify unknown_token when word being out of Vocabulary (default={}).
   /// \param[in] data_type type of the tensor after lookup, typically int32.
   explicit Lookup(const std::shared_ptr<Vocab> &vocab, const std::optional<std::string> &unknown_token = {},
-                  const std::string &data_type = "int32");
+                  const std::string &data_type = "int32")
+      : Lookup(vocab, OptionalStringToChar(unknown_token), StringToChar(data_type)) {}
+
+  explicit Lookup(const std::shared_ptr<Vocab> &vocab, const std::optional<std::vector<char>> &unknown_token,
+                  const std::vector<char> &data_type);
 
   /// \brief Destructor
   ~Lookup() = default;
@@ -185,9 +188,8 @@ class Lookup : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::shared_ptr<Vocab> vocab_;
+  struct Data;
-  std::optional<std::string> unknown_token_;
+  std::shared_ptr<Data> data_;
-  std::string data_type_;
 };
 
 /// \brief TensorOp to generate n-gram from a 1-D string Tensor.
@@ -203,7 +205,11 @@ class Ngram : public TensorTransform {
   ///   be capped at n-1. right_pad=("-":2) would pad right side of the sequence with "--" (default={"", 0}}).
   /// \param[in] separator Symbol used to join strings together (default=" ").
   explicit Ngram(const std::vector<int32_t> &ngrams, const std::pair<std::string, int32_t> &left_pad = {"", 0},
-                 const std::pair<std::string, int32_t> &right_pad = {"", 0}, const std::string &separator = " ");
+                 const std::pair<std::string, int32_t> &right_pad = {"", 0}, const std::string &separator = " ")
+      : Ngram(ngrams, PairStringToChar(left_pad), PairStringToChar(right_pad), StringToChar(separator)) {}
+
+  explicit Ngram(const std::vector<int32_t> &ngrams, const std::pair<std::vector<char>, int32_t> &left_pad,
+                 const std::pair<std::vector<char>, int32_t> &right_pad, const std::vector<char> &separator);
 
   /// \brief Destructor
   ~Ngram() = default;
@@ -213,10 +219,8 @@ class Ngram : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> ngrams_;
+  struct Data;
-  std::pair<std::string, int32_t> left_pad_;
+  std::shared_ptr<Data> data_;
-  std::pair<std::string, int32_t> right_pad_;
-  std::string separator_;
 };
 
 #ifndef _WIN32
@@ -243,7 +247,8 @@ class NormalizeUTF8 : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  NormalizeForm normalize_form_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Replace UTF-8 string tensor with 'replace' according to regular expression 'pattern'.
@@ -254,7 +259,10 @@ class RegexReplace : public TensorTransform {
   /// \param[in] replace The string to replace matched element.
   /// \param[in] replace_all Confirm whether to replace all. If false, only replace first matched element;
   ///   if true, replace all matched elements (default=true).
-  explicit RegexReplace(std::string pattern, std::string replace, bool replace_all = true);
+  explicit RegexReplace(std::string pattern, std::string replace, bool replace_all = true)
+      : RegexReplace(StringToChar(pattern), StringToChar(replace), replace_all) {}
+
+  explicit RegexReplace(const std::vector<char> &pattern, const std::vector<char> &replace, bool replace_all);
 
   /// \brief Destructor
   ~RegexReplace() = default;
@@ -264,9 +272,8 @@ class RegexReplace : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::string pattern_;
+  struct Data;
-  std::string replace_;
+  std::shared_ptr<Data> data_;
-  bool replace_all_;
 };
 
 /// \brief Tokenize a scalar tensor of UTF-8 string by regex expression pattern.
@@ -278,7 +285,11 @@ class RegexTokenizer : public TensorTransform {
   ///   matched by 'keep_delim_pattern'. The default value is an empty string ("")
   ///   which means that delimiters will not be kept as an output token (default="").
   /// \param[in] with_offsets If or not output offsets of tokens (default=false).
-  explicit RegexTokenizer(std::string delim_pattern, std::string keep_delim_pattern = "", bool with_offsets = false);
+  explicit RegexTokenizer(std::string delim_pattern, std::string keep_delim_pattern = "", bool with_offsets = false)
+      : RegexTokenizer(StringToChar(delim_pattern), StringToChar(keep_delim_pattern), with_offsets) {}
+
+  explicit RegexTokenizer(const std::vector<char> &delim_pattern, const std::vector<char> &keep_delim_pattern,
+                          bool with_offsets);
 
   /// \brief Destructor
   ~RegexTokenizer() = default;
@@ -288,9 +299,8 @@ class RegexTokenizer : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::string delim_pattern_;
+  struct Data;
-  std::string keep_delim_pattern_;
+  std::shared_ptr<Data> data_;
-  bool with_offsets_;
 };
 #endif
 
@@ -306,7 +316,10 @@ class SentencePieceTokenizer : public TensorTransform {
   /// \brief Constructor.
   /// \param[in] vocab_path vocab model file path.
   /// \param[in] out_type The type of output.
-  SentencePieceTokenizer(const std::string &vocab_path, mindspore::dataset::SPieceTokenizerOutType out_type);
+  SentencePieceTokenizer(const std::string &vocab_path, mindspore::dataset::SPieceTokenizerOutType out_type)
+      : SentencePieceTokenizer(StringToChar(vocab_path), out_type) {}
+
+  SentencePieceTokenizer(const std::vector<char> &vocab_path, mindspore::dataset::SPieceTokenizerOutType out_type);
 
   /// \brief Destructor
   ~SentencePieceTokenizer() = default;
@@ -316,10 +329,8 @@ class SentencePieceTokenizer : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::shared_ptr<SentencePieceVocab> vocab_;
+  struct Data;
-  std::string vocab_path_;
+  std::shared_ptr<Data> data_;
-  SPieceTokenizerLoadType load_type_;
-  SPieceTokenizerOutType out_type_;
 };
 
 /// \brief TensorOp to construct a tensor from data (only 1-D for now), where each element in the dimension
@@ -340,8 +351,8 @@ class SlidingWindow : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  int32_t width_;
+  struct Data;
-  int32_t axis_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Tensor operation to convert every element of a string tensor to a number.
@@ -353,7 +364,9 @@ class ToNumber : public TensorTransform {
  public:
   /// \brief Constructor.
   /// \param[in] data_type of the tensor to be casted to. Must be a numeric type.
-  explicit ToNumber(const std::string &data_type);
+  explicit ToNumber(const std::string &data_type) : ToNumber(StringToChar(data_type)) {}
+
+  explicit ToNumber(const std::vector<char> &data_type);
 
   /// \brief Destructor
   ~ToNumber() = default;
@@ -363,7 +376,8 @@ class ToNumber : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::string data_type_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Truncate a pair of rank-1 tensors such that the total length is less than max_length.
@@ -381,7 +395,8 @@ class TruncateSequencePair : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  int32_t max_length_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Tokenize a scalar tensor of UTF-8 string to Unicode characters.
@@ -399,7 +414,8 @@ class UnicodeCharTokenizer : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  bool with_offsets_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 #ifndef _WIN32
@@ -419,8 +435,8 @@ class UnicodeScriptTokenizer : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  bool keep_whitespace_;
+  struct Data;
-  bool with_offsets_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Tokenize a scalar tensor of UTF-8 string on ICU4C defined whitespaces.
@@ -438,7 +454,8 @@ class WhitespaceTokenizer : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  bool with_offsets_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 #endif
 }  // namespace text

mindspore/ccsrc/minddata/dataset/include/transforms.h

@@ -22,6 +22,7 @@
 #include <string>
 #include <vector>
 
+#include "include/api/dual_abi_helper.h"
 #include "include/api/status.h"
 #include "minddata/dataset/include/constants.h"
 
@@ -72,7 +73,8 @@ class Compose : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<std::shared_ptr<TensorOperation>> transforms_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Duplicate Op.
@@ -107,7 +109,8 @@ class OneHot : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float num_classes_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomApply Op.
@@ -129,8 +132,8 @@ class RandomApply : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<std::shared_ptr<TensorOperation>> transforms_;
+  struct Data;
-  double prob_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomChoice Op.
@@ -151,7 +154,8 @@ class RandomChoice : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<std::shared_ptr<TensorOperation>> transforms_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief TypeCast Op.
@@ -160,7 +164,9 @@ class TypeCast : public TensorTransform {
  public:
   /// \brief Constructor.
   /// \param[in] data_type mindspore.dtype to be cast to.
-  explicit TypeCast(std::string data_type);
+  explicit TypeCast(std::string data_type) : TypeCast(StringToChar(data_type)) {}
+
+  explicit TypeCast(const std::vector<char> &data_type);
 
   /// \brief Destructor
   ~TypeCast() = default;
@@ -170,7 +176,8 @@ class TypeCast : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::string data_type_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Unique Op.

mindspore/ccsrc/minddata/dataset/include/vision.h

@@ -23,6 +23,7 @@
 #include <utility>
 #include <vector>
 
+#include "include/api/dual_abi_helper.h"
 #include "include/api/status.h"
 #include "minddata/dataset/include/constants.h"
 #include "minddata/dataset/include/transforms.h"
@@ -52,8 +53,8 @@ class AutoContrast : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float cutoff_;
+  struct Data;
-  std::vector<uint32_t> ignore_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief BoundingBoxAugment TensorTransform.
@@ -83,8 +84,8 @@ class BoundingBoxAugment : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::shared_ptr<TensorOperation> transform_;
+  struct Data;
-  float ratio_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Constructor to apply CutMix on a batch of images
@@ -106,9 +107,8 @@ class CutMixBatch : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float alpha_;
+  struct Data;
-  float prob_;
+  std::shared_ptr<Data> data_;
-  ImageBatchFormat image_batch_format_;
 };
 
 /// \brief CutOut TensorOp
@@ -128,8 +128,8 @@ class CutOut : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  int32_t length_;
+  struct Data;
-  int32_t num_patches_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Equalize TensorTransform.
@@ -194,7 +194,8 @@ class MixUpBatch : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float alpha_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief NormalizePad TensorTransform.
@@ -210,7 +211,10 @@ class NormalizePad : public TensorTransform {
   /// \param[in] dtype The output datatype of Tensor.
   ///     The standard deviation values must be "float32" or "float16"（default = "float32"）
   explicit NormalizePad(const std::vector<float> &mean, const std::vector<float> &std,
-                        const std::string &dtype = "float32");
+                        const std::string &dtype = "float32")
+      : NormalizePad(mean, std, StringToChar(dtype)) {}
+
+  explicit NormalizePad(const std::vector<float> &mean, const std::vector<float> &std, const std::vector<char> &dtype);
 
   /// \brief Destructor.
   ~NormalizePad() = default;
@@ -220,9 +224,8 @@ class NormalizePad : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<float> mean_;
+  struct Data;
-  std::vector<float> std_;
+  std::shared_ptr<Data> data_;
-  std::string dtype_;
 };
 
 /// \brief Pad TensorOp
@@ -257,9 +260,8 @@ class Pad : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> padding_;
+  struct Data;
-  std::vector<uint8_t> fill_value_;
+  std::shared_ptr<Data> data_;
-  BorderType padding_mode_;
 };
 
 /// \brief Blends an image with its grayscale version with random weights
@@ -280,8 +282,8 @@ class RandomColor : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float t_lb_;
+  struct Data;
-  float t_ub_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomColorAdjust TensorTransform.
@@ -309,10 +311,8 @@ class RandomColorAdjust : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<float> brightness_;
+  struct Data;
-  std::vector<float> contrast_;
+  std::shared_ptr<Data> data_;
-  std::vector<float> saturation_;
-  std::vector<float> hue_;
 };
 
 /// \brief RandomCrop TensorTransform.
@@ -346,11 +346,8 @@ class RandomCrop : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
-  std::vector<int32_t> padding_;
+  std::shared_ptr<Data> data_;
-  bool pad_if_needed_;
-  std::vector<uint8_t> fill_value_;
-  BorderType padding_mode_;
 };
 
 /// \brief RandomCropDecodeResize TensorTransform.
@@ -381,11 +378,8 @@ class RandomCropDecodeResize : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
-  std::vector<float> scale_;
+  std::shared_ptr<Data> data_;
-  std::vector<float> ratio_;
-  InterpolationMode interpolation_;
-  int32_t max_attempts_;
 };
 
 /// \brief RandomCropWithBBox TensorTransform.
@@ -421,11 +415,8 @@ class RandomCropWithBBox : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
-  std::vector<int32_t> padding_;
+  std::shared_ptr<Data> data_;
-  bool pad_if_needed_;
-  std::vector<uint8_t> fill_value_;
-  BorderType padding_mode_;
 };
 
 /// \brief RandomHorizontalFlip TensorTransform.
@@ -444,7 +435,8 @@ class RandomHorizontalFlip : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float probability_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomHorizontalFlipWithBBox TensorTransform.
@@ -463,7 +455,8 @@ class RandomHorizontalFlipWithBBox : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float probability_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomPosterize TensorTransform.
@@ -482,7 +475,8 @@ class RandomPosterize : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<uint8_t> bit_range_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomResize TensorTransform.
@@ -503,7 +497,8 @@ class RandomResize : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomResizeWithBBox TensorTransform.
@@ -525,7 +520,8 @@ class RandomResizeWithBBox : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomResizedCrop TensorTransform.
@@ -555,11 +551,8 @@ class RandomResizedCrop : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
-  std::vector<float> scale_;
+  std::shared_ptr<Data> data_;
-  std::vector<float> ratio_;
-  InterpolationMode interpolation_;
-  int32_t max_attempts_;
 };
 
 /// \brief RandomResizedCropWithBBox TensorTransform.
@@ -589,11 +582,8 @@ class RandomResizedCropWithBBox : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
-  std::vector<float> scale_;
+  std::shared_ptr<Data> data_;
-  std::vector<float> ratio_;
-  InterpolationMode interpolation_;
-  int32_t max_attempts_;
 };
 
 /// \brief RandomRotation TensorOp
@@ -620,11 +610,8 @@ class RandomRotation : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<float> degrees_;
+  struct Data;
-  InterpolationMode interpolation_mode_;
+  std::shared_ptr<Data> data_;
-  std::vector<float> center_;
-  bool expand_;
-  std::vector<uint8_t> fill_value_;
 };
 
 /// \brief RandomSelectSubpolicy TensorTransform.
@@ -655,7 +642,8 @@ class RandomSelectSubpolicy : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<std::vector<std::pair<std::shared_ptr<TensorOperation>, double>>> policy_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomSharpness TensorTransform.
@@ -675,7 +663,8 @@ class RandomSharpness : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<float> degrees_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomSolarize TensorTransform.
@@ -695,7 +684,8 @@ class RandomSolarize : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<uint8_t> threshold_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomVerticalFlip TensorTransform.
@@ -714,7 +704,8 @@ class RandomVerticalFlip : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float probability_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomVerticalFlipWithBBox TensorTransform.
@@ -733,7 +724,8 @@ class RandomVerticalFlipWithBBox : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float probability_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RescaleOperation TensorTransform.
@@ -753,8 +745,8 @@ class Rescale : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float rescale_;
+  struct Data;
-  float shift_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief ResizeWithBBox TensorTransform.
@@ -776,8 +768,8 @@ class ResizeWithBBox : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
-  InterpolationMode interpolation_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RgbaToBgr TensorTransform.
@@ -833,10 +825,8 @@ class SoftDvppDecodeRandomCropResizeJpeg : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
-  std::vector<float> scale_;
+  std::shared_ptr<Data> data_;
-  std::vector<float> ratio_;
-  int32_t max_attempts_;
 };
 
 /// \brief SoftDvppDecodeResizeJpeg TensorTransform.
@@ -864,7 +854,8 @@ class SoftDvppDecodeResizeJpeg : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief SwapRedBlue TensorOp
@@ -909,8 +900,8 @@ class UniformAugment : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<std::shared_ptr<TensorOperation>> transforms_;
+  struct Data;
-  int32_t num_ops_;
+  std::shared_ptr<Data> data_;
 };
 
 }  // namespace vision

mindspore/ccsrc/minddata/dataset/include/vision_ascend.h

@@ -50,7 +50,8 @@ class DvppDecodeResizeJpeg : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) override;
 
  private:
-  std::vector<uint32_t> resize_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 class DvppDecodeResizeCropJpeg : public TensorTransform {
@@ -70,8 +71,8 @@ class DvppDecodeResizeCropJpeg : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) override;
 
  private:
-  std::vector<uint32_t> crop_;
+  struct Data;
-  std::vector<uint32_t> resize_;
+  std::shared_ptr<Data> data_;
 };
 
 class DvppDecodePng : public TensorTransform {

mindspore/ccsrc/minddata/dataset/include/vision_lite.h

@@ -62,12 +62,8 @@ class Affine : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  float degrees_;
+  struct Data;
-  std::vector<float> translation_;
+  std::shared_ptr<Data> data_;
-  float scale_;
-  std::vector<float> shear_;
-  InterpolationMode interpolation_;
-  std::vector<uint8_t> fill_value_;
 };
 
 /// \brief CenterCrop TensorTransform.
@@ -90,7 +86,8 @@ class CenterCrop : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Crop TensorTransform.
@@ -112,8 +109,8 @@ class Crop : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<int32_t> coordinates_;
+  struct Data;
-  std::vector<int32_t> size_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Decode TensorTransform.
@@ -134,7 +131,8 @@ class Decode : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) override;
 
  private:
-  bool rgb_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Normalize TensorTransform.
@@ -158,8 +156,8 @@ class Normalize : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) override;
 
  private:
-  std::vector<float> mean_;
+  struct Data;
-  std::vector<float> std_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief RandomAffine TensorTransform.
@@ -196,12 +194,8 @@ class RandomAffine : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<float_t> degrees_;          // min_degree, max_degree
+  struct Data;
-  std::vector<float_t> translate_range_;  // maximum x translation percentage, maximum y translation percentage
+  std::shared_ptr<Data> data_;
-  std::vector<float_t> scale_range_;      // min_scale, max_scale
-  std::vector<float_t> shear_ranges_;     // min_x_shear, max_x_shear, min_y_shear, max_y_shear
-  InterpolationMode interpolation_;
-  std::vector<uint8_t> fill_value_;
 };
 
 /// \brief Resize TensorTransform.
@@ -225,8 +219,8 @@ class Resize : public TensorTransform {
   std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) override;
 
  private:
-  std::vector<int32_t> size_;
+  struct Data;
-  InterpolationMode interpolation_;
+  std::shared_ptr<Data> data_;
 };
 
 /// \brief Rotate TensorTransform.

mindspore/ccsrc/minddata/dataset/liteapi/include/datasets.h

@@ -19,6 +19,7 @@
 
 #include <sys/stat.h>
 #include <unistd.h>
+#include <algorithm>
 #include <map>
 #include <memory>
 #include <set>
@@ -28,8 +29,10 @@
 #include <utility>
 #include <vector>
 
+#include "include/api/dual_abi_helper.h"
 #include "include/iterator.h"
 #include "include/samplers.h"
+#include "include/transforms.h"
 
 namespace mindspore {
 namespace dataset {
@@ -94,11 +97,13 @@ class Dataset : public std::enable_shared_from_this<Dataset> {
 
   /// \brief Gets the column names
   /// \return Names of the columns. If failed, return an empty vector
-  std::vector<std::string> GetColumnNames();
+  std::vector<std::string> GetColumnNames() { return VectorCharToString(GetColumnNamesCharIF()); }
 
   /// \brief Gets the class indexing
   /// \return a map of ClassIndexing. If failed, return an empty map
-  std::vector<std::pair<std::string, std::vector<int32_t>>> GetClassIndexing();
+  std::vector<std::pair<std::string, std::vector<int32_t>>> GetClassIndexing() {
+    return ClassIndexCharToString(GetClassIndexingCharIF());
+  }
 
   /// \brief Setter function for runtime number of workers
   /// \param[in] num_workers The number of threads in this operator
@@ -110,7 +115,9 @@ class Dataset : public std::enable_shared_from_this<Dataset> {
   /// \param[in] num_epochs Number of epochs to run through the pipeline, default -1 which means infinite epochs.
   ///     An empty row is returned at the end of each epoch
   /// \return Shared pointer to the Iterator
-  std::shared_ptr<Iterator> CreateIterator(std::vector<std::string> columns = {}, int32_t num_epochs = -1);
+  std::shared_ptr<Iterator> CreateIterator(std::vector<std::string> columns = {}, int32_t num_epochs = -1) {
+    return CreateIteratorCharIF(VectorStringToChar(columns), num_epochs);
+  }
 
   /// \brief Function to create a BatchDataset
   /// \notes Combines batch_size number of consecutive rows into batches
@@ -138,14 +145,49 @@ class Dataset : public std::enable_shared_from_this<Dataset> {
   /// \param[in] project_columns A list of column names to project
   /// \param[in] cache Tensor cache to use. (default=nullptr which means no cache is used).
   /// \return Shared pointer to the current MapDataset
-  std::shared_ptr<MapDataset> Map(std::vector<std::shared_ptr<TensorOperation>> operations,
+  std::shared_ptr<MapDataset> Map(std::vector<TensorTransform *> operations,
                                   const std::vector<std::string> &input_columns = {},
                                   const std::vector<std::string> &output_columns = {},
                                   const std::vector<std::string> &project_columns = {},
                                   const std::shared_ptr<DatasetCache> &cache = nullptr,
                                   std::vector<std::shared_ptr<DSCallback>> callbacks = {}) {
-    return std::make_shared<MapDataset>(shared_from_this(), operations, input_columns, output_columns, project_columns,
+    std::vector<std::shared_ptr<TensorOperation>> transform_ops;
-                                        cache, callbacks);
+    (void)std::transform(
+      operations.begin(), operations.end(), std::back_inserter(transform_ops),
+      [](TensorTransform *op) -> std::shared_ptr<TensorOperation> { return op != nullptr ? op->Parse() : nullptr; });
+    return std::make_shared<MapDataset>(shared_from_this(), transform_ops, VectorStringToChar(input_columns),
+                                        VectorStringToChar(output_columns), VectorStringToChar(project_columns), cache,
+                                        callbacks);
+  }
+
+  std::shared_ptr<MapDataset> Map(std::vector<std::shared_ptr<TensorTransform>> operations,
+                                  const std::vector<std::string> &input_columns = {},
+                                  const std::vector<std::string> &output_columns = {},
+                                  const std::vector<std::string> &project_columns = {},
+                                  const std::shared_ptr<DatasetCache> &cache = nullptr,
+                                  std::vector<std::shared_ptr<DSCallback>> callbacks = {}) {
+    std::vector<std::shared_ptr<TensorOperation>> transform_ops;
+    (void)std::transform(operations.begin(), operations.end(), std::back_inserter(transform_ops),
+                         [](std::shared_ptr<TensorTransform> op) -> std::shared_ptr<TensorOperation> {
+                           return op != nullptr ? op->Parse() : nullptr;
+                         });
+    return std::make_shared<MapDataset>(shared_from_this(), transform_ops, VectorStringToChar(input_columns),
+                                        VectorStringToChar(output_columns), VectorStringToChar(project_columns), cache,
+                                        callbacks);
+  }
+
+  std::shared_ptr<MapDataset> Map(const std::vector<std::reference_wrapper<TensorTransform>> operations,
+                                  const std::vector<std::string> &input_columns = {},
+                                  const std::vector<std::string> &output_columns = {},
+                                  const std::vector<std::string> &project_columns = {},
+                                  const std::shared_ptr<DatasetCache> &cache = nullptr,
+                                  std::vector<std::shared_ptr<DSCallback>> callbacks = {}) {
+    std::vector<std::shared_ptr<TensorOperation>> transform_ops;
+    (void)std::transform(operations.begin(), operations.end(), std::back_inserter(transform_ops),
+                         [](TensorTransform &op) -> std::shared_ptr<TensorOperation> { return op.Parse(); });
+    return std::make_shared<MapDataset>(shared_from_this(), transform_ops, VectorStringToChar(input_columns),
+                                        VectorStringToChar(output_columns), VectorStringToChar(project_columns), cache,
+                                        callbacks);
   }
 
   /// \brief Function to create a Project Dataset
@@ -153,7 +195,7 @@ class Dataset : public std::enable_shared_from_this<Dataset> {
   /// \param[in] columns The name of columns to project
   /// \return Shared pointer to the current Dataset
   std::shared_ptr<ProjectDataset> Project(const std::vector<std::string> &columns) {
-    return std::make_shared<ProjectDataset>(shared_from_this(), columns);
+    return std::make_shared<ProjectDataset>(shared_from_this(), VectorStringToChar(columns));
   }
 
   /// \brief Function to create a Shuffle Dataset
@@ -169,6 +211,16 @@ class Dataset : public std::enable_shared_from_this<Dataset> {
  protected:
   std::shared_ptr<TreeGetters> tree_getters_;
   std::shared_ptr<DatasetNode> ir_node_;
+
+ private:
+  // Char interface(CharIF) of GetColumnNames
+  std::vector<std::vector<char>> GetColumnNamesCharIF();
+
+  // Char interface(CharIF) of GetClassIndexing
+  std::vector<std::pair<std::vector<char>, std::vector<int32_t>>> GetClassIndexingCharIF();
+
+  // Char interface(CharIF) of CreateIterator
+  std::shared_ptr<Iterator> CreateIteratorCharIF(std::vector<std::vector<char>> columns, int32_t num_epochs);
 };
 
 class BatchDataset : public Dataset {
@@ -180,15 +232,15 @@ class BatchDataset : public Dataset {
 class MapDataset : public Dataset {
  public:
   MapDataset(std::shared_ptr<Dataset> input, std::vector<std::shared_ptr<TensorOperation>> operations,
-             const std::vector<std::string> &input_columns, const std::vector<std::string> &output_columns,
+             const std::vector<std::vector<char>> &input_columns, const std::vector<std::vector<char>> &output_columns,
-             const std::vector<std::string> &project_columns, const std::shared_ptr<DatasetCache> &cache,
+             const std::vector<std::vector<char>> &project_columns, const std::shared_ptr<DatasetCache> &cache,
              std::vector<std::shared_ptr<DSCallback>> callbacks);
   ~MapDataset() = default;
 };
 
 class ProjectDataset : public Dataset {
  public:
-  ProjectDataset(std::shared_ptr<Dataset> input, const std::vector<std::string> &columns);
+  ProjectDataset(std::shared_ptr<Dataset> input, const std::vector<std::vector<char>> &columns);
   ~ProjectDataset() = default;
 };
 
@@ -201,14 +253,22 @@ class ShuffleDataset : public Dataset {
 /// \brief Function to create a SchemaObj
 /// \param[in] schema_file Path of schema file
 /// \return Shared pointer to the current schema
-std::shared_ptr<SchemaObj> Schema(const std::string &schema_file = "");
+std::shared_ptr<SchemaObj> SchemaCharIF(const std::vector<char> &schema_file);
 
+inline std::shared_ptr<SchemaObj> Schema(const std::string &schema_file = "") {
+  return SchemaCharIF(StringToChar(schema_file));
+}
 class AlbumDataset : public Dataset {
  public:
-  AlbumDataset(const std::string &dataset_dir, const std::string &data_schema,
+  AlbumDataset(const std::vector<char> &dataset_dir, const std::vector<char> &data_schema,
-               const std::vector<std::string> &column_names = {}, bool decode = false,
+               const std::vector<std::vector<char>> &column_names, bool decode, const std::shared_ptr<Sampler> &sampler,
-               const std::shared_ptr<SamplerObj> &sampler = RandomSampler(),
+               const std::shared_ptr<DatasetCache> &cache);
-               const std::shared_ptr<DatasetCache> &cache = nullptr);
+  AlbumDataset(const std::vector<char> &dataset_dir, const std::vector<char> &data_schema,
+               const std::vector<std::vector<char>> &column_names, bool decode, Sampler *sampler,
+               const std::shared_ptr<DatasetCache> &cache);
+  AlbumDataset(const std::vector<char> &dataset_dir, const std::vector<char> &data_schema,
+               const std::vector<std::vector<char>> &column_names, bool decode,
+               const std::reference_wrapper<Sampler> sampler, const std::shared_ptr<DatasetCache> &cache);
   ~AlbumDataset() = default;
 };
 
@@ -219,20 +279,58 @@ class AlbumDataset : public Dataset {
 /// \param[in] column_names Column names used to specify columns to load, if empty, will read all columns.
 ///     (default = {})
 /// \param[in] decode the option to decode the images in dataset (default = false)
-/// \param[in] sampler Object used to choose samples from the dataset. If sampler is not given,
+/// \param[in] sampler Shared pointer to a sampler object used to choose samples from the dataset. If sampler is not
+/// given,
 ///     a `RandomSampler` will be used to randomly iterate the entire dataset (default = RandomSampler())
 /// \param[in] cache Tensor cache to use. (default=nullptr which means no cache is used).
 /// \return Shared pointer to the current Dataset
-std::shared_ptr<AlbumDataset> Album(const std::string &dataset_dir, const std::string &data_schema,
+inline std::shared_ptr<AlbumDataset> Album(const std::string &dataset_dir, const std::string &data_schema,
-                                    const std::vector<std::string> &column_names = {}, bool decode = false,
+                                           const std::vector<std::string> &column_names = {}, bool decode = false,
-                                    const std::shared_ptr<SamplerObj> &sampler = RandomSampler(),
+                                           const std::shared_ptr<Sampler> &sampler = std::make_shared<RandomSampler>(),
-                                    const std::shared_ptr<DatasetCache> &cache = nullptr);
+                                           const std::shared_ptr<DatasetCache> &cache = nullptr) {
+  return std::make_shared<AlbumDataset>(StringToChar(dataset_dir), StringToChar(data_schema),
+                                        VectorStringToChar(column_names), decode, sampler, cache);
+}
+/// \brief Function to create an AlbumDataset
+/// \notes The generated dataset is specified through setting a schema
+/// \param[in] dataset_dir Path to the root directory that contains the dataset
+/// \param[in] data_schema Path to dataset schema file
+/// \param[in] column_names Column names used to specify columns to load
+/// \param[in] decode the option to decode the images in dataset
+/// \param[in] sampler Raw pointer to a sampler object used to choose samples from the dataset.
+/// \param[in] cache Tensor cache to use. (default=nullptr which means no cache is used).
+/// \return Shared pointer to the current Dataset
+inline std::shared_ptr<AlbumDataset> Album(const std::string &dataset_dir, const std::string &data_schema,
+                                           const std::vector<std::string> &column_names, bool decode, Sampler *sampler,
+                                           const std::shared_ptr<DatasetCache> &cache = nullptr) {
+  return std::make_shared<AlbumDataset>(StringToChar(dataset_dir), StringToChar(data_schema),
+                                        VectorStringToChar(column_names), decode, sampler, cache);
+}
+/// \brief Function to create an AlbumDataset
+/// \notes The generated dataset is specified through setting a schema
+/// \param[in] dataset_dir Path to the root directory that contains the dataset
+/// \param[in] data_schema Path to dataset schema file
+/// \param[in] column_names Column names used to specify columns to load
+/// \param[in] decode the option to decode the images in dataset
+/// \param[in] sampler Sampler object used to choose samples from the dataset.
+/// \param[in] cache Tensor cache to use. (default=nullptr which means no cache is used).
+/// \return Shared pointer to the current Dataset
+inline std::shared_ptr<AlbumDataset> Album(const std::string &dataset_dir, const std::string &data_schema,
+                                           const std::vector<std::string> &column_names, bool decode,
+                                           const std::reference_wrapper<Sampler> sampler,
+                                           const std::shared_ptr<DatasetCache> &cache = nullptr) {
+  return std::make_shared<AlbumDataset>(StringToChar(dataset_dir), StringToChar(data_schema),
+                                        VectorStringToChar(column_names), decode, sampler, cache);
+}
 
 class MnistDataset : public Dataset {
  public:
-  explicit MnistDataset(const std::string &dataset_dir, const std::string &usage = "all",
+  explicit MnistDataset(const std::vector<char> &dataset_dir, const std::vector<char> &usage,
-                        const std::shared_ptr<SamplerObj> &sampler = RandomSampler(),
+                        const std::shared_ptr<Sampler> &sampler, const std::shared_ptr<DatasetCache> &cache);
-                        const std::shared_ptr<DatasetCache> &cache = nullptr);
+  explicit MnistDataset(const std::vector<char> &dataset_dir, const std::vector<char> &usage, Sampler *sampler,
+                        const std::shared_ptr<DatasetCache> &cache);
+  explicit MnistDataset(const std::vector<char> &dataset_dir, const std::vector<char> &usage,
+                        const std::reference_wrapper<Sampler> sampler, const std::shared_ptr<DatasetCache> &cache);
   ~MnistDataset() = default;
 };
 
@@ -240,13 +338,41 @@ class MnistDataset : public Dataset {
 /// \notes The generated dataset has two columns ["image", "label"]
 /// \param[in] dataset_dir Path to the root directory that contains the dataset
 /// \param[in] usage of MNIST, can be "train", "test" or "all" (default = "all").
-/// \param[in] sampler Object used to choose samples from the dataset. If sampler is not given,
+/// \param[in] sampler Shared pointer to a sampler object used to choose samples from the dataset. If sampler is not
+/// given,
 ///     a `RandomSampler` will be used to randomly iterate the entire dataset (default = RandomSampler())
 /// \param[in] cache Tensor cache to use. (default=nullptr which means no cache is used).
 /// \return Shared pointer to the current MnistDataset
-std::shared_ptr<MnistDataset> Mnist(const std::string &dataset_dir, const std::string &usage = "all",
+inline std::shared_ptr<MnistDataset> Mnist(const std::string &dataset_dir, const std::string &usage = "all",
-                                    const std::shared_ptr<SamplerObj> &sampler = RandomSampler(),
+                                           const std::shared_ptr<Sampler> &sampler = std::make_shared<RandomSampler>(),
-                                    const std::shared_ptr<DatasetCache> &cache = nullptr);
+                                           const std::shared_ptr<DatasetCache> &cache = nullptr) {
+  return std::make_shared<MnistDataset>(StringToChar(dataset_dir), StringToChar(usage), sampler, cache);
+}
+
+/// \brief Function to create a MnistDataset
+/// \notes The generated dataset has two columns ["image", "label"]
+/// \param[in] dataset_dir Path to the root directory that contains the dataset
+/// \param[in] usage of MNIST, can be "train", "test" or "all"
+/// \param[in] sampler Raw pointer to a sampler object used to choose samples from the dataset.
+/// \param[in] cache Tensor cache to use. (default=nullptr which means no cache is used).
+/// \return Shared pointer to the current MnistDataset
+inline std::shared_ptr<MnistDataset> Mnist(const std::string &dataset_dir, const std::string &usage, Sampler *sampler,
+                                           const std::shared_ptr<DatasetCache> &cache = nullptr) {
+  return std::make_shared<MnistDataset>(StringToChar(dataset_dir), StringToChar(usage), sampler, cache);
+}
+
+/// \brief Function to create a MnistDataset
+/// \notes The generated dataset has two columns ["image", "label"]
+/// \param[in] dataset_dir Path to the root directory that contains the dataset
+/// \param[in] usage of MNIST, can be "train", "test" or "all"
+/// \param[in] sampler Sampler object used to choose samples from the dataset.
+/// \param[in] cache Tensor cache to use. (default=nullptr which means no cache is used).
+/// \return Shared pointer to the current MnistDataset
+inline std::shared_ptr<MnistDataset> Mnist(const std::string &dataset_dir, const std::string &usage,
+                                           const std::reference_wrapper<Sampler> sampler,
+                                           const std::shared_ptr<DatasetCache> &cache = nullptr) {
+  return std::make_shared<MnistDataset>(StringToChar(dataset_dir), StringToChar(usage), sampler, cache);
+}
 
 }  // namespace dataset
 }  // namespace mindspore

mindspore/ccsrc/minddata/dataset/liteapi/include/execute.h

@@ -26,16 +26,27 @@
 namespace mindspore {
 namespace dataset {
 
+class DeviceResource;
 // class to run tensor operations in eager mode
 class Execute {
  public:
   /// \brief Constructor
-  explicit Execute(std::shared_ptr<TensorOperation> op);
+  // FIXME - Temporarily overload Execute to support both TensorOperation and TensorTransform
+  explicit Execute(std::shared_ptr<TensorOperation> op, MapTargetDevice deviceType = MapTargetDevice::kCpu);
+  explicit Execute(std::shared_ptr<TensorTransform> op, MapTargetDevice deviceType = MapTargetDevice::kCpu);
+  // explicit Execute(TensorTransform op, MapTargetDevice deviceType = MapTargetDevice::KCpu);
+  explicit Execute(TensorTransform *op, MapTargetDevice deviceType = MapTargetDevice::kCpu);
 
-  explicit Execute(std::vector<std::shared_ptr<TensorOperation>> ops);
+  explicit Execute(std::vector<std::shared_ptr<TensorOperation>> ops,
+                   MapTargetDevice deviceType = MapTargetDevice::kCpu);
+  explicit Execute(std::vector<std::shared_ptr<TensorTransform>> ops,
+                   MapTargetDevice deviceType = MapTargetDevice::kCpu);
+  explicit Execute(const std::vector<std::reference_wrapper<TensorTransform>> ops,
+                   MapTargetDevice deviceType = MapTargetDevice::kCpu);
+  explicit Execute(std::vector<TensorTransform *> ops, MapTargetDevice deviceType = MapTargetDevice::kCpu);
 
   /// \brief Destructor
-  ~Execute() = default;
+  ~Execute();
 
   /// \brief callable function to execute the TensorOperation in eager mode
   /// \param[in] input Tensor to be transformed
@@ -49,8 +60,16 @@ class Execute {
   /// \return - Status
   Status operator()(const std::vector<mindspore::MSTensor> &input_tensor_list, std::vector<mindspore::MSTensor> *out);
 
+  Status DeviceMemoryRelease();
+
  private:
+  Status validate_device_();
+
   std::vector<std::shared_ptr<TensorOperation>> ops_;
+
+  MapTargetDevice device_type_;
+
+  std::shared_ptr<DeviceResource> device_resource_;
 };
 
 }  // namespace dataset

mindspore/ccsrc/minddata/dataset/liteapi/include/iterator.h

@@ -17,10 +17,12 @@
 #ifndef MINDSPORE_CCSRC_MINDDATA_DATASET_INCLUDE_ITERATOR_H_
 #define MINDSPORE_CCSRC_MINDDATA_DATASET_INCLUDE_ITERATOR_H_
 
+#include <map>
 #include <memory>
 #include <string>
 #include <unordered_map>
 #include <vector>
+#include "include/api/dual_abi_helper.h"
 #include "include/api/status.h"
 #include "include/api/types.h"
 
@@ -39,6 +41,7 @@ class IteratorConsumer;
 class Dataset;
 
 using MSTensorMap = std::unordered_map<std::string, mindspore::MSTensor>;
+using MSTensorMapChar = std::map<std::vector<char>, mindspore::MSTensor>;
 using MSTensorVec = std::vector<mindspore::MSTensor>;
 
 // Abstract class for iterating over the dataset.
@@ -60,7 +63,18 @@ class Iterator {
   /// \note Type of return data is a map(with column name).
   /// \param[out] row - the output tensor row.
   /// \return - a Status error code, returns OK if no error encountered.
-  Status GetNextRow(MSTensorMap *row);
+  Status GetNextRow(MSTensorMap *row) {
+    MSTensorMapChar row_;
+    row_.clear();
+    row->clear();
+    Status s = GetNextRowCharIF(&row_);
+    TensorMapCharToString(&row_, row);
+    return s;
+  }
+
+  // Char interface(CharIF) of GetNextRow
+  // This api exists because std::string will constrained by ABI compile macro but char don't.
+  Status GetNextRowCharIF(MSTensorMapChar *row);
 
   /// \brief Function to get the next row from the data pipeline.
   /// \note Type of return data is a vector(without column name).

mindspore/ccsrc/minddata/dataset/liteapi/include/samplers.h

@@ -26,143 +26,59 @@
 namespace mindspore {
 namespace dataset {
 
-// Internal Sampler class forward declaration
+class SamplerObj;
-class SamplerRT;
+
+// Abstract class to represent a sampler in the data pipeline.
+/// \class Sampler samplers.h
+/// \brief An abstract base class to represent a sampler in the data pipeline.
+class Sampler : std::enable_shared_from_this<Sampler> {
+  friend class AlbumDataset;
+  friend class MindDataDataset;
+  friend std::shared_ptr<SamplerObj> SelectSampler(int64_t, bool, int32_t, int32_t);
 
-class SamplerObj {
  public:
   /// \brief Constructor
-  SamplerObj();
+  Sampler() {}
 
   /// \brief Destructor
-  ~SamplerObj() = default;
+  ~Sampler() = default;
 
-  /// \brief Pure virtual function for derived class to implement parameters validation
+  /// \brief A virtual function to add a child sampler.
-  /// \return The Status code of the function. It returns OK status if parameters are valid.
+  /// \param[in] child The child sampler to be added as a children of this sampler.
-  virtual Status ValidateParams() = 0;
+  virtual void AddChild(std::shared_ptr<Sampler> child) { children_.push_back(child); }
-
-  /// \brief Pure virtual function to convert a SamplerObj class into a runtime sampler object
-  /// \return Shared pointers to the newly created Sampler
-  virtual std::shared_ptr<SamplerRT> SamplerBuild() = 0;
-
-  /// \brief Pure virtual function to copy a SamplerObj class
-  /// \return Shared pointers to the newly copied SamplerObj
-  virtual std::shared_ptr<SamplerObj> SamplerCopy() = 0;
-
-  /// \brief Function for derived class to get the shard id of sampler
-  /// \return The shard id of the derived sampler
-  virtual int64_t ShardId() { return 0; }
-
-  /// \brief Adds a child to the sampler
-  /// \param[in] child The sampler to be added as child
-  /// \return the Status code returned
-  Status AddChildSampler(std::shared_ptr<SamplerObj> child);
-
-  std::vector<std::shared_ptr<SamplerObj>> GetChild() { return children_; }
 
  protected:
-  /// \brief A function that calls build on the children of this sampler
+  /// \brief Pure virtual function to convert a Sampler class into an IR Sampler object.
-  /// \param[in] sampler The samplerRT object built from this sampler
+  /// \return shared pointer to the newly created TensorOperation.
-  void BuildChildren(std::shared_ptr<SamplerRT> sampler);
+  virtual std::shared_ptr<SamplerObj> Parse() = 0;
 
-  std::vector<std::shared_ptr<SamplerObj>> children_;
+  std::vector<std::shared_ptr<Sampler>> children_;
 };
 
-class DistributedSamplerObj;
+/// \brief A class to represent a Distributed Sampler in the data pipeline.
-class PKSamplerObj;
+/// \notes A Sampler that accesses a shard of the dataset.
-class PreBuiltSamplerObj;
+class DistributedSampler : public Sampler {
-class RandomSamplerObj;
+  friend std::shared_ptr<SamplerObj> SelectSampler(int64_t, bool, int32_t, int32_t);
-class SequentialSamplerObj;
-class SubsetSamplerObj;
-class SubsetRandomSamplerObj;
-class WeightedRandomSamplerObj;
 
-/// Function to create a Distributed Sampler.
-/// \notes A Sampler that access a shard of the dataset.
-/// \param[in] num_shards - Number of shards to divide the dataset into.
-/// \param[in] shard_id - Shard ID of the current shard within num_shards.
-/// \param[in] shuffle - If true, the indices are shuffled.
-/// \param[in] num_samples - The number of samples to draw (default to all elements).
-/// \param[in] seed - The seed in use when shuffle is true.
-/// \param[in] offset - The starting position where access to elements in the dataset begins.
-/// \param[in] even_dist - If true, each shard would return the same number of rows (default to true).
-///     If false the total rows returned by all the shards would not have overlap.
-/// \return Shared pointer to the current Sampler.
-std::shared_ptr<DistributedSamplerObj> DistributedSampler(int64_t num_shards, int64_t shard_id, bool shuffle = true,
-                                                          int64_t num_samples = 0, uint32_t seed = 1,
-                                                          int64_t offset = -1, bool even_dist = true);
-
-/// Function to create a PK Sampler.
-/// \notes Samples K elements for each P class in the dataset.
-///        This will sample all classes.
-/// \param[in] num_val - Number of elements to sample for each class.
-/// \param[in] shuffle - If true, the class IDs are shuffled.
-/// \param[in] num_samples - The number of samples to draw (default to all elements).
-/// \return Shared pointer to the current Sampler.
-std::shared_ptr<PKSamplerObj> PKSampler(int64_t num_val, bool shuffle = false, int64_t num_samples = 0);
-
-/// Function to create a Random Sampler.
-/// \notes Samples the elements randomly.
-/// \param[in] replacement - If true, put the sample ID back for the next draw.
-/// \param[in] num_samples - The number of samples to draw (default to all elements).
-/// \return Shared pointer to the current Sampler.
-std::shared_ptr<RandomSamplerObj> RandomSampler(bool replacement = false, int64_t num_samples = 0);
-
-/// Function to create a Sequential Sampler.
-/// \notes Samples the dataset elements sequentially, same as not having a sampler.
-/// \param[in] start_index - Index to start sampling at (default to start at first id).
-/// \param[in] num_samples - The number of samples to draw (default to all elements).
-/// \return Shared pointer to the current Sampler.
-std::shared_ptr<SequentialSamplerObj> SequentialSampler(int64_t start_index = 0, int64_t num_samples = 0);
-
-/// Function to create a Subset  Sampler.
-/// \notes Samples the elements from a sequence of indices.
-/// \param[in] indices - A vector sequence of indices.
-/// \param[in] num_samples - The number of samples to draw (default to all elements).
-/// \return Shared pointer to the current Sampler.
-std::shared_ptr<SubsetSamplerObj> SubsetSampler(std::vector<int64_t> indices, int64_t num_samples = 0);
-
-/// Function to create a Subset Random Sampler.
-/// \notes Samples the elements randomly from a sequence of indices.
-/// \param[in] indices - A vector sequence of indices.
-/// \param[in] num_samples - The number of samples to draw (default to all elements).
-/// \return Shared pointer to the current Sampler.
-std::shared_ptr<SubsetRandomSamplerObj> SubsetRandomSampler(std::vector<int64_t> indices, int64_t num_samples = 0);
-
-/// Function to create a Weighted Random Sampler.
-/// \notes Samples the elements from [0, len(weights) - 1] randomly with the given
-///        weights (probabilities).
-/// \param[in] weights - A vector sequence of weights, not necessarily summing up to 1.
-/// \param[in] num_samples - The number of samples to draw (default to all elements).
-/// \param[in] replacement - If true, put the sample ID back for the next draw.
-/// \return Shared pointer to the current Sampler.
-std::shared_ptr<WeightedRandomSamplerObj> WeightedRandomSampler(std::vector<double> weights, int64_t num_samples = 0,
-                                                                bool replacement = true);
-
-/* ####################################### Derived Sampler classes ################################# */
-class DistributedSamplerObj : public SamplerObj {
  public:
-  DistributedSamplerObj(int64_t num_shards, int64_t shard_id, bool shuffle, int64_t num_samples, uint32_t seed,
+  /// \brief Constructor
-                        int64_t offset, bool even_dist);
+  /// \param[in] num_shards - Number of shards to divide the dataset into.
+  /// \param[in] shard_id - Shard ID of the current shard within num_shards.
+  /// \param[in] shuffle - If true, the indices are shuffled.
+  /// \param[in] num_samples - The number of samples to draw (default to all elements).
+  /// \param[in] seed - The seed in use when shuffle is true.
+  /// \param[in] offset - The starting position where access to elements in the dataset begins.
+  /// \param[in] even_dist - If true, each shard would return the same number of rows (default to true).
+  ///     If false the total rows returned by all the shards would not have overlap.
+  explicit DistributedSampler(int64_t num_shards, int64_t shard_id, bool shuffle = true, int64_t num_samples = 0,
+                              uint32_t seed = 1, int64_t offset = -1, bool even_dist = true);
+  /// \brief Destructor.
+  ~DistributedSampler() = default;
 
-  virtual ~DistributedSamplerObj() = default;
+ protected:
-
+  /// \brief Function to convert a Sampler into an IR SamplerObj.
-  std::shared_ptr<SamplerRT> SamplerBuild() override;
+  /// \return shared pointer to the newly created SamplerObj.
-
+  std::shared_ptr<SamplerObj> Parse() override;
-  std::shared_ptr<SamplerObj> SamplerCopy() override {
-    auto sampler = std::make_shared<DistributedSamplerObj>(num_shards_, shard_id_, shuffle_, num_samples_, seed_,
-                                                           offset_, even_dist_);
-    for (auto child : children_) {
-      sampler->AddChildSampler(child);
-    }
-    return sampler;
-  }
-
-  Status ValidateParams() override;
-
-  /// \brief Function to get the shard id of sampler
-  /// \return The shard id of sampler
-  int64_t ShardId() override { return shard_id_; }
 
  private:
   int64_t num_shards_;
@@ -174,23 +90,26 @@ class DistributedSamplerObj : public SamplerObj {
   bool even_dist_;
 };
 
-class PKSamplerObj : public SamplerObj {
+/// \brief A class to represent a PK Sampler in the data pipeline.
+/// \notes Samples K elements for each P class in the dataset.
+///        This will sample all classes.
+class PKSampler : public Sampler {
+  friend std::shared_ptr<SamplerObj> SelectSampler(int64_t, bool, int32_t, int32_t);
+
  public:
-  PKSamplerObj(int64_t num_val, bool shuffle, int64_t num_samples);
+  /// \brief Constructor
+  /// \param[in] num_val - Number of elements to sample for each class.
+  /// \param[in] shuffle - If true, the class IDs are shuffled.
+  /// \param[in] num_samples - The number of samples to draw (default to all elements).
+  explicit PKSampler(int64_t num_val, bool shuffle = false, int64_t num_samples = 0);
 
-  virtual ~PKSamplerObj() = default;
+  /// \brief Destructor.
+  ~PKSampler() = default;
 
-  std::shared_ptr<SamplerRT> SamplerBuild() override;
+ protected:
-
+  /// \brief Function to convert a Sampler into an IR SamplerObj.
-  std::shared_ptr<SamplerObj> SamplerCopy() override {
+  /// \return shared pointer to the newly created SamplerObj.
-    auto sampler = std::make_shared<PKSamplerObj>(num_val_, shuffle_, num_samples_);
+  std::shared_ptr<SamplerObj> Parse() override;
-    for (auto child : children_) {
-      sampler->AddChildSampler(child);
-    }
-    return sampler;
-  }
-
-  Status ValidateParams() override;
 
  private:
   int64_t num_val_;
@@ -198,131 +117,120 @@ class PKSamplerObj : public SamplerObj {
   int64_t num_samples_;
 };
 
-class PreBuiltSamplerObj : public SamplerObj {
+/// \brief A class to represent a Random Sampler in the data pipeline.
+/// \notes Samples the elements randomly.
+class RandomSampler : public Sampler {
+  friend std::shared_ptr<SamplerObj> SelectSampler(int64_t, bool, int32_t, int32_t);
+
  public:
-  explicit PreBuiltSamplerObj(std::shared_ptr<SamplerRT> sampler);
+  /// \brief Constructor
+  /// \param[in] replacement - If true, put the sample ID back for the next draw.
+  /// \param[in] num_samples - The number of samples to draw (default to all elements).
+  explicit RandomSampler(bool replacement = false, int64_t num_samples = 0);
 
-  ~PreBuiltSamplerObj() = default;
+  /// \brief Destructor.
+  ~RandomSampler() = default;
 
-  std::shared_ptr<SamplerRT> SamplerBuild() override;
+ protected:
-
+  /// \brief Function to convert a Sampler into an IR SamplerObj.
-  std::shared_ptr<SamplerObj> SamplerCopy() override;
+  /// \return shared pointer to the newly created SamplerObj.
-
+  std::shared_ptr<SamplerObj> Parse() override;
-  Status ValidateParams() override;
-
- private:
-  std::shared_ptr<SamplerRT> sp_;
-};
-
-class RandomSamplerObj : public SamplerObj {
- public:
-  RandomSamplerObj(bool replacement, int64_t num_samples, bool reshuffle_each_epoch = true);
-
-  virtual ~RandomSamplerObj() = default;
-
-  std::shared_ptr<SamplerRT> SamplerBuild() override;
-
-  std::shared_ptr<SamplerObj> SamplerCopy() override {
-    auto sampler = std::make_shared<RandomSamplerObj>(replacement_, num_samples_, reshuffle_each_epoch_);
-    for (auto child : children_) {
-      sampler->AddChildSampler(child);
-    }
-    return sampler;
-  }
-
-  Status ValidateParams() override;
 
  private:
   bool replacement_;
   int64_t num_samples_;
-  bool reshuffle_each_epoch_;
 };
 
-class SequentialSamplerObj : public SamplerObj {
+/// \brief A class to represent a Sequential Sampler in the data pipeline.
+/// \notes Samples the dataset elements sequentially, same as not having a sampler.
+class SequentialSampler : public Sampler {
+  friend std::shared_ptr<SamplerObj> SelectSampler(int64_t, bool, int32_t, int32_t);
+
  public:
-  SequentialSamplerObj(int64_t start_index, int64_t num_samples);
+  /// \brief Constructor
+  /// \param[in] start_index - Index to start sampling at (default to start at first id).
+  /// \param[in] num_samples - The number of samples to draw (default to all elements).
+  explicit SequentialSampler(int64_t start_index = 0, int64_t num_samples = 0);
 
-  virtual ~SequentialSamplerObj() = default;
+  /// \brief Destructor.
+  ~SequentialSampler() = default;
 
-  std::shared_ptr<SamplerRT> SamplerBuild() override;
+ protected:
-
+  /// \brief Function to convert a Sampler into an IR SamplerObj.
-  std::shared_ptr<SamplerObj> SamplerCopy() override {
+  /// \return shared pointer to the newly created SamplerObj.
-    auto sampler = std::make_shared<SequentialSamplerObj>(start_index_, num_samples_);
+  std::shared_ptr<SamplerObj> Parse() override;
-    for (auto child : children_) {
-      sampler->AddChildSampler(child);
-    }
-    return sampler;
-  }
-
-  Status ValidateParams() override;
 
  private:
   int64_t start_index_;
   int64_t num_samples_;
 };
 
-class SubsetSamplerObj : public SamplerObj {
+/// \brief A class to represent a Subset Sampler in the data pipeline.
+/// \notes Samples the elements from a sequence of indices.
+class SubsetSampler : public Sampler {
+  friend std::shared_ptr<SamplerObj> SelectSampler(int64_t, bool, int32_t, int32_t);
+
  public:
-  SubsetSamplerObj(std::vector<int64_t> indices, int64_t num_samples);
+  /// \brief Constructor
+  /// \param[in] indices - A vector sequence of indices.
+  /// \param[in] num_samples - The number of samples to draw (default to all elements).
+  explicit SubsetSampler(std::vector<int64_t> indices, int64_t num_samples = 0);
 
-  virtual ~SubsetSamplerObj() = default;
+  /// \brief Destructor.
-
+  ~SubsetSampler() = default;
-  std::shared_ptr<SamplerRT> SamplerBuild() override;
-
-  std::shared_ptr<SamplerObj> SamplerCopy() override {
-    auto sampler = std::make_shared<SubsetSamplerObj>(indices_, num_samples_);
-    for (auto child : children_) {
-      sampler->AddChildSampler(child);
-    }
-    return sampler;
-  }
-
-  Status ValidateParams() override;
 
  protected:
-  const std::vector<int64_t> indices_;
+  /// \brief Function to convert a Sampler into an IR SamplerObj.
+  /// \return shared pointer to the newly created SamplerObj.
+  std::shared_ptr<SamplerObj> Parse() override;
+
+  std::vector<int64_t> indices_;
   int64_t num_samples_;
 };
 
-class SubsetRandomSamplerObj : public SubsetSamplerObj {
+/// \brief A class to represent a Subset Random Sampler in the data pipeline.
+/// \notes Samples the elements randomly from a sequence of indices.
+class SubsetRandomSampler : public SubsetSampler {
+  friend std::shared_ptr<SamplerObj> SelectSampler(int64_t, bool, int32_t, int32_t);
+
  public:
-  SubsetRandomSamplerObj(std::vector<int64_t> indices, int64_t num_samples);
+  /// \brief Constructor
+  /// \param[in] indices - A vector sequence of indices.
+  /// \param[in] num_samples - The number of samples to draw (default to all elements).
+  explicit SubsetRandomSampler(std::vector<int64_t> indices, int64_t num_samples = 0);
 
-  ~SubsetRandomSamplerObj() = default;
+  /// \brief Destructor.
+  ~SubsetRandomSampler() = default;
 
-  std::shared_ptr<SamplerRT> SamplerBuild() override;
+ protected:
-
+  /// \brief Function to convert a Sampler into an IR SamplerObj.
-  std::shared_ptr<SamplerObj> SamplerCopy() override {
+  /// \return shared pointer to the newly created SamplerObj.
-    auto sampler = std::make_shared<SubsetRandomSamplerObj>(indices_, num_samples_);
+  std::shared_ptr<SamplerObj> Parse() override;
-    for (auto child : children_) {
-      sampler->AddChildSampler(child);
-    }
-    return sampler;
-  }
-
- private:
 };
 
-class WeightedRandomSamplerObj : public SamplerObj {
+/// \brief A class to represent a Weighted Random Sampler in the data pipeline.
+/// \notes Samples the elements from [0, len(weights) - 1] randomly with the given
+///        weights (probabilities).
+class WeightedRandomSampler : public Sampler {
+  friend std::shared_ptr<SamplerObj> SelectSampler(int64_t, bool, int32_t, int32_t);
+
  public:
-  explicit WeightedRandomSamplerObj(std::vector<double> weights, int64_t num_samples = 0, bool replacement = true);
+  /// \brief Constructor
+  /// \param[in] weights - A vector sequence of weights, not necessarily summing up to 1.
+  /// \param[in] num_samples - The number of samples to draw (default to all elements).
+  /// \param[in] replacement - If true, put the sample ID back for the next draw.
+  explicit WeightedRandomSampler(std::vector<double> weights, int64_t num_samples = 0, bool replacement = true);
 
-  virtual ~WeightedRandomSamplerObj() = default;
+  /// \brief Destructor.
+  ~WeightedRandomSampler() = default;
 
-  std::shared_ptr<SamplerRT> SamplerBuild() override;
+ protected:
-
+  /// \brief Function to convert a Sampler into an IR SamplerObj.
-  std::shared_ptr<SamplerObj> SamplerCopy() override {
+  /// \return shared pointer to the newly created SamplerObj.
-    auto sampler = std::make_shared<WeightedRandomSamplerObj>(weights_, num_samples_, replacement_);
+  std::shared_ptr<SamplerObj> Parse() override;
-    for (auto child : children_) {
-      sampler->AddChildSampler(child);
-    }
-    return sampler;
-  }
-
-  Status ValidateParams() override;
 
  private:
-  const std::vector<double> weights_;
+  std::vector<double> weights_;
   int64_t num_samples_;
   bool replacement_;
 };

mindspore/ccsrc/minddata/dataset/liteapi/include/transforms.h

@@ -20,231 +20,178 @@
 #include <memory>
 #include <string>
 #include <vector>
-#include "include/constants.h"
+#include "include/api/dual_abi_helper.h"
 #include "include/api/status.h"
+#include "include/constants.h"
 
 namespace mindspore {
 namespace dataset {
 
-class TensorOp;
+class TensorOperation;
 
-// Char arrays storing name of corresponding classes (in alphabetical order)
+// Abstract class to represent a tensor transform operation in the data pipeline.
-constexpr char kComposeOperation[] = "Compose";
+/// \class TensorTransform transforms.h
-constexpr char kDuplicateOperation[] = "Duplicate";
+/// \brief A base class to represent a tensor transform operation in the data pipeline.
-constexpr char kOneHotOperation[] = "OneHot";
+class TensorTransform : public std::enable_shared_from_this<TensorTransform> {
-constexpr char kPreBuiltOperation[] = "PreBuilt";
-constexpr char kRandomApplyOperation[] = "RandomApply";
-constexpr char kRandomChoiceOperation[] = "RandomChoice";
-constexpr char kRandomSelectSubpolicyOperation[] = "RandomSelectSubpolicy";
-constexpr char kTypeCastOperation[] = "TypeCast";
-constexpr char kUniqueOperation[] = "Unique";
-
-// Abstract class to represent a dataset in the data pipeline.
-class TensorOperation : public std::enable_shared_from_this<TensorOperation> {
  public:
   /// \brief Constructor
-  TensorOperation() : random_op_(false) {}
+  TensorTransform() {}
-
-  /// \brief Constructor
-  explicit TensorOperation(bool random) : random_op_(random) {}
 
   /// \brief Destructor
-  ~TensorOperation() = default;
+  ~TensorTransform() = default;
 
-  /// \brief Pure virtual function to convert a TensorOperation class into a runtime TensorOp object.
+  /// \brief Pure virtual function to convert a TensorTransform class into a IR TensorOperation object.
-  /// \return shared pointer to the newly created TensorOp.
+  /// \return shared pointer to the newly created TensorOperation.
-  virtual std::shared_ptr<TensorOp> Build() = 0;
+  virtual std::shared_ptr<TensorOperation> Parse() = 0;
 
-  virtual Status ValidateParams() = 0;
+  /// \brief Virtual function to convert a TensorTransform class into a IR TensorOperation object.
-
+  /// \param[in] env A string to determine the running environment
-  virtual std::string Name() const = 0;
+  /// \return shared pointer to the newly created TensorOperation.
-
+  virtual std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) { return nullptr; }
-  /// \brief Check whether the operation is deterministic.
-  /// \return true if this op is a random op (returns non-deterministic result e.g. RandomCrop)
-  bool IsRandomOp() const { return random_op_; }
-
- protected:
-  bool random_op_;
 };
 
-// Helper function to validate fill value
-Status ValidateVectorFillvalue(const std::string &transform_name, const std::vector<uint8_t> &fill_value);
-
-// Helper function to validate probability
-Status ValidateProbability(const std::string &transform_name, const float &probability);
-
-// Helper function to validate padding
-Status ValidateVectorPadding(const std::string &transform_name, const std::vector<int32_t> &padding);
-
-// Helper function to validate size
-Status ValidateVectorPositive(const std::string &transform_name, const std::vector<int32_t> &size);
-
-// Helper function to validate transforms
-Status ValidateVectorTransforms(const std::string &transform_name,
-                                const std::vector<std::shared_ptr<TensorOperation>> &transforms);
-
-// Helper function to compare float value
-bool CmpFloat(const float &a, const float &b, float epsilon = 0.0000000001f);
-
 // Transform operations for performing data transformation.
 namespace transforms {
 
-// Transform Op classes (in alphabetical order)
+/// \brief Compose Op.
-class ComposeOperation;
-class DuplicateOperation;
-class OneHotOperation;
-class PreBuiltOperation;
-class RandomApplyOperation;
-class RandomChoiceOperation;
-class TypeCastOperation;
-
-/// \brief Function to create a Compose TensorOperation.
 /// \notes Compose a list of transforms into a single transform.
-/// \param[in] transforms A vector of transformations to be applied.
+class Compose : public TensorTransform {
-/// \return Shared pointer to the current TensorOperation.
+ public:
-std::shared_ptr<ComposeOperation> Compose(const std::vector<std::shared_ptr<TensorOperation>> &transforms);
+  /// \brief Constructor.
+  /// \param[in] transforms A vector of transformations to be applied.
+  explicit Compose(const std::vector<TensorTransform *> &transforms);
+  explicit Compose(const std::vector<std::shared_ptr<TensorTransform>> &transforms);
+  explicit Compose(const std::vector<std::reference_wrapper<TensorTransform>> &transforms);
 
-/// \brief Function to create a Duplicate TensorOperation.
+  /// \brief Destructor
+  ~Compose() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+/// \brief Duplicate Op.
 /// \notes Duplicate the input tensor to a new output tensor.
 ///     The input tensor is carried over to the output list.
-/// \return Shared pointer to the current TensorOperation.
+class Duplicate : public TensorTransform {
-std::shared_ptr<DuplicateOperation> Duplicate();
+ public:
+  /// \brief Constructor.
+  Duplicate();
 
-/// \brief Function to create a OneHot TensorOperation.
+  /// \brief Destructor
+  ~Duplicate() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+};
+
+/// \brief OneHot Op.
 /// \notes Convert the labels into OneHot format.
-/// \param[in] num_classes number of classes.
+class OneHot : public TensorTransform {
-/// \return Shared pointer to the current TensorOperation.
+ public:
-std::shared_ptr<OneHotOperation> OneHot(int32_t num_classes);
+  /// \brief Constructor.
+  /// \param[in] num_classes number of classes.
+  explicit OneHot(int32_t num_classes);
 
-/// \brief Function to create a RandomApply TensorOperation.
+  /// \brief Destructor
+  ~OneHot() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+/// \brief RandomApply Op.
 /// \notes Randomly perform a series of transforms with a given probability.
-/// \param[in] transforms A vector of transformations to be applied.
+class RandomApply : public TensorTransform {
-/// \param[in] prob The probability to apply the transformation list (default=0.5)
+ public:
-/// \return Shared pointer to the current TensorOperation.
+  /// \brief Constructor.
-std::shared_ptr<RandomApplyOperation> RandomApply(const std::vector<std::shared_ptr<TensorOperation>> &transforms,
+  /// \param[in] transforms A vector of transformations to be applied.
-                                                  double prob = 0.5);
+  /// \param[in] prob The probability to apply the transformation list (default=0.5)
+  explicit RandomApply(const std::vector<TensorTransform *> &transforms, double prob = 0.5);
+  explicit RandomApply(const std::vector<std::shared_ptr<TensorTransform>> &transforms, double prob = 0.5);
+  explicit RandomApply(const std::vector<std::reference_wrapper<TensorTransform>> &transforms, double prob = 0.5);
 
-/// \brief Function to create a RandomChoice TensorOperation.
+  /// \brief Destructor
+  ~RandomApply() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+/// \brief RandomChoice Op.
 /// \notes Randomly selects one transform from a list of transforms to perform operation.
-/// \param[in] transforms A vector of transformations to be chosen from to apply.
+class RandomChoice : public TensorTransform {
-/// \return Shared pointer to the current TensorOperation.
+ public:
-std::shared_ptr<RandomChoiceOperation> RandomChoice(const std::vector<std::shared_ptr<TensorOperation>> &transforms);
+  /// \brief Constructor.
+  /// \param[in] transforms A vector of transformations to be chosen from to apply.
+  explicit RandomChoice(const std::vector<TensorTransform *> &transforms);
+  explicit RandomChoice(const std::vector<std::shared_ptr<TensorTransform>> &transforms);
+  explicit RandomChoice(const std::vector<std::reference_wrapper<TensorTransform>> &transforms);
 
-/// \brief Function to create a TypeCast TensorOperation.
+  /// \brief Destructor
+  ~RandomChoice() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+/// \brief TypeCast Op.
 /// \notes Tensor operation to cast to a given MindSpore data type.
-/// \param[in] data_type mindspore.dtype to be cast to.
+class TypeCast : public TensorTransform {
-/// \return Shared pointer to the current TensorOperation.
-std::shared_ptr<TypeCastOperation> TypeCast(std::string data_type);
-
-/* ####################################### Derived TensorOperation classes ################################# */
-
-class ComposeOperation : public TensorOperation {
  public:
-  explicit ComposeOperation(const std::vector<std::shared_ptr<TensorOperation>> &transforms);
+  /// \brief Constructor.
+  /// \param[in] data_type mindspore.dtype to be cast to.
+  explicit TypeCast(std::string data_type) : TypeCast(StringToChar(data_type)) {}
 
-  ~ComposeOperation() = default;
+  explicit TypeCast(const std::vector<char> &data_type);
 
-  std::shared_ptr<TensorOp> Build() override;
+  /// \brief Destructor
+  ~TypeCast() = default;
 
-  Status ValidateParams() override;
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
-
+  /// \return Shared pointer to TensorOperation object.
-  std::string Name() const override { return kComposeOperation; }
+  std::shared_ptr<TensorOperation> Parse() override;
 
  private:
-  std::vector<std::shared_ptr<TensorOperation>> transforms_;
+  struct Data;
+  std::shared_ptr<Data> data_;
 };
 
-class DuplicateOperation : public TensorOperation {
+/// \brief Unique Op.
+/// \notes Return an output tensor containing all the unique elements of the input tensor in
+///     the same order that they occur in the input tensor.
+class Unique : public TensorTransform {
  public:
-  DuplicateOperation() = default;
+  /// \brief Constructor.
+  Unique();
 
-  ~DuplicateOperation() = default;
+  /// \brief Destructor
+  ~Unique() = default;
 
-  std::shared_ptr<TensorOp> Build() override;
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
-
+  /// \return Shared pointer to TensorOperation object.
-  Status ValidateParams() override;
+  std::shared_ptr<TensorOperation> Parse() override;
-
-  std::string Name() const override { return kDuplicateOperation; }
-};
-
-class OneHotOperation : public TensorOperation {
- public:
-  explicit OneHotOperation(int32_t num_classes_);
-
-  ~OneHotOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kOneHotOperation; }
-
- private:
-  float num_classes_;
-};
-
-class PreBuiltOperation : public TensorOperation {
- public:
-  explicit PreBuiltOperation(std::shared_ptr<TensorOp> tensor_op);
-
-  ~PreBuiltOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kPreBuiltOperation; }
-
- private:
-  std::shared_ptr<TensorOp> op_;
-};
-
-class RandomApplyOperation : public TensorOperation {
- public:
-  explicit RandomApplyOperation(const std::vector<std::shared_ptr<TensorOperation>> &transforms, double prob);
-
-  ~RandomApplyOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kRandomApplyOperation; }
-
- private:
-  std::vector<std::shared_ptr<TensorOperation>> transforms_;
-  double prob_;
-};
-
-class RandomChoiceOperation : public TensorOperation {
- public:
-  explicit RandomChoiceOperation(const std::vector<std::shared_ptr<TensorOperation>> &transforms);
-
-  ~RandomChoiceOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kRandomChoiceOperation; }
-
- private:
-  std::vector<std::shared_ptr<TensorOperation>> transforms_;
-};
-class TypeCastOperation : public TensorOperation {
- public:
-  explicit TypeCastOperation(std::string data_type);
-
-  ~TypeCastOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kTypeCastOperation; }
-
- private:
-  std::string data_type_;
 };
 }  // namespace transforms
 }  // namespace dataset

mindspore/ccsrc/minddata/dataset/liteapi/include/vision_lite.h

@@ -22,7 +22,7 @@
 #include <string>
 #include <utility>
 #include <vector>
-
+#include "include/constants.h"
 #include "include/transforms.h"
 
 namespace mindspore {
@@ -31,167 +31,210 @@ namespace dataset {
 // Transform operations for performing computer vision.
 namespace vision {
 
-// Char arrays storing name of corresponding classes (in alphabetical order)
+// Forward Declarations
-constexpr char kCenterCropOperation[] = "CenterCrop";
-constexpr char kCropOperation[] = "Crop";
-constexpr char kDecodeOperation[] = "Decode";
-constexpr char kNormalizeOperation[] = "Normalize";
-constexpr char kResizeOperation[] = "Resize";
-constexpr char kRotateOperation[] = "Rotate";
-// Transform Op classes (in alphabetical order)
-class CenterCropOperation;
-class CropOperation;
-class DecodeOperation;
-class NormalizeOperation;
-class ResizeOperation;
 class RotateOperation;
 
-/// \brief Function to create a CenterCrop TensorOperation.
+/// \brief Affine TensorTransform.
+/// \notes Apply affine transform on input image.
+class Affine : public TensorTransform {
+ public:
+  /// \brief Constructor.
+  /// \param[in] degrees The degrees to rotate the image by
+  /// \param[in] translation The value representing vertical and horizontal translation (default = {0.0, 0.0})
+  ///     The first value represent the x axis translation while the second represents y axis translation.
+  /// \param[in] scale The scaling factor for the image (default = 0.0)
+  /// \param[in] shear A float vector of size 2, representing the shear degrees (default = {0.0, 0.0})
+  /// \param[in] interpolation An enum for the mode of interpolation
+  /// \param[in] fill_value A vector representing the value to fill the area outside the transform
+  ///    in the output image. If 1 value is provided, it is used for all RGB channels.
+  ///    If 3 values are provided, it is used to fill R, G, B channels respectively.
+  explicit Affine(float_t degrees, const std::vector<float> &translation = {0.0, 0.0}, float scale = 0.0,
+                  const std::vector<float> &shear = {0.0, 0.0},
+                  InterpolationMode interpolation = InterpolationMode::kNearestNeighbour,
+                  const std::vector<uint8_t> &fill_value = {0, 0, 0});
+
+  /// \brief Destructor.
+  ~Affine() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+/// \brief CenterCrop TensorTransform.
 /// \notes Crops the input image at the center to the given size.
-/// \param[in] size A vector representing the output size of the cropped image.
+class CenterCrop : public TensorTransform {
-///     If size is a single value, a square crop of size (size, size) is returned.
+ public:
-///     If size has 2 values, it should be (height, width).
+  /// \brief Constructor.
-/// \return Shared pointer to the current TensorOperation.
+  /// \param[in] size A vector representing the output size of the cropped image.
-std::shared_ptr<CenterCropOperation> CenterCrop(std::vector<int32_t> size);
+  ///     If size is a single value, a square crop of size (size, size) is returned.
+  ///     If size has 2 values, it should be (height, width).
+  explicit CenterCrop(std::vector<int32_t> size);
 
-/// \brief Function to create a Crop TensorOp
+  /// \brief Destructor.
+  ~CenterCrop() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+  std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+/// \brief Crop TensorTransform.
 /// \notes Crop an image based on location and crop size
-/// \param[in] coordinates Starting location of crop. Must be a vector of two values, in the form of {x_coor, y_coor}
+class Crop : public TensorTransform {
-/// \param[in] size Size of the cropped area.
+ public:
-///     If size is a single value, a square crop of size (size, size) is returned.
+  /// \brief Constructor.
-///     If size has 2 values, it should be (height, width).
+  /// \param[in] coordinates Starting location of crop. Must be a vector of two values, in the form of {x_coor, y_coor}
-/// \return Shared pointer to the current TensorOp
+  /// \param[in] size Size of the cropped area.
-std::shared_ptr<CropOperation> Crop(std::vector<int32_t> coordinates, std::vector<int32_t> size);
+  ///     If size is a single value, a square crop of size (size, size) is returned.
+  ///     If size has 2 values, it should be (height, width).
+  Crop(std::vector<int32_t> coordinates, std::vector<int32_t> size);
 
-/// \brief Function to create a Decode TensorOperation.
+  /// \brief Destructor.
+  ~Crop() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+/// \brief Decode TensorTransform.
 /// \notes Decode the input image in RGB mode.
-/// \param[in] rgb A boolean of whether to decode in RGB mode or not.
+class Decode : public TensorTransform {
-/// \return Shared pointer to the current TensorOperation.
+ public:
-std::shared_ptr<DecodeOperation> Decode(bool rgb = true);
+  /// \brief Constructor.
+  /// \param[in] rgb A boolean of whether to decode in RGB mode or not.
+  explicit Decode(bool rgb = true);
 
-/// \brief Function to create a Normalize TensorOperation.
+  /// \brief Destructor.
+  ~Decode() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+  std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+/// \brief Normalize TensorTransform.
 /// \notes Normalize the input image with respect to mean and standard deviation.
-/// \param[in] mean A vector of mean values for each channel, w.r.t channel order.
+class Normalize : public TensorTransform {
-///     The mean values must be in range [0.0, 255.0].
+ public:
-/// \param[in] std A vector of standard deviations for each channel, w.r.t. channel order.
+  /// \brief Constructor.
-///     The standard deviation values must be in range (0.0, 255.0]
+  /// \param[in] mean A vector of mean values for each channel, w.r.t channel order.
-/// \return Shared pointer to the current TensorOperation.
+  ///     The mean values must be in range [0.0, 255.0].
-std::shared_ptr<NormalizeOperation> Normalize(std::vector<float> mean, std::vector<float> std);
+  /// \param[in] std A vector of standard deviations for each channel, w.r.t. channel order.
+  ///     The standard deviation values must be in range (0.0, 255.0]
+  Normalize(std::vector<float> mean, std::vector<float> std);
 
-/// \brief Function to create a Resize TensorOperation.
+  /// \brief Destructor.
+  ~Normalize() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+class RandomAffine : public TensorTransform {
+ public:
+  /// \brief Constructor.
+  /// \param[in] degrees A float vector of size 2, representing the starting and ending degree
+  /// \param[in] translate_range A float vector of size 2 or 4, representing percentages of translation on x and y axes.
+  ///    if size is 2, (min_dx, max_dx, 0, 0)
+  ///    if size is 4, (min_dx, max_dx, min_dy, max_dy)
+  ///    all values are in range [-1, 1]
+  /// \param[in] scale_range A float vector of size 2, representing the starting and ending scales in the range.
+  /// \param[in] shear_ranges A float vector of size 2 or 4, representing the starting and ending shear degrees
+  ///    vertically and horizontally.
+  ///    if size is 2, (min_shear_x, max_shear_x, 0, 0)
+  ///    if size is 4, (min_shear_x, max_shear_x, min_shear_y, max_shear_y)
+  /// \param[in] interpolation An enum for the mode of interpolation
+  /// \param[in] fill_value A vector representing the value to fill the area outside the transform
+  ///    in the output image. If 1 value is provided, it is used for all RGB channels.
+  ///    If 3 values are provided, it is used to fill R, G, B channels respectively.
+  explicit RandomAffine(const std::vector<float_t> &degrees,
+                        const std::vector<float_t> &translate_range = {0.0, 0.0, 0.0, 0.0},
+                        const std::vector<float_t> &scale_range = {1.0, 1.0},
+                        const std::vector<float_t> &shear_ranges = {0.0, 0.0, 0.0, 0.0},
+                        InterpolationMode interpolation = InterpolationMode::kNearestNeighbour,
+                        const std::vector<uint8_t> &fill_value = {0, 0, 0});
+
+  /// \brief Destructor.
+  ~RandomAffine() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+/// \brief Resize TensorTransform.
 /// \notes Resize the input image to the given size.
-/// \param[in] size A vector representing the output size of the resized image.
+class Resize : public TensorTransform {
-///     If size is a single value, the image will be resized to this value with
+ public:
-///     the same image aspect ratio. If size has 2 values, it should be (height, width).
+  /// \brief Constructor.
-/// \param[in] interpolation An enum for the mode of interpolation
+  /// \param[in] size A vector representing the output size of the resized image.
-/// \return Shared pointer to the current TensorOperation.
+  ///     If size is a single value, the image will be resized to this value with
-std::shared_ptr<ResizeOperation> Resize(std::vector<int32_t> size,
+  ///     the same image aspect ratio. If size has 2 values, it should be (height, width).
-                                        InterpolationMode interpolation = InterpolationMode::kLinear);
+  /// \param[in] interpolation An enum for the mode of interpolation
-/// \brief Applies an rotate transformation to an image.
+  explicit Resize(std::vector<int32_t> size, InterpolationMode interpolation = InterpolationMode::kLinear);
+
+  /// \brief Destructor.
+  ~Resize() = default;
+
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
+  /// \return Shared pointer to TensorOperation object.
+  std::shared_ptr<TensorOperation> Parse() override;
+
+  std::shared_ptr<TensorOperation> Parse(const MapTargetDevice &env) override;
+
+ private:
+  struct Data;
+  std::shared_ptr<Data> data_;
+};
+
+/// \brief Rotate TensorTransform.
 /// \notes Rotate the input image using a specified angle id.
-/// \return Shared pointer to the current TensorOperation.
+class Rotate : public TensorTransform {
-std::shared_ptr<RotateOperation> Rotate();
-
-class CenterCropOperation : public TensorOperation {
  public:
-  explicit CenterCropOperation(std::vector<int32_t> size);
+  /// \brief Constructor.
+  Rotate();
 
-  ~CenterCropOperation() = default;
+  /// \brief Destructor.
+  ~Rotate() = default;
 
-  std::shared_ptr<TensorOp> Build() override;
+  /// \brief Function to convert TensorTransform object into a TensorOperation object.
-
+  /// \return Shared pointer to TensorOperation object.
-  Status ValidateParams() override;
+  std::shared_ptr<TensorOperation> Parse() override;
-
-  std::string Name() const override { return kCenterCropOperation; }
 
  private:
-  std::vector<int32_t> size_;
+  std::shared_ptr<RotateOperation> op_;
 };
 
-class CropOperation : public TensorOperation {
- public:
-  CropOperation(std::vector<int32_t> coordinates, std::vector<int32_t> size);
-
-  ~CropOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kCropOperation; }
-
- private:
-  std::vector<int32_t> coordinates_;
-  std::vector<int32_t> size_;
-};
-class DecodeOperation : public TensorOperation {
- public:
-  explicit DecodeOperation(bool rgb = true);
-
-  ~DecodeOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kDecodeOperation; }
-
- private:
-  bool rgb_;
-};
-
-class NormalizeOperation : public TensorOperation {
- public:
-  NormalizeOperation(std::vector<float> mean, std::vector<float> std);
-
-  ~NormalizeOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kNormalizeOperation; }
-
- private:
-  std::vector<float> mean_;
-  std::vector<float> std_;
-};
-
-class ResizeOperation : public TensorOperation {
- public:
-  explicit ResizeOperation(std::vector<int32_t> size,
-                           InterpolationMode interpolation_mode = InterpolationMode::kLinear);
-
-  ~ResizeOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kResizeOperation; }
-
- private:
-  std::vector<int32_t> size_;
-  InterpolationMode interpolation_;
-};
-
-class RotateOperation : public TensorOperation {
- public:
-  RotateOperation();
-
-  ~RotateOperation() = default;
-
-  std::shared_ptr<TensorOp> Build() override;
-
-  Status ValidateParams() override;
-
-  std::string Name() const override { return kRotateOperation; }
-
-  void setAngle(uint64_t angle_id);
-
- private:
-  std::shared_ptr<TensorOp> rotate_op;
-};
 }  // namespace vision
 }  // namespace dataset
 }  // namespace mindspore