!182 Tuning mindrecord writer performance

Merge pull request !182 from jonathan_yan/write_a

example/convert_to_mindrecord/README.md

@@ -0,0 +1,46 @@
+# MindRecord generating guidelines
+
+<!-- TOC -->
+
+- [MindRecord generating guidelines](#mindrecord-generating-guidelines)
+    - [Create work space](#create-work-space)
+    - [Implement data generator](#implement-data-generator)
+    - [Run data generator](#run-data-generator)
+
+<!-- /TOC -->
+
+## Create work space
+
+Assume the dataset name is 'xyz'
+* Create work space from template
+    ```shell
+    cd ${your_mindspore_home}/example/convert_to_mindrecord
+    cp -r template xyz
+    ```
+
+## Implement data generator 
+
+Edit dictionary data generator  
+* Edit file 
+    ```shell
+    cd ${your_mindspore_home}/example/convert_to_mindrecord
+    vi xyz/mr_api.py
+    ```
+
+ Two API, 'mindrecord_task_number' and 'mindrecord_dict_data', must be implemented
+- 'mindrecord_task_number()' returns number of tasks. Return 1 if data row is generated serially. Return N if generator can be split into N parallel-run tasks.
+- 'mindrecord_dict_data(task_id)' yields dictionary data row by row. 'task_id' is 0..N-1, if N is return value of mindrecord_task_number()
+
+
+Tricky for parallel run
+- For imagenet, one directory can be a task.
+- For TFRecord with multiple files, each file can be a task.
+- For TFRecord with 1 file only, it could also be split into N tasks. Task_id=K means: data row is picked only if (count % N == K) 
+
+
+## Run data generator 
+* run python script 
+    ```shell
+    cd ${your_mindspore_home}/example/convert_to_mindrecord
+    python writer.py --mindrecord_script imagenet [...]
+    ```

example/convert_to_mindrecord/imagenet/mr_api.py

@@ -0,0 +1,122 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""
+User-defined API for MindRecord writer.
+Two API must be implemented,
+  1. mindrecord_task_number()
+       # Return number of parallel tasks. return 1 if no parallel
+  2. mindrecord_dict_data(task_id)
+       # Yield data for one task
+       # task_id is 0..N-1, if N is return value of mindrecord_task_number()
+"""
+import argparse
+import os
+import pickle
+
+######## mindrecord_schema begin ##########
+mindrecord_schema = {"label": {"type": "int64"},
+                     "data": {"type": "bytes"},
+                     "file_name": {"type": "string"}}
+######## mindrecord_schema end ##########
+
+######## Frozen code begin ##########
+with open('mr_argument.pickle', 'rb') as mindrecord_argument_file_handle:
+    ARG_LIST = pickle.load(mindrecord_argument_file_handle)
+######## Frozen code end ##########
+
+parser = argparse.ArgumentParser(description='Mind record imagenet example')
+parser.add_argument('--label_file', type=str, default="", help='label file')
+parser.add_argument('--image_dir', type=str, default="", help='images directory')
+
+######## Frozen code begin ##########
+args = parser.parse_args(ARG_LIST)
+print(args)
+######## Frozen code end ##########
+
+
+def _user_defined_private_func():
+    """
+    Internal function for tasks list
+
+    Return:
+       tasks list
+    """
+    if not os.path.exists(args.label_file):
+        raise IOError("map file {} not exists".format(args.label_file))
+
+    label_dict = {}
+    with open(args.label_file) as file_handle:
+        line = file_handle.readline()
+        while line:
+            labels = line.split(" ")
+            label_dict[labels[1]] = labels[0]
+            line = file_handle.readline()
+    # get all the dir which are n02087046, n02094114, n02109525
+    dir_paths = {}
+    for item in label_dict:
+        real_path = os.path.join(args.image_dir, label_dict[item])
+        if not os.path.isdir(real_path):
+            print("{} dir is not exist".format(real_path))
+            continue
+        dir_paths[item] = real_path
+
+    if not dir_paths:
+        print("not valid image dir in {}".format(args.image_dir))
+        return {}, {}
+
+    dir_list = []
+    for label in dir_paths:
+        dir_list.append(label)
+    return dir_list, dir_paths
+
+
+dir_list_global, dir_paths_global = _user_defined_private_func()
+
+def mindrecord_task_number():
+    """
+    Get task size.
+
+    Return:
+       number of tasks
+    """
+    return len(dir_list_global)
+
+
+def mindrecord_dict_data(task_id):
+    """
+    Get data dict.
+
+    Yields:
+        data (dict): data row which is dict.
+    """
+
+    # get the filename, label and image binary as a dict
+    label = dir_list_global[task_id]
+    for item in os.listdir(dir_paths_global[label]):
+        file_name = os.path.join(dir_paths_global[label], item)
+        if not item.endswith("JPEG") and not item.endswith(
+                "jpg") and not item.endswith("jpeg"):
+            print("{} file is not suffix with JPEG/jpg, skip it.".format(file_name))
+            continue
+        data = {}
+        data["file_name"] = str(file_name)
+        data["label"] = int(label)
+
+        # get the image data
+        image_file = open(file_name, "rb")
+        image_bytes = image_file.read()
+        image_file.close()
+        data["data"] = image_bytes
+        yield data

example/convert_to_mindrecord/run_imagenet.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+rm /tmp/imagenet/mr/*
+
+python writer.py --mindrecord_script imagenet \
+--mindrecord_file "/tmp/imagenet/mr/m" \
+--mindrecord_partitions 16 \
+--label_file "/tmp/imagenet/label.txt" \
+--image_dir "/tmp/imagenet/jpeg"

example/convert_to_mindrecord/run_template.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+rm /tmp/template/*
+
+python writer.py --mindrecord_script template \
+--mindrecord_file "/tmp/template/m" \
+--mindrecord_partitions 4

example/convert_to_mindrecord/template/mr_api.py

@@ -0,0 +1,73 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""
+User-defined API for MindRecord writer.
+Two API must be implemented,
+  1. mindrecord_task_number()
+       # Return number of parallel tasks. return 1 if no parallel
+  2. mindrecord_dict_data(task_id)
+       # Yield data for one task
+       # task_id is 0..N-1, if N is return value of mindrecord_task_number()
+"""
+import argparse
+import pickle
+
+# ## Parse argument
+
+with open('mr_argument.pickle', 'rb') as mindrecord_argument_file_handle:  # Do NOT change this line
+    ARG_LIST = pickle.load(mindrecord_argument_file_handle)                # Do NOT change this line
+parser = argparse.ArgumentParser(description='Mind record api template')   # Do NOT change this line
+
+# ## Your arguments below
+# parser.add_argument(...)
+
+args = parser.parse_args(ARG_LIST)  # Do NOT change this line
+print(args)                         # Do NOT change this line
+
+
+# ## Default mindrecord vars. Comment them unless default value has to be changed.
+# mindrecord_index_fields = ['label']
+# mindrecord_header_size = 1 << 24
+# mindrecord_page_size = 1 << 25
+
+
+# define global vars here if necessary
+
+
+# ####### Your code below ##########
+mindrecord_schema = {"label": {"type": "int32"}}
+
+def mindrecord_task_number():
+    """
+    Get task size.
+
+    Return:
+       number of tasks
+    """
+    return 1
+
+
+def mindrecord_dict_data(task_id):
+    """
+    Get data dict.
+
+    Yields:
+        data (dict): data row which is dict.
+    """
+    print("task is {}".format(task_id))
+    for i in range(256):
+        data = {}
+        data['label'] = i
+        yield data

example/convert_to_mindrecord/writer.py

@@ -0,0 +1,149 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""
+######################## write mindrecord example ########################
+Write mindrecord by data dictionary:
+python writer.py --mindrecord_script /YourScriptPath ...
+"""
+import argparse
+import os
+import pickle
+import time
+from importlib import import_module
+from multiprocessing import Pool
+
+from mindspore.mindrecord import FileWriter
+
+
+def _exec_task(task_id, parallel_writer=True):
+    """
+    Execute task with specified task id
+    """
+    print("exec task {}, parallel: {} ...".format(task_id, parallel_writer))
+    imagenet_iter = mindrecord_dict_data(task_id)
+    batch_size = 2048
+    transform_count = 0
+    while True:
+        data_list = []
+        try:
+            for _ in range(batch_size):
+                data_list.append(imagenet_iter.__next__())
+                transform_count += 1
+            writer.write_raw_data(data_list, parallel_writer=parallel_writer)
+            print("transformed {} record...".format(transform_count))
+        except StopIteration:
+            if data_list:
+                writer.write_raw_data(data_list, parallel_writer=parallel_writer)
+                print("transformed {} record...".format(transform_count))
+            break
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Mind record writer')
+    parser.add_argument('--mindrecord_script', type=str, default="template",
+                        help='path where script is saved')
+
+    parser.add_argument('--mindrecord_file', type=str, default="/tmp/mindrecord",
+                        help='written file name prefix')
+
+    parser.add_argument('--mindrecord_partitions', type=int, default=1,
+                        help='number of written files')
+
+    parser.add_argument('--mindrecord_workers', type=int, default=8,
+                        help='number of parallel workers')
+
+    args = parser.parse_known_args()
+
+    args, other_args = parser.parse_known_args()
+
+    print(args)
+    print(other_args)
+
+    with open('mr_argument.pickle', 'wb') as file_handle:
+        pickle.dump(other_args, file_handle)
+
+    try:
+        mr_api = import_module(args.mindrecord_script + '.mr_api')
+    except ModuleNotFoundError:
+        raise RuntimeError("Unknown module path: {}".format(args.mindrecord_script + '.mr_api'))
+
+    num_tasks = mr_api.mindrecord_task_number()
+
+    print("Write mindrecord ...")
+
+    mindrecord_dict_data = mr_api.mindrecord_dict_data
+
+    # get number of files
+    writer = FileWriter(args.mindrecord_file, args.mindrecord_partitions)
+
+    start_time = time.time()
+
+    # set the header size
+    try:
+        header_size = mr_api.mindrecord_header_size
+        writer.set_header_size(header_size)
+    except AttributeError:
+        print("Default header size: {}".format(1 << 24))
+
+    # set the page size
+    try:
+        page_size = mr_api.mindrecord_page_size
+        writer.set_page_size(page_size)
+    except AttributeError:
+        print("Default page size: {}".format(1 << 25))
+
+    # get schema
+    try:
+        mindrecord_schema = mr_api.mindrecord_schema
+    except AttributeError:
+        raise RuntimeError("mindrecord_schema is not defined in mr_api.py.")
+
+    # create the schema
+    writer.add_schema(mindrecord_schema, "mindrecord_schema")
+
+    # add the index
+    try:
+        index_fields = mr_api.mindrecord_index_fields
+        writer.add_index(index_fields)
+    except AttributeError:
+        print("Default index fields: all simple fields are indexes.")
+
+    writer.open_and_set_header()
+
+    task_list = list(range(num_tasks))
+
+    # set number of workers
+    num_workers = args.mindrecord_workers
+
+    if num_tasks < 1:
+        num_tasks = 1
+
+    if num_workers > num_tasks:
+        num_workers = num_tasks
+
+    if num_tasks > 1:
+        with Pool(num_workers) as p:
+            p.map(_exec_task, task_list)
+    else:
+        _exec_task(0, False)
+
+    ret = writer.commit()
+
+    os.remove("{}".format("mr_argument.pickle"))
+
+    end_time = time.time()
+    print("--------------------------------------------")
+    print("END. Total time: {}".format(end_time - start_time))
+    print("--------------------------------------------")

mindspore/ccsrc/mindrecord/common/shard_pybind.cc

@@ -75,12 +75,9 @@ void BindShardWriter(py::module *m) {
     .def("set_header_size", &ShardWriter::set_header_size)
     .def("set_page_size", &ShardWriter::set_page_size)
     .def("set_shard_header", &ShardWriter::SetShardHeader)
-    .def("write_raw_data",
-         (MSRStatus(ShardWriter::*)(std::map<uint64_t, std::vector<py::handle>> &, vector<vector<uint8_t>> &, bool)) &
-           ShardWriter::WriteRawData)
-    .def("write_raw_nlp_data", (MSRStatus(ShardWriter::*)(std::map<uint64_t, std::vector<py::handle>> &,
-                                                          std::map<uint64_t, std::vector<py::handle>> &, bool)) &
-                                 ShardWriter::WriteRawData)
+    .def("write_raw_data", (MSRStatus(ShardWriter::*)(std::map<uint64_t, std::vector<py::handle>> &,
+                                                      vector<vector<uint8_t>> &, bool, bool)) &
+                             ShardWriter::WriteRawData)
     .def("commit", &ShardWriter::Commit);
 }
 

mindspore/ccsrc/mindrecord/include/shard_header.h

@@ -121,6 +121,10 @@ class ShardHeader {
 
   std::vector<std::string> SerializeHeader();
 
+  MSRStatus PagesToFile(const std::string dump_file_name);
+
+  MSRStatus FileToPages(const std::string dump_file_name);
+
  private:
   MSRStatus InitializeHeader(const std::vector<json> &headers);
 

mindspore/ccsrc/mindrecord/include/shard_writer.h

@@ -18,6 +18,7 @@
 #define MINDRECORD_INCLUDE_SHARD_WRITER_H_
 
 #include <libgen.h>
+#include <sys/file.h>
 #include <unistd.h>
 #include <algorithm>
 #include <array>
@@ -87,7 +88,7 @@ class ShardWriter {
   /// \param[in] sign validate data or not
   /// \return MSRStatus the status of MSRStatus to judge if write successfully
   MSRStatus WriteRawData(std::map<uint64_t, std::vector<json>> &raw_data, vector<vector<uint8_t>> &blob_data,
-                         bool sign = true);
+                         bool sign = true, bool parallel_writer = false);
 
   /// \brief write raw data by group size for call from python
   /// \param[in] raw_data the vector of raw json data, python-handle format
@@ -95,7 +96,7 @@ class ShardWriter {
   /// \param[in] sign validate data or not
   /// \return MSRStatus the status of MSRStatus to judge if write successfully
   MSRStatus WriteRawData(std::map<uint64_t, std::vector<py::handle>> &raw_data, vector<vector<uint8_t>> &blob_data,
-                         bool sign = true);
+                         bool sign = true, bool parallel_writer = false);
 
   /// \brief write raw data by group size for call from python
   /// \param[in] raw_data the vector of raw json data, python-handle format
@@ -103,7 +104,8 @@ class ShardWriter {
   /// \param[in] sign validate data or not
   /// \return MSRStatus the status of MSRStatus to judge if write successfully
   MSRStatus WriteRawData(std::map<uint64_t, std::vector<py::handle>> &raw_data,
-                         std::map<uint64_t, std::vector<py::handle>> &blob_data, bool sign = true);
+                         std::map<uint64_t, std::vector<py::handle>> &blob_data, bool sign = true,
+                         bool parallel_writer = false);
 
  private:
   /// \brief write shard header data to disk
@@ -201,7 +203,34 @@ class ShardWriter {
   MSRStatus CheckDataTypeAndValue(const std::string &key, const json &value, const json &data, const int &i,
                                   std::map<int, std::string> &err_raw_data);
 
+  /// \brief Lock writer and save pages info
+  int LockWriter(bool parallel_writer = false);
+
+  /// \brief Unlock writer and save pages info
+  MSRStatus UnlockWriter(int fd, bool parallel_writer = false);
+
+  /// \brief Check raw data before writing
+  MSRStatus WriteRawDataPreCheck(std::map<uint64_t, std::vector<json>> &raw_data, vector<vector<uint8_t>> &blob_data,
+                                 bool sign, int *schema_count, int *row_count);
+
+  /// \brief Get full path from file name
+  MSRStatus GetFullPathFromFileName(const std::vector<std::string> &paths);
+
+  /// \brief Open files
+  MSRStatus OpenDataFiles(bool append);
+
+  /// \brief Remove lock file
+  MSRStatus RemoveLockFile();
+
+  /// \brief Remove lock file
+  MSRStatus InitLockFile();
+
  private:
+  const std::string kLockFileSuffix = "_Locker";
+  const std::string kPageFileSuffix = "_Pages";
+  std::string lock_file_;   // lock file for parallel run
+  std::string pages_file_;  // temporary file of pages info for parallel run
+
   int shard_count_;        // number of files
   uint64_t header_size_;   // header size
   uint64_t page_size_;     // page size
@@ -211,7 +240,7 @@ class ShardWriter {
   std::vector<uint64_t> raw_data_size_;   // Raw data size
   std::vector<uint64_t> blob_data_size_;  // Blob data size
 
-  std::vector<string> file_paths_;                           // file paths
+  std::vector<std::string> file_paths_;                      // file paths
   std::vector<std::shared_ptr<std::fstream>> file_streams_;  // file handles
   std::shared_ptr<ShardHeader> shard_header_;                // shard headers
 

mindspore/ccsrc/mindrecord/io/shard_index_generator.cc

@@ -520,13 +520,16 @@ MSRStatus ShardIndexGenerator::ExecuteTransaction(const int &shard_no, const std
   for (int raw_page_id : raw_page_ids) {
     auto sql = GenerateRawSQL(fields_);
     if (sql.first != SUCCESS) {
+      MS_LOG(ERROR) << "Generate raw SQL failed";
       return FAILED;
     }
     auto data = GenerateRowData(shard_no, blob_id_to_page_id, raw_page_id, in);
     if (data.first != SUCCESS) {
+      MS_LOG(ERROR) << "Generate raw data failed";
       return FAILED;
     }
     if (BindParameterExecuteSQL(db.second, sql.second, data.second) == FAILED) {
+      MS_LOG(ERROR) << "Execute SQL failed";
       return FAILED;
     }
     MS_LOG(INFO) << "Insert " << data.second.size() << " rows to index db.";

mindspore/ccsrc/mindrecord/meta/shard_header.cc

@@ -677,5 +677,43 @@ std::pair<std::shared_ptr<Statistics>, MSRStatus> ShardHeader::GetStatisticByID(
   }
   return std::make_pair(statistics_.at(statistic_id), SUCCESS);
 }
+
+MSRStatus ShardHeader::PagesToFile(const std::string dump_file_name) {
+  // write header content to file, dump whatever is in the file before
+  std::ofstream page_out_handle(dump_file_name.c_str(), std::ios_base::trunc | std::ios_base::out);
+  if (page_out_handle.fail()) {
+    MS_LOG(ERROR) << "Failed in opening page file";
+    return FAILED;
+  }
+
+  auto pages = SerializePage();
+  for (const auto &shard_pages : pages) {
+    page_out_handle << shard_pages << "\n";
+  }
+
+  page_out_handle.close();
+  return SUCCESS;
+}
+
+MSRStatus ShardHeader::FileToPages(const std::string dump_file_name) {
+  for (auto &v : pages_) {  // clean pages
+    v.clear();
+  }
+  // attempt to open the file contains the page in json
+  std::ifstream page_in_handle(dump_file_name.c_str());
+
+  if (!page_in_handle.good()) {
+    MS_LOG(INFO) << "No page file exists.";
+    return SUCCESS;
+  }
+
+  std::string line;
+  while (std::getline(page_in_handle, line)) {
+    ParsePage(json::parse(line));
+  }
+
+  page_in_handle.close();
+  return SUCCESS;
+}
 }  // namespace mindrecord
 }  // namespace mindspore

mindspore/mindrecord/filewriter.py

@@ -200,13 +200,24 @@ class FileWriter:
             raw_data.pop(i)
             logger.warning(v)
 
-    def write_raw_data(self, raw_data):
+    def open_and_set_header(self):
+        """
+        Open writer and set header
+
+        """
+        if not self._writer.is_open:
+            self._writer.open(self._paths)
+        if not self._writer.get_shard_header():
+            self._writer.set_shard_header(self._header)
+
+    def write_raw_data(self, raw_data, parallel_writer=False):
         """
         Write raw data and generate sequential pair of MindRecord File and \
         validate data based on predefined schema by default.
 
         Args:
            raw_data (list[dict]): List of raw data.
+           parallel_writer (bool, optional): Load data parallel if it equals to True (default=False).
 
         Raises:
             ParamTypeError: If index field is invalid.
@@ -225,7 +236,7 @@ class FileWriter:
             if not isinstance(each_raw, dict):
                 raise ParamTypeError('raw_data item', 'dict')
         self._verify_based_on_schema(raw_data)
-        return self._writer.write_raw_data(raw_data, True)
+        return self._writer.write_raw_data(raw_data, True, parallel_writer)
 
     def set_header_size(self, header_size):
         """

mindspore/mindrecord/shardwriter.py

@@ -135,7 +135,7 @@ class ShardWriter:
     def get_shard_header(self):
         return self._header
 
-    def write_raw_data(self, data, validate=True):
+    def write_raw_data(self, data, validate=True, parallel_writer=False):
         """
         Write raw data of cv dataset.
 
@@ -145,6 +145,7 @@ class ShardWriter:
         Args:
            data (list[dict]): List of raw data.
            validate (bool, optional): verify data according schema if it equals to True.
+           parallel_writer (bool, optional): Load data parallel if it equals to True.
 
         Returns:
             MSRStatus, SUCCESS or FAILED.
@@ -165,7 +166,7 @@ class ShardWriter:
             if row_raw:
                 raw_data.append(row_raw)
         raw_data = {0: raw_data} if raw_data else {}
-        ret = self._writer.write_raw_data(raw_data, blob_data, validate)
+        ret = self._writer.write_raw_data(raw_data, blob_data, validate, parallel_writer)
         if ret != ms.MSRStatus.SUCCESS:
             logger.error("Failed to write dataset.")
             raise MRMWriteDatasetError