!8538 fix get_num_class in imagefolder

From: @ziruiwu
Reviewed-by: @robingrosman,@nsyca
Signed-off-by: @nsyca

mindspore/ccsrc/minddata/dataset/api/python/bindings/dataset/engine/datasetops/source/bindings.cc

@@ -91,11 +91,17 @@ PYBIND_REGISTER(CocoOp, 1, ([](const py::module *m) {
 
 PYBIND_REGISTER(ImageFolderOp, 1, ([](const py::module *m) {
                   (void)py::class_<ImageFolderOp, DatasetOp, std::shared_ptr<ImageFolderOp>>(*m, "ImageFolderOp")
-                    .def_static("get_num_rows_and_classes", [](const std::string &path) {
-                      int64_t count = 0, num_classes = 0;
-                      THROW_IF_ERROR(
-                        ImageFolderOp::CountRowsAndClasses(path, std::set<std::string>{}, &count, &num_classes));
-                      return py::make_tuple(count, num_classes);
+                    .def_static("get_num_rows",
+                                [](const std::string &path) {
+                                  int64_t count = 0;
+                                  THROW_IF_ERROR(ImageFolderOp::CountRowsAndClasses(path, {}, &count, nullptr, {}));
+                                  return count;
+                                })
+                    .def_static("get_num_classes", [](const std::string &path,
+                                                      const std::map<std::string, int32_t> class_index) {
+                      int64_t num_classes = 0;
+                      THROW_IF_ERROR(ImageFolderOp::CountRowsAndClasses(path, {}, nullptr, &num_classes, class_index));
+                      return num_classes;
                     });
                 }));
 

mindspore/ccsrc/minddata/dataset/engine/datasetops/source/image_folder_op.cc

@@ -15,7 +15,7 @@
  */
 #include "minddata/dataset/engine/datasetops/source/image_folder_op.h"
 #include <fstream>
-#include <iomanip>
+#include <unordered_set>
 #include "utils/ms_utils.h"
 #include "minddata/dataset/core/config_manager.h"
 #include "minddata/dataset/core/tensor_shape.h"
@@ -280,7 +280,7 @@ Status ImageFolderOp::GetClassIds(std::map<int32_t, std::vector<int64_t>> *cls_i
       RETURN_STATUS_UNEXPECTED("No images found in dataset, please check if Op read images successfully or not.");
     } else {
       RETURN_STATUS_UNEXPECTED(
-        "Map for storaging image-index pair is nullptr or has been set in other place,"
+        "Map containing image-index pair is nullptr or has been set in other place,"
         "it must be empty before using GetClassIds.");
     }
   }
@@ -294,14 +294,14 @@ Status ImageFolderOp::GetClassIds(std::map<int32_t, std::vector<int64_t>> *cls_i
 }
 
 // Worker Entry for pre-scanning all the folders and do the 1st level shuffle
-// Worker pull a file name from mFoldernameQueue (which is a Queue), walks all the images under that foldername
+// Worker pull a file name from folder_name_queue_ (which is a Queue), walks all the images under that foldername
 // After walking is complete, sort all the file names (relative path to all jpeg files under the same directory )
 // (Sort is automatically conducted using a set which is implemented using a Red-Black Tree)
 // Add the sorted filenames in to a queue. The make a pair (foldername, queue<filenames>*),
 // foldername is used for 2nd level sorting.
 // FYI: 1st level sorting: sort all images under the same directory.
 // FYI: 2nd level sorting: sort all folder names
-// push this pair to mImagenameQueue (which is again a Queue)
+// push this pair to image_name_queue (which is again a Queue)
 Status ImageFolderOp::PrescanWorkerEntry(int32_t worker_id) {
   TaskManager::FindMe()->Post();
   std::string folder_name;
@@ -334,7 +334,7 @@ Status ImageFolderOp::PrescanWorkerEntry(int32_t worker_id) {
   return Status::OK();
 }
 
-// This helper function recursively walks all foldernames, and send each foldername to mFoldernameQueue
+// This helper function recursively walks all folder_paths, and send each foldername to folder_name_queue_
 // if mRecursive == false, don't go into folder of folders
 Status ImageFolderOp::RecursiveWalkFolder(Path *dir) {
   std::shared_ptr<Path::DirIterator> dir_itr = Path::DirIterator::OpenDirectory(dir);
@@ -355,7 +355,7 @@ Status ImageFolderOp::RecursiveWalkFolder(Path *dir) {
 }
 
 // A thread that calls RecursiveWalkFolder
-Status ImageFolderOp::startAsyncWalk() {
+Status ImageFolderOp::StartAsyncWalk() {
   TaskManager::FindMe()->Post();
   Path dir(folder_path_);
   if (dir.Exists() == false || dir.IsDirectory() == false) {
@@ -363,8 +363,8 @@ Status ImageFolderOp::startAsyncWalk() {
   }
   dirname_offset_ = folder_path_.length();
   RETURN_IF_NOT_OK(RecursiveWalkFolder(&dir));
-  // send out num_workers_ end signal to mFoldernameQueue, 1 for each worker.
-  // Upon receiving end Signal, worker quits and set another end Signal to mImagenameQueue.
+  // send out num_workers_ end signal to folder_name_queue_, 1 for each worker.
+  // Upon receiving end Signal, worker quits and set another end Signal to image_name_queue.
   for (int32_t ind = 0; ind < num_workers_; ++ind) {
     RETURN_IF_NOT_OK(folder_name_queue_->EmplaceBack(""));  // end signal
   }
@@ -372,19 +372,17 @@ Status ImageFolderOp::startAsyncWalk() {
 }
 
 Status ImageFolderOp::LaunchThreadsAndInitOp() {
-  if (tree_ == nullptr) {
-    RETURN_STATUS_UNEXPECTED("Pipeline init failed, Execution tree not set.");
-  }
+  RETURN_UNEXPECTED_IF_NULL(tree_);
   // Registers QueueList and individual Queues for interrupt services
   RETURN_IF_NOT_OK(io_block_queues_.Register(tree_->AllTasks()));
   RETURN_IF_NOT_OK(folder_name_queue_->Register(tree_->AllTasks()));
   RETURN_IF_NOT_OK(image_name_queue_->Register(tree_->AllTasks()));
   RETURN_IF_NOT_OK(wait_for_workers_post_.Register(tree_->AllTasks()));
   // The following code launch 3 threads group
-  // 1) A thread that walks all folders and push the folder names to a util:Queue mFoldernameQueue.
-  // 2) Workers that pull foldername from mFoldernameQueue, walk it and return the sorted images to mImagenameQueue
+  // 1) A thread that walks all folders and push the folder names to a util:Queue folder_name_queue_.
+  // 2) Workers that pull foldername from folder_name_queue_, walk it and return the sorted images to image_name_queue
   // 3) Launch main workers that load DataBuffers by reading all images
-  RETURN_IF_NOT_OK(tree_->AllTasks()->CreateAsyncTask("walk dir", std::bind(&ImageFolderOp::startAsyncWalk, this)));
+  RETURN_IF_NOT_OK(tree_->AllTasks()->CreateAsyncTask("walk dir", std::bind(&ImageFolderOp::StartAsyncWalk, this)));
   RETURN_IF_NOT_OK(
     tree_->LaunchWorkers(num_workers_, std::bind(&ImageFolderOp::PrescanWorkerEntry, this, std::placeholders::_1)));
   RETURN_IF_NOT_OK(
@@ -397,42 +395,53 @@ Status ImageFolderOp::LaunchThreadsAndInitOp() {
 }
 
 Status ImageFolderOp::CountRowsAndClasses(const std::string &path, const std::set<std::string> &exts, int64_t *num_rows,
-                                          int64_t *num_classes, int64_t dev_id, int64_t num_dev) {
+                                          int64_t *num_classes, std::map<std::string, int32_t> class_index) {
   Path dir(path);
   std::string err_msg = "";
   int64_t row_cnt = 0;
   err_msg += (dir.Exists() == false || dir.IsDirectory() == false)
-               ? "Invalid parameter, image folde path is invalid or not set, path: " + path
+               ? "Invalid parameter, image folder path is invalid or not set, path: " + path
                : "";
   err_msg +=
-    (num_classes == nullptr || num_rows == nullptr) ? "Invalid parameter, num_class or num_rows cannot be null.\n" : "";
-  err_msg += (dev_id >= num_dev || num_dev <= 0)
-               ? "Invalid parameter, num_shard must be greater than shard_id and greater than 0, got num_shard: " +
-                   std::to_string(num_dev) + ", shard_id: " + std::to_string(dev_id) + ".\n"
-               : "";
+    (num_classes == nullptr && num_rows == nullptr) ? "Invalid parameter, num_class and num_rows are null.\n" : "";
   if (err_msg.empty() == false) {
     RETURN_STATUS_UNEXPECTED(err_msg);
   }
-  std::queue<std::string> foldernames;
+  std::queue<std::string> folder_paths;
   std::shared_ptr<Path::DirIterator> dir_itr = Path::DirIterator::OpenDirectory(&dir);
+  std::unordered_set<std::string> folder_names;
   while (dir_itr->hasNext()) {
     Path subdir = dir_itr->next();
     if (subdir.IsDirectory()) {
-      foldernames.push(subdir.toString());
+      folder_paths.push(subdir.toString());
+      if (!class_index.empty()) folder_names.insert(subdir.Basename());
     }
   }
-  (*num_classes) = foldernames.size();
-  while (foldernames.empty() == false) {
-    Path subdir(foldernames.front());
+  if (num_classes != nullptr) {
+    // if class index is empty, get everything on disk
+    if (class_index.empty()) {
+      *num_classes = folder_paths.size();
+    } else {
+      for (const auto &p : class_index) {
+        CHECK_FAIL_RETURN_UNEXPECTED(folder_names.find(p.first) != folder_names.end(),
+                                     "folder: " + p.first + " doesn't exist in " + path + " .");
+      }
+      (*num_classes) = class_index.size();
+    }
+  }
+  // return here if only num_class is needed
+  RETURN_OK_IF_TRUE(num_rows == nullptr);
+  while (folder_paths.empty() == false) {
+    Path subdir(folder_paths.front());
     dir_itr = Path::DirIterator::OpenDirectory(&subdir);
     while (dir_itr->hasNext()) {
       if (exts.empty() || exts.find(subdir.Extension()) != exts.end()) {
         ++row_cnt;
       }
     }
-    foldernames.pop();
+    folder_paths.pop();
   }
-  (*num_rows) = (row_cnt / num_dev) + (row_cnt % num_dev == 0 ? 0 : 1);
+  (*num_rows) = row_cnt;
   return Status::OK();
 }
 
@@ -460,9 +469,12 @@ Status ImageFolderOp::GetDatasetSize(int64_t *dataset_size) {
     *dataset_size = dataset_size_;
     return Status::OK();
   }
-  int64_t sample_size, num_rows, num_classes;
+  int64_t sample_size, num_rows;
   num_rows = num_rows_;
-  if (num_rows_ <= 0) RETURN_IF_NOT_OK(CountRowsAndClasses(folder_path_, extensions_, &num_rows, &num_classes));
+  if (num_rows_ <= 0) {
+    // GetDatasetSize will not be impacted by class_index_
+    RETURN_IF_NOT_OK(CountRowsAndClasses(folder_path_, extensions_, &num_rows, nullptr, {}));
+  }
   sample_size = sampler_->GetNumSamples();
   *dataset_size = sample_size > 0 ? std::min(num_rows, sample_size) : num_rows;
   dataset_size_ = *dataset_size;
@@ -475,8 +487,7 @@ Status ImageFolderOp::GetNumClasses(int64_t *num_classes) {
     *num_classes = num_classes_;
     return Status::OK();
   }
-  int64_t num_rows = num_rows_;
-  RETURN_IF_NOT_OK(CountRowsAndClasses(folder_path_, extensions_, &num_rows, num_classes));
+  RETURN_IF_NOT_OK(CountRowsAndClasses(folder_path_, extensions_, nullptr, num_classes, class_index_));
   num_classes_ = *num_classes;
   return Status::OK();
 }

mindspore/ccsrc/minddata/dataset/engine/datasetops/source/image_folder_op.h

@@ -205,7 +205,7 @@ class ImageFolderOp : public ParallelOp, public RandomAccessOp {
   // returned by this function may not be consistent with what image_folder_op is going to return
   // user this at your own risk!
   static Status CountRowsAndClasses(const std::string &path, const std::set<std::string> &exts, int64_t *num_rows,
-                                    int64_t *num_classes, int64_t dev_id = 0, int64_t num_dev = 1);
+                                    int64_t *num_classes, std::map<std::string, int32_t> class_index);
 
   // Base-class override for NodePass visitor acceptor.
   // @param p - Pointer to the NodePass to be accepted.
@@ -251,7 +251,7 @@ class ImageFolderOp : public ParallelOp, public RandomAccessOp {
 
   // start walking of all dirs
   // @return
-  Status startAsyncWalk();
+  Status StartAsyncWalk();
 
   // Called first when function is called
   // @return

mindspore/ccsrc/minddata/dataset/engine/ir/datasetops/batch_node.cc

@@ -81,14 +81,16 @@ std::vector<std::shared_ptr<DatasetOp>> BatchNode::Build() {
   std::vector<std::shared_ptr<DatasetOp>> node_ops;
 
 #ifdef ENABLE_PYTHON
-  node_ops.push_back(std::make_shared<BatchOp>(batch_size_, drop_remainder_, pad_, connector_que_size_, num_workers_,
-                                               in_col_names_, out_col_names_, batch_size_func_, batch_map_func_,
-                                               pad_map_));
-  // need to insert a project when per_batch_func changes the number of columns
+  // if col_order_ isn't empty, then a project node needs to be attached after batch node. (same as map)
+  // this means project_node needs to be the parent of batch_node. this means node_ops = [project_node, batch_node]
   if (!col_order_.empty()) {
     auto project_op = std::make_shared<ProjectOp>(col_order_);
     node_ops.push_back(project_op);
   }
+
+  node_ops.push_back(std::make_shared<BatchOp>(batch_size_, drop_remainder_, pad_, connector_que_size_, num_workers_,
+                                               in_col_names_, out_col_names_, batch_size_func_, batch_map_func_,
+                                               pad_map_));
 #else
   node_ops.push_back(std::make_shared<BatchOp>(batch_size_, drop_remainder_, pad_, connector_que_size_, num_workers_,
                                                in_col_names_, pad_map_));

mindspore/dataset/engine/datasets.py

@@ -2891,7 +2891,7 @@ class ImageFolderDataset(MappableDataset):
             Number, number of batches.
         """
         if self.dataset_size is None:
-            num_rows = ImageFolderOp.get_num_rows_and_classes(self.dataset_dir)[0]
+            num_rows = ImageFolderOp.get_num_rows(self.dataset_dir)
             self.dataset_size = get_num_rows(num_rows, self.num_shards)
             rows_from_sampler = self._get_sampler_dataset_size()
             if rows_from_sampler is not None and rows_from_sampler < self.dataset_size:
@@ -2905,7 +2905,8 @@ class ImageFolderDataset(MappableDataset):
         Return:
             Number, number of classes.
         """
-        return ImageFolderOp.get_num_rows_and_classes(self.dataset_dir)[1]
+        class_index = self.class_indexing if self.class_indexing else {}
+        return ImageFolderOp.get_num_classes(self.dataset_dir, class_index)
 
     def is_shuffled(self):
         if self.shuffle_level is None:

tests/ut/cpp/dataset/c_api_dataset_save.cc

@@ -65,7 +65,9 @@ TEST_F(MindDataTestPipeline, TestSaveCifar10AndLoad) {
   std::string temp_file = datasets_root_path_ + "/testCifar10Data/mind.mind";
   std::string temp_file_db = datasets_root_path_ + "/testCifar10Data/mind.mind.db";
   bool rc = ds->Save(temp_file);
-  EXPECT_EQ(rc, true);
+  // if save fails, no need to continue the execution
+  // save could fail if temp_file already exists
+  ASSERT_EQ(rc, true);
 
   // Stage 3: Load dataset from file output by stage 2
   // Create a MindData Dataset

tests/ut/cpp/dataset/c_api_datasets_test.cc

@@ -305,3 +305,21 @@ TEST_F(MindDataTestPipeline, TestMnistFailWithNullSamplerFail) {
   // Expect failure: invalid Mnist input, sampler cannot be nullptr
   EXPECT_EQ(iter, nullptr);
 }
+
+TEST_F(MindDataTestPipeline, TestImageFolderClassIndexDatasetSize) {
+  std::string folder_path = datasets_root_path_ + "/testPK/data";
+  std::map<std::string, int32_t> class_index;
+  class_index["class1"] = 111;
+  class_index["class2"] = 333;
+  auto ds = ImageFolder(folder_path, false, RandomSampler(), {}, class_index);
+  EXPECT_EQ(ds->GetNumClasses(), 2);
+}
+
+TEST_F(MindDataTestPipeline, TestImageFolderClassIndexDatasetSizeFail) {
+  std::string folder_path = datasets_root_path_ + "/testPK/data";
+  std::map<std::string, int32_t> class_index;
+  class_index["class1"] = 111;
+  class_index["wrong class"] = 333;
+  auto ds = ImageFolder(folder_path, false, RandomSampler(), {}, class_index);
+  EXPECT_EQ(ds->GetNumClasses(), -1);
+}

tests/ut/cpp/dataset/image_folder_op_test.cc

@@ -38,9 +38,9 @@
 namespace common = mindspore::common;
 
 using namespace mindspore::dataset;
-using mindspore::MsLogLevel::ERROR;
-using mindspore::ExceptionType::NoExceptionType;
 using mindspore::LogStream;
+using mindspore::ExceptionType::NoExceptionType;
+using mindspore::MsLogLevel::ERROR;
 
 std::shared_ptr<BatchOp> Batch(int batch_size = 1, bool drop = false, int rows_per_buf = 2);
 
@@ -54,14 +54,17 @@ std::shared_ptr<ImageFolderOp> ImageFolder(int64_t num_works, int64_t rows, int6
   std::shared_ptr<ImageFolderOp> so;
   ImageFolderOp::Builder builder;
   Status rc = builder.SetNumWorkers(num_works)
-                     .SetImageFolderDir(path)
-                     .SetRowsPerBuffer(rows)
-                     .SetOpConnectorSize(conns)
-                     .SetExtensions({".jpg", ".JPEG"})
-                     .SetSampler(std::move(sampler))
-                     .SetClassIndex(map)
-                     .SetDecode(decode)
-                     .Build(&so);
+                .SetImageFolderDir(path)
+                .SetRowsPerBuffer(rows)
+                .SetOpConnectorSize(conns)
+                .SetExtensions({".jpg", ".JPEG"})
+                .SetSampler(std::move(sampler))
+                .SetClassIndex(map)
+                .SetDecode(decode)
+                .Build(&so);
+  if (rc.IsError()) {
+    MS_LOG(ERROR) << "Fail to build ImageFolderOp: " << rc.ToString() << "\n";
+  }
   return so;
 }
 
@@ -166,9 +169,9 @@ TEST_F(MindDataTestImageFolderSampler, TestSequentialImageFolderWithRepeatBatch)
   auto tree = Build({ImageFolder(16, 2, 32, folder_path, false), Repeat(2), Batch(11)});
   tree->Prepare();
   int32_t res[4][11] = {{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
-                         {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1},
-                         {2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2},
-                         {3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3}};
+                        {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+                        {2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2},
+                        {3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3}};
   Status rc = tree->Launch();
   if (rc.IsError()) {
     MS_LOG(ERROR) << "Return code error detected during tree launch: " << common::SafeCStr(rc.ToString()) << ".";
@@ -184,7 +187,7 @@ TEST_F(MindDataTestImageFolderSampler, TestSequentialImageFolderWithRepeatBatch)
       Create1DTensor(&label, 11, reinterpret_cast<unsigned char *>(res[i % 4]), DataType::DE_INT32);
       EXPECT_TRUE((*label) == (*tensor_map["label"]));
       MS_LOG(DEBUG) << "row: " << i << " " << tensor_map["image"]->shape() << " (*label):" << (*label)
-                << " *tensor_map[label]: " << *tensor_map["label"] << std::endl;
+                    << " *tensor_map[label]: " << *tensor_map["label"] << std::endl;
       i++;
       di.GetNextAsMap(&tensor_map);
     }
@@ -373,8 +376,8 @@ TEST_F(MindDataTestImageFolderSampler, TestImageFolderDecode) {
     while (tensor_map.size() != 0) {
       tensor_map["label"]->GetItemAt<int32_t>(&label, {});
       EXPECT_TRUE(label == res[i / 11]);
-      EXPECT_TRUE(
-        tensor_map["image"]->shape() == TensorShape({2268, 4032, 3}));  // verify shapes are correct after decode
+      EXPECT_TRUE(tensor_map["image"]->shape() ==
+                  TensorShape({2268, 4032, 3}));  // verify shapes are correct after decode
       MS_LOG(DEBUG) << "row: " << i << "\t" << tensor_map["image"]->shape() << "label:" << label << "\n";
       i++;
       di.GetNextAsMap(&tensor_map);

tests/ut/python/dataset/test_get_size.py

@@ -158,13 +158,23 @@ def test_imagefolder():
     assert data.get_dataset_size() == 10
     assert data.num_classes() == 4
 
+    data = ds.ImageFolderDataset("../data/dataset/testPK/data/", class_indexing={"class1": 1, "class2": 22})
+    assert data.num_classes() == 2
+
+    data = ds.ImageFolderDataset("../data/dataset/testPK/data/", class_indexing={"class1": 1, "wrong name": 22})
+    err_msg = ""
+    try:
+        data.num_classes()
+    except RuntimeError as e:
+        err_msg = str(e)
+    assert "wrong name doesn't exist" in err_msg
+
 
 if __name__ == '__main__':
-    # test_compare_v1_and_2()
-    # test_imagefolder()
-    # test_manifest()
+    test_manifest()
     test_case1()
-    # test_case2()
-    # test_case3()
-    # test_case4()
-    # test_case5()
+    test_case2()
+    test_case3()
+    test_case4()
+    test_case5()
+    test_imagefolder()