Refactor tests (open-mmlab#283)

* fix sdk model's pipeline.json

* resize INT64 mask

* refactor unit tests

* fix api in model.h

* remove 'customs' from meta info

* fix zip model

* fix clang-format issue

* put tc on each backend into a SECTION

* change SECTION title

* add DYNAMIC_SECTION for capi unit test

* change 'devices' to 'device_names'

* change trt to tensorrt

* remove uncessary check

* add color_type 'color_ignore_orientation' which is used in ocr

* 'min_width', 'max_width' and 'backend' might be null in pipeline config

* fix clang-format issue

* remove useless code

.gitignore

@@ -128,3 +128,6 @@ work_dirs/
 
 # install directory
 /install
+
+# the generated header files
+/tests/test_csrc/test_define.h

csrc/apis/c/model.h

@@ -9,6 +9,6 @@ int mmdeploy_model_create_by_path(const char* path, mm_model_t* model);
 
 int mmdeploy_model_create(const void* buffer, int size, mm_model_t* model);
 
-void mmdeploy_model_destroy(mm_model_t* model);
+void mmdeploy_model_destroy(mm_model_t model);
 
 #endif // MMDEPLOY_SRC_APIS_C_MODEL_H_

csrc/codebase/mmocr/resize_ocr.cpp

@@ -19,12 +19,18 @@ class ResizeOCRImpl : public Module {
  public:
  explicit ResizeOCRImpl(const Value& args) noexcept {
  height_ = args.value("height", height_);
- min_width_ = args.value("min_width", min_width_);
- max_width_ = args.value("max_width", max_width_);
+ min_width_ = args.contains("min_width") && args["min_width"].is_number_integer()
+ ? args["min_width"].get<int>()
+ : min_width_;
+ max_width_ = args.contains("max_width") && args["max_width"].is_number_integer()
+ ? args["max_width"].get<int>()
+ : max_width_;
  keep_aspect_ratio_ = args.value("keep_aspect_ratio", keep_aspect_ratio_);
+ backend_ = args.contains("backend") && args["backend"].is_string()
+ ? args["backend"].get<string>()
+ : backend_;
  img_pad_value_ = args.value("img_pad_value", img_pad_value_);
  width_downsample_ratio_ = args.value("width_downsample_ratio", width_downsample_ratio_);
- backend_ = args.value("backend", backend_);
  stream_ = args["context"]["stream"].get<Stream>();
  }
 

csrc/codebase/mmseg/segment.cpp

@@ -9,60 +9,66 @@
 
 namespace mmdeploy::mmseg {
 
-static Result<void> VisualizeMask(const std::string &image_name, const Tensor &mask, int height,
- int width, Stream &stream) {
- Device cpu_device{"cpu"};
- OUTCOME_TRY(auto host_mask, MakeAvailableOnDevice(mask, cpu_device, stream));
- OUTCOME_TRY(stream.Wait());
- // cv::Mat mask_image(height, width, CV_32SC1, host_mask.data<int>());
- // cv::imwrite(image_name + ".png", mask_image * 10);
- // ofstream ofs(image_name + ".data");
- // auto _data_ptr = host_mask.data<int>();
- // for (auto i = 0; i < height; ++i) {
- // for (auto j = 0; j < width; ++j) {
- // ofs << *_data_ptr++ << ", ";
- // }
- // ofs << "\n";
- // }
- return success();
-}
-
 class ResizeMask : public MMSegmentation {
  public:
  explicit ResizeMask(const Value &cfg) : MMSegmentation(cfg) {
- classes_ = cfg["params"]["num_classes"].get<int>();
+ try {
+ classes_ = cfg["params"]["num_classes"].get<int>();
+ } catch (const std::exception &e) {
+ ERROR("no ['params']['num_classes'] is specified in cfg: {}", cfg);
+ throw_exception(eInvalidArgument);
+ }
  }
 
  Result<Value> operator()(const Value &preprocess_result, const Value &inference_result) {
  DEBUG("preprocess: {}\ninference: {}", preprocess_result, inference_result);
 
  auto mask = inference_result["output"].get<Tensor>();
- INFO("tensor.name: {}, tensor.shape: {}", mask.name(), mask.shape());
- assert(mask.data_type() == DataType::kINT32);
+ INFO("tensor.name: {}, tensor.shape: {}, tensor.data_type: {}", mask.name(), mask.shape(),
+ mask.data_type());
+ assert(mask.data_type() == DataType::kINT32 || mask.data_type() == DataType::kINT64);
  assert(mask.shape(0) == 1);
  assert(mask.shape(1) == 1);
 
  auto height = (int)mask.shape(2);
  auto width = (int)mask.shape(3);
  auto input_height = preprocess_result["img_metas"]["ori_shape"][1].get<int>();
  auto input_width = preprocess_result["img_metas"]["ori_shape"][2].get<int>();
- if (height == input_height && width == input_width) {
- SegmentorOutput output{mask, input_height, input_width, classes_};
- return to_value(output);
+ Device host{"cpu"};
+ OUTCOME_TRY(auto host_tensor, MakeAvailableOnDevice(mask, host, stream_));
+ stream_.Wait().value();
+ if (mask.data_type() == DataType::kINT64) {
+ // change kINT64 to 2 INT32
+ TensorDesc desc{.device = host_tensor.device(),
+ .data_type = DataType::kINT32,
+ .shape = {1, 2, height, width},
+ .name = host_tensor.name()};
+ Tensor _host_tensor(desc, mask.buffer());
+ return MaskResize(_host_tensor, input_height, input_width);
  } else {
- Device host{"cpu"};
-
- OUTCOME_TRY(auto host_tensor, MakeAvailableOnDevice(mask, host, stream_));
- host_tensor.Reshape({1, height, width, 1});
- auto mat = cpu::Tensor2CVMat(host_tensor);
- auto dst = cpu::Resize(mat, input_height, input_width, "nearest");
- auto output_tensor = cpu::CVMat2Tensor(dst);
+ return MaskResize(host_tensor, input_height, input_width);
+ }
+ }
 
- SegmentorOutput output{output_tensor, input_height, input_width, classes_};
+ private:
+ Result<Value> MaskResize(Tensor &tensor, int dst_height, int dst_width) {
+ auto channel = tensor.shape(1);
+ auto height = tensor.shape(2);
+ auto width = tensor.shape(3);
 
- // OUTCOME_TRY(
- // VisualizeMask("resize_mask", output_tensor, input_height, input_width,
- // stream_));
+ // reshape tensor to convert it to cv::Mat
+ tensor.Reshape({1, height, width, channel});
+ auto mat = cpu::Tensor2CVMat(tensor);
+ auto dst = cpu::Resize(mat, dst_height, dst_width, "nearest");
+ if (channel == 1) {
+ auto output_tensor = cpu::CVMat2Tensor(dst);
+ SegmentorOutput output{output_tensor, dst_height, dst_width, classes_};
+ return to_value(output);
+ } else {
+ cv::Mat _dst;
+ cv::extractChannel(dst, _dst, 0);
+ auto output_tensor = cpu::CVMat2Tensor(_dst);
+ SegmentorOutput output{output_tensor, dst_height, dst_width, classes_};
  return to_value(output);
  }
  }

csrc/core/model.h

@@ -27,8 +27,7 @@ struct model_meta_info_t {
 struct deploy_meta_info_t {
  std::string version;
  std::vector<model_meta_info_t> models;
- std::vector<std::string> customs;
- MMDEPLOY_ARCHIVE_MEMBERS(version, models, customs);
+ MMDEPLOY_ARCHIVE_MEMBERS(version, models);
 };
 
 class ModelImpl;

csrc/model/zip_model_impl.cpp

@@ -1,7 +1,5 @@
 // Copyright (c) OpenMMLab. All rights reserved.
 
-#include <string.h>
-
 #include <fstream>
 #include <map>
 
@@ -10,6 +8,13 @@
 #include "core/model.h"
 #include "core/model_impl.h"
 #include "zip.h"
+#if __GNUC__ >= 8
+#include <filesystem>
+namespace fs = std::filesystem;
+#else
+#include <experimental/filesystem>
+namespace fs = std::experimental::filesystem;
+#endif
 
 using nlohmann::json;
 
@@ -62,6 +67,7 @@ class ZipModelImpl : public ModelImpl {
  Result<std::string> ReadFile(const std::string& file_path) const override {
  int ret = 0;
  int index = -1;
+
  auto iter = file_index_.find(file_path);
  if (iter == file_index_.end()) {
  ERROR("cannot find file {} under dir {}", file_path.c_str(), root_dir_.c_str());
@@ -103,16 +109,20 @@ class ZipModelImpl : public ModelImpl {
  Result<void> InitZip() {
  int files = zip_get_num_files(zip_);
  INFO("there are {} files in sdk model file", files);
-
+ if (files == 0) {
+ return Status(eFail);
+ }
  for (int i = 0; i < files; ++i) {
  struct zip_stat stat;
  zip_stat_init(&stat);
  zip_stat_index(zip_, i, 0, &stat);
- if (stat.name[strlen(stat.name) - 1] == '/') {
+ fs::path path(stat.name);
+ auto file_name = path.filename().string();
+ if (file_name == ".") {
  DEBUG("{}-th file name is: {}， which is a directory", i, stat.name);
  } else {
  DEBUG("{}-th file name is: {}， which is a file", i, stat.name);
- file_index_[stat.name] = i;
+ file_index_[file_name] = i;
  }
  }
  return success();

csrc/net/trt/trt_net.cpp

@@ -212,7 +212,7 @@ Result<void> TRTNet::ForwardAsync(Event* event) { return Status(eNotSupported);
 
 class TRTNetCreator : public Creator<Net> {
  public:
- const char* GetName() const override { return "trt"; }
+ const char* GetName() const override { return "tensorrt"; }
  int GetVersion() const override { return 0; }
  std::unique_ptr<Net> Create(const Value& args) override {
  auto p = std::make_unique<TRTNet>();

csrc/preprocess/transform/load.cpp

@@ -38,7 +38,9 @@ Result<Value> PrepareImageImpl::Process(const Value& input) {
  Value output = input;
 
  Mat src_mat = input["ori_img"].get<Mat>();
- auto res = (arg_.color_type == "color" ? ConvertToBGR(src_mat) : ConvertToGray(src_mat));
+ auto res = (arg_.color_type == "color" || arg_.color_type == "color_ignore_orientation"
+ ? ConvertToBGR(src_mat)
+ : ConvertToGray(src_mat));
 
  OUTCOME_TRY(auto tensor, std::move(res));
 

demo/csrc/config/resnet50_ort/pipeline.json

@@ -80,8 +80,8 @@
  {
  "name": "postprocess",
  "type": "Task",
- "module": "MMClsPostprocess",
- "postprocess_type": "SoftmaxPost",
+ "module": "mmcls",
+ "component": "LinearClsHead",
  "input": [
  "data",
  "cls_res"

demo/csrc/config/retinanet_ort/pipeline.json

@@ -81,18 +81,21 @@
  }
  },
  {
- "name": "retinanet_post",
+ "name": "postprocess",
  "type": "Task",
- "module": "MMDetPostprocess",
- "postprocess_type": "SingleStagePost",
+ "module": "mmdet",
+ "component": "ResizeBBox",
  "input": [
  "prep_res",
  "infer_res"
  ],
  "output": [
  "bboxes"
  ],
- "score_thr": 0.3
+ "params": {
+ "score_thr": 0.3,
+ "min_bbox_size": 10
+ }
  }
  ]
  }