update

Author: lzhangzz

csrc/mmdeploy/net/trt/trt_net.cpp

@@ -8,6 +8,7 @@
 #include "mmdeploy/core/model.h"
 #include "mmdeploy/core/module.h"
 #include "mmdeploy/core/utils/formatter.h"
+#include "mmdeploy/device/cuda/cuda_device.h"
 
 namespace mmdeploy::framework {
 

csrc/mmdeploy/triton/CMakeLists.txt

@@ -61,13 +61,11 @@ FetchContent_Declare(
         repo-core
         GIT_REPOSITORY https://github.com/triton-inference-server/core.git
         GIT_TAG ${TRITON_CORE_REPO_TAG}
-        GIT_SHALLOW ON
 )
 FetchContent_Declare(
         repo-backend
         GIT_REPOSITORY https://github.com/triton-inference-server/backend.git
         GIT_TAG ${TRITON_BACKEND_REPO_TAG}
-        GIT_SHALLOW ON
 )
 FetchContent_MakeAvailable(repo-common repo-core repo-backend)
 
@@ -98,8 +96,9 @@ target_link_libraries(
 mmdeploy_load_static(triton-mmdeploy-backend MMDeployStaticModules)
 target_link_libraries(triton-mmdeploy-backend PRIVATE MMDeployLibs)
 
-mmdeploy_export(triton-mmdeploy-backend)
+set_target_properties(triton-mmdeploy-backend PROPERTIES INSTALL_RPATH "\$ORIGIN")
 
+install(TARGETS triton-mmdeploy-backend DESTINATION backend/mmdeploy)
 
 if (WIN32)
     set_target_properties(

csrc/mmdeploy/triton/convert.cpp

@@ -90,6 +90,7 @@ void ConvertSegmentation(const Value& item, std::vector<Tensor>& tensors) {
     auto desc = seg.mask.desc();
     desc.name = "mask";
     tensors.emplace_back(desc, seg.mask.buffer());
+    tensors.back().Squeeze();
   }
 }
 

csrc/mmdeploy/triton/instance_state.cpp

@@ -172,7 +172,7 @@ TRITONSERVER_Error* ModelInstanceState::Execute(TRITONBACKEND_Request** requests
     input_args.push_back(std::move(input_tensors_array));
   }
 
-  MMDEPLOY_ERROR("input: {}", input_args);
+  MMDEPLOY_DEBUG("input: {}", input_args);
 
   uint64_t compute_start_ns = 0;
   SET_TIMESTAMP(compute_start_ns);
@@ -187,23 +187,20 @@ TRITONSERVER_Error* ModelInstanceState::Execute(TRITONBACKEND_Request** requests
   SET_TIMESTAMP(compute_end_ns);
 
   std::vector<std::unique_ptr<BackendOutputResponder>> responders(request_count);
-  MMDEPLOY_ERROR("request_count {}", request_count);
+  MMDEPLOY_DEBUG("request_count {}", request_count);
   for (uint32_t request_index = 0; request_index < request_count; ++request_index) {
     responders[request_index] = std::make_unique<BackendOutputResponder>(
         &requests[request_index], 1, &response_vecs[request_index],
         model_state->TritonMemoryManager(), false, false, nullptr);
-    for (const auto& name : model_state->output_names()) {
-      MMDEPLOY_ERROR("name {}", name);
-    }
     for (size_t output_id = 0; output_id < model_state->output_names().size(); ++output_id) {
       auto output_name = model_state->output_names()[output_id];
-      MMDEPLOY_ERROR("output name {}", output_name);
+      MMDEPLOY_DEBUG("output name {}", output_name);
       auto output_data_type = model_state->output_data_types()[output_id];
       for (const auto& tensor : output_tensors[request_index]) {
         if (tensor.name() == output_name) {
           if (output_data_type != TRITONSERVER_TYPE_BYTES) {
             auto shape = tensor.shape();
-            MMDEPLOY_ERROR("name {}, shape {}", tensor.name(), shape);
+            MMDEPLOY_DEBUG("name {}, shape {}", tensor.name(), shape);
             auto memory_type = TRITONSERVER_MEMORY_CPU;
             int64_t memory_type_id = 0;
             if (not tensor.device().is_host()) {

demo/python/triton_client.py

@@ -61,6 +61,10 @@ def vis_rdet(img, bboxes, labels):
 
 
 def vis_seg(img, mask, scores):
+    if mask is not None:
+        print(f'mask {mask.shape}')
+    if scores is not None:
+        print(f'scores {scores.shape}')
     if mask is None:
         mask = np.argmax(scores, axis=0)
 
@@ -127,8 +131,9 @@ def main():
     vis = visualize(img, *map(response.as_numpy, output_names))
 
     if vis is not None:
-        cv2.imshow('', vis)
-        cv2.waitKey(0)
+        cv2.imwrite('vis.jpg', vis)
+        # cv2.imshow('', vis)
+        # cv2.waitKey(0)
 
 
 if __name__ == '__main__':