fix reshape trt condition (#34007)

PaddlePaddle · Jul 7, 2021 · 0914ff9 · 0914ff9
1 parent cb73fee
commit 0914ff9
Show file tree

Hide file tree

Showing 4 changed files with 60 additions and 18 deletions.
diff --git a/paddle/fluid/inference/goapi/README.md b/paddle/fluid/inference/goapi/README.md
@@ -11,8 +11,8 @@ Paddle Inference golang API 基于 [capi](../capi_exp) 和 cgo 实现，需要
 2. 使用`go get`获取golang paddle api
 
 ```
-# 此处使用上一步记录的CommitId，假设为76e5724
-COMMITID=76e5724
+# 此处使用上一步记录的CommitId，假设为0722297
+COMMITID=0722297
 go get -d -v github.com/paddlepaddle/paddle/paddle/fluid/inference/goapi@${COMMITID}
 ```
 
@@ -28,7 +28,7 @@ go1.15新增了`GOMODCACHE`环境变量，`go get`默认会将代码下载到`GO
 ```bash
 eval $(go env | grep GOMODCACHE)
 # 按需修改最后的goapi版本号
-cd ${GOMODCACHE}/github.com/paddlepaddle/paddle/paddle/fluid/inference/goapi\@v0.0.0-20210517084506-76e5724c16a5/
+cd ${GOMODCACHE}/github.com/paddlepaddle/paddle/paddle/fluid/inference/goapi\@v0.0.0-20210623023452-0722297d9b8c/
 ln -s ${PADDLE_C_DOWNLOAD_DIR}/paddle_inference_c_install_dir paddle_inference_c
 ```
 

diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc
@@ -692,15 +692,16 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
     if (op_type == "reshape" || op_type == "reshape2") {
       if (!desc.HasAttr("shape")) {
         return false;
-        // Paddle-TRT does not support the input tensors: Shape and ShapeTensor
-      } else if (desc.Input("Shape").size() >= 1 ||
-                 desc.Input("ShapeTensor").size() >= 1) {
+      }
+      // Paddle-TRT does not support the input tensors: Shape and ShapeTensor
+      if (desc.Input("Shape").size() >= 1 ||
+          desc.Input("ShapeTensor").size() >= 1) {
         return false;
-      } else {
-        std::vector<int> shape =
-            BOOST_GET_CONST(std::vector<int>, desc.GetAttr("shape"));
-        if (shape.size() >= nvinfer1::Dims::MAX_DIMS) return false;
       }
+      std::vector<int> shape =
+          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("shape"));
+      if (shape.size() >= nvinfer1::Dims::MAX_DIMS) return false;
+      if (!with_dynamic_shape && shape[0] == -1) return false;
     }
 
     if (op_type == "reduce_sum") {

diff --git a/python/paddle/fluid/tests/unittests/ir/inference/inference_pass_test.py b/python/paddle/fluid/tests/unittests/ir/inference/inference_pass_test.py
@@ -72,9 +72,7 @@ def _save_models(self, executor, program, scope):
                 feeded_var_names=list(self.feeds.keys()),
                 target_vars=self.fetch_list,
                 executor=executor,
-                main_program=program,
-                model_filename="model",
-                params_filename="params")
+                main_program=program)
 
         return outs
 
@@ -111,8 +109,7 @@ def _get_analysis_config(self,
         '''
         Return a new object of AnalysisConfig. 
         '''
-        config = AnalysisConfig(
-            os.path.join(self.path, "model"), os.path.join(self.path, "params"))
+        config = AnalysisConfig(self.path)
         config.disable_gpu()
         config.switch_specify_input_names(True)
         config.switch_ir_optim(True)

diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_yolo_box_op.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_yolo_box_op.py
@@ -32,8 +32,6 @@ def setUp(self):
             image_size = fluid.data(
                 name='image_size', shape=[self.bs, 2], dtype='int32')
             boxes, scores = self.append_yolobox(image, image_size)
-            scores = fluid.layers.reshape(scores, (self.bs, -1))
-            out = fluid.layers.batch_norm(scores, is_test=True)
 
         self.feeds = {
             'image': np.random.random(image_shape).astype('float32'),
@@ -43,7 +41,7 @@ def setUp(self):
         self.enable_trt = True
         self.trt_parameters = TRTYoloBoxTest.TensorRTParam(
             1 << 30, self.bs, 1, AnalysisConfig.Precision.Float32, False, False)
-        self.fetch_list = [out, boxes]
+        self.fetch_list = [scores, boxes]
 
     def set_params(self):
         self.bs = 4
@@ -72,5 +70,51 @@ def test_check_output(self):
                 PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
 
 
+class TRTYoloBoxFP16Test(InferencePassTest):
+    def setUp(self):
+        self.set_params()
+        with fluid.program_guard(self.main_program, self.startup_program):
+            image_shape = [self.bs, self.channel, self.height, self.width]
+            image = fluid.data(name='image', shape=image_shape, dtype='float32')
+            image_size = fluid.data(
+                name='image_size', shape=[self.bs, 2], dtype='int32')
+            boxes, scores = self.append_yolobox(image, image_size)
+
+        self.feeds = {
+            'image': np.random.random(image_shape).astype('float32'),
+            'image_size': np.array([[416, 416]]).astype('int32'),
+        }
+        self.enable_trt = True
+        self.trt_parameters = TRTYoloBoxFP16Test.TensorRTParam(
+            1 << 30, self.bs, 1, AnalysisConfig.Precision.Half, False, False)
+        self.fetch_list = [scores, boxes]
+
+    def set_params(self):
+        self.bs = 1
+        self.height = 13
+        self.width = 13
+        self.class_num = 1
+        self.anchors = [106, 148, 92, 300, 197, 334]
+        self.channel = 18
+        self.conf_thresh = .05
+        self.downsample_ratio = 32
+
+    def append_yolobox(self, image, image_size):
+        return fluid.layers.yolo_box(
+            x=image,
+            img_size=image_size,
+            class_num=self.class_num,
+            anchors=self.anchors,
+            conf_thresh=self.conf_thresh,
+            downsample_ratio=self.downsample_ratio)
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            use_gpu = True
+            self.check_output_with_option(use_gpu, flatten=True, rtol=1e-1)
+            self.assertTrue(
+                PassVersionChecker.IsCompatible('tensorrt_subgraph_pass'))
+
+
 if __name__ == "__main__":
     unittest.main()