fix wrong trt dim when input dim is 2 (#36614) (#36732)

* fix wrong trt dim when input dim is 2

* update leaky_relu and instance_norm converter unit test

* add instance_norm input dim check

paddle/fluid/inference/tensorrt/engine.h

@@ -116,6 +116,17 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector<T>& shape, std::string input,
             input, ShapeStr(shape)));
       }
       return nvinfer1::Dims2(shape[1], shape[2]);
+    } else if (shape.size() == 2UL) {
+      if (shape[1] == -1) {
+        PADDLE_THROW(platform::errors::InvalidArgument(
+            "The input [%s] shape of trt subgraph is %s, please enable "
+            "trt dynamic_shape mode by SetTRTDynamicShapeInfo.",
+            input, ShapeStr(shape)));
+      }
+      nvinfer1::Dims dims;
+      dims.nbDims = 1;
+      dims.d[0] = shape[1];
+      return dims;
     }
     return nvinfer1::Dims3(shape[1], 1, 1);
   } else {

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -1036,6 +1036,22 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
                 << desc.Output("Y").size();
         return false;
       }
+
+      auto* block = desc.Block();
+      if (block == nullptr) {
+        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
+                   "Developers need to check whether block_desc is passed in "
+                   "the pass.";
+        return false;
+      }
+      auto x_var_name = desc.Input("X")[0];
+      auto* x_var_desc = block->FindVar(x_var_name);
+      const auto x_shape = x_var_desc->GetShape();
+      if (x_shape.size() != 4) {
+        VLOG(3) << "The instance_norm op only support 4-dimensional input in "
+                   "tensorrt.";
+        return false;
+      }
     }
 
     if (op_type == "leaky_relu") {

paddle/fluid/inference/tensorrt/plugin/instance_norm_op_plugin.cu

@@ -65,11 +65,6 @@ int InstanceNormPlugin::enqueue(int batch_size, const void *const *inputs,
 #endif
                                 cudaStream_t stream) TRT_NOEXCEPT {
   const auto &input_dims = this->getInputDims(0);
-
-  PADDLE_ENFORCE_EQ(input_dims.nbDims, 3,
-                    platform::errors::InvalidArgument(
-                        "Input Dims should be 3 (except the batch), got %d",
-                        input_dims.nbDims));
   int n = batch_size;
   int c = input_dims.d[0];
   int h = input_dims.d[1];

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -555,10 +555,6 @@ if(WITH_GPU AND TENSORRT_FOUND)
     if (NOT EXISTS ${TEST_SPLIT_CONVERTER_MODEL}/split_converter.tgz)
         inference_download_and_uncompress_without_verify(${TEST_SPLIT_CONVERTER_MODEL} ${INFERENCE_URL}/tensorrt_test "split_converter.tgz")
     endif()
-    set(TEST_INSTANCE_NORM_MODEL "${TRT_MODEL_INSTALL_DIR}/trt_instance_norm_test")
-    if (NOT EXISTS ${TEST_INSTANCE_NORM_MODEL}/instance_norm.tgz)
-        inference_download_and_uncompress_without_verify(${TEST_INSTANCE_NORM_MODEL} ${INFERENCE_URL}/tensorrt_test "instance_norm.tgz")
-    endif()
     inference_analysis_test(trt_mobilenet_test SRCS trt_mobilenet_test.cc
             EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
             ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_inference_test_models)
@@ -577,9 +573,6 @@ if(WITH_GPU AND TENSORRT_FOUND)
     inference_analysis_test(trt_split_converter_test SRCS trt_split_converter_test.cc
             EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} 
             ARGS --infer_model=${TEST_SPLIT_CONVERTER_MODEL}/)
-    inference_analysis_test(trt_instance_norm_test SRCS trt_instance_norm_converter_test.cc
-            EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} 
-            ARGS --infer_model=${TEST_INSTANCE_NORM_MODEL}/)
     inference_analysis_test(test_analyzer_capi_exp_gpu SRCS analyzer_capi_exp_gpu_tester.cc
             EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} paddle_inference_c
             ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_inference_test_models)

python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_instance_norm.py

@@ -24,8 +24,6 @@ import unittest
 
 class TrtConvertInstanceNormTest(TrtLayerAutoScanTest):
     def is_program_valid(self, program_config: ProgramConfig) -> bool:
-        inputs = program_config.inputs
-        weights = program_config.weights
         attrs = [
             program_config.ops[i].attrs
             for i in range(len(program_config.ops))
@@ -38,14 +36,17 @@ class TrtConvertInstanceNormTest(TrtLayerAutoScanTest):
 
     def sample_program_configs(self):
         def generate_input1(attrs: List[Dict[str, Any]], shape_input):
-            return np.ones(shape_input).astype(np.float32)
+            return np.random.random(shape_input).astype(np.float32)
 
         def generate_input2(attrs: List[Dict[str, Any]], shape_input):
-            return np.ones(len(shape_input) - 1).astype(np.float32)
+            return np.random.random(shape_input[1]).astype(np.float32)
 
+        for batch in [1, 2, 4]:
+            for shape_input in [[batch, 16], [batch, 32, 64],
+                                [batch, 16, 32, 64]]:
+                self.in_dim = len(shape_input)
                 for epsilon in [0.0005, -1, 1]:
                     dics = [{"epsilon": epsilon}]
-
                     ops_config = [{
                         "op_type": "instance_norm",
                         "op_inputs": {
@@ -61,7 +62,6 @@ class TrtConvertInstanceNormTest(TrtLayerAutoScanTest):
                         "op_attrs": dics[0]
                     }]
                     ops = self.generate_op_config(ops_config)
-            shape_input = [1, 3, 64, 64]
                     program_config = ProgramConfig(
                         ops=ops,
                         weights={
@@ -81,9 +81,26 @@ class TrtConvertInstanceNormTest(TrtLayerAutoScanTest):
     def sample_predictor_configs(
             self, program_config) -> (paddle_infer.Config, List[int], float):
         def generate_dynamic_shape(attrs):
-            self.dynamic_shape.min_input_shape = {"input_data": [1, 3, 32, 32]}
-            self.dynamic_shape.max_input_shape = {"input_data": [4, 3, 64, 64]}
-            self.dynamic_shape.opt_input_shape = {"input_data": [1, 3, 64, 64]}
+            if self.in_dim == 2:
+                self.dynamic_shape.min_input_shape = {"input_data": [1, 4]}
+                self.dynamic_shape.max_input_shape = {"input_data": [4, 64]}
+                self.dynamic_shape.opt_input_shape = {"input_data": [2, 16]}
+            elif self.in_dim == 3:
+                self.dynamic_shape.min_input_shape = {"input_data": [1, 1, 4]}
+                self.dynamic_shape.max_input_shape = {
+                    "input_data": [4, 32, 256]
+                }
+                self.dynamic_shape.opt_input_shape = {"input_data": [2, 3, 32]}
+            elif self.in_dim == 4:
+                self.dynamic_shape.min_input_shape = {
+                    "input_data": [1, 1, 4, 4]
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_data": [4, 32, 128, 256]
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_data": [2, 3, 32, 32]
+                }
 
         def clear_dynamic_shape():
             self.dynamic_shape.min_input_shape = {}
@@ -91,8 +108,7 @@ class TrtConvertInstanceNormTest(TrtLayerAutoScanTest):
             self.dynamic_shape.opt_input_shape = {}
 
         def generate_trt_nodes_num(attrs, dynamic_shape):
-            inputs = program_config.inputs
-            if dynamic_shape:
+            if dynamic_shape or self.in_dim != 4:
                 return 0, 3
             return 1, 2
 
@@ -108,7 +124,7 @@ class TrtConvertInstanceNormTest(TrtLayerAutoScanTest):
             attrs, False), 1e-5
         self.trt_param.precision = paddle_infer.PrecisionType.Half
         yield self.create_inference_config(), generate_trt_nodes_num(
-            attrs, False), 1e-2
+            attrs, False), 1e-5
 
         # for dynamic_shape
         generate_dynamic_shape(attrs)
@@ -117,7 +133,7 @@ class TrtConvertInstanceNormTest(TrtLayerAutoScanTest):
                                                                      True), 1e-5
         self.trt_param.precision = paddle_infer.PrecisionType.Half
         yield self.create_inference_config(), generate_trt_nodes_num(attrs,
-                                                                     True), 1e-2
+                                                                     True), 1e-5
 
     def test(self):
         self.run_test()

python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_leaky_relu.py

@@ -27,18 +27,14 @@ class TrtConvertLeakyReluTest(TrtLayerAutoScanTest):
         return True
 
     def sample_program_configs(self):
-        def generate_input1(attrs: List[Dict[str, Any]]):
-            return np.ones([1, 3, 64, 64]).astype(np.float32)
+        def generate_input1(shape):
+            return np.random.random(shape).astype(np.float32)
 
+        for batch in [1, 2]:
+            for shape in [[batch, 64], [batch, 32, 64], [batch, 8, 32, 32]]:
+                self.input_dim = len(shape)
                 for alpha in [0.02, 1.0, 100.0, -1.0, 0.0]:
-            for X_scale in [1.0, 100.0, 0.01, -0.1, 0.0]:
-                dics = [{
-                    "alpha": alpha,
-                    "use_mkldnn": True,
-                    "enable_int8": True,
-                    "X_scale": X_scale
-                }]
-
+                    dics = [{"alpha": alpha}]
                     ops_config = [{
                         "op_type": "leaky_relu",
                         "op_inputs": {
@@ -54,8 +50,8 @@ class TrtConvertLeakyReluTest(TrtLayerAutoScanTest):
                         ops=ops,
                         weights={},
                         inputs={
-                        "input_data":
-                        TensorConfig(data_gen=partial(generate_input1, dics))
+                            "input_data": TensorConfig(data_gen=partial(
+                                generate_input1, shape))
                         },
                         outputs=["y_data"])
 
@@ -64,9 +60,26 @@ class TrtConvertLeakyReluTest(TrtLayerAutoScanTest):
     def sample_predictor_configs(
             self, program_config) -> (paddle_infer.Config, List[int], float):
         def generate_dynamic_shape(attrs):
-            self.dynamic_shape.min_input_shape = {"input_data": [1, 3, 32, 32]}
-            self.dynamic_shape.max_input_shape = {"input_data": [4, 3, 64, 64]}
-            self.dynamic_shape.opt_input_shape = {"input_data": [4, 3, 64, 64]}
+            if self.input_dim == 2:
+                self.dynamic_shape.min_input_shape = {"input_data": [1, 8]}
+                self.dynamic_shape.max_input_shape = {"input_data": [64, 128]}
+                self.dynamic_shape.opt_input_shape = {"input_data": [2, 16]}
+            elif self.input_dim == 3:
+                self.dynamic_shape.min_input_shape = {"input_data": [1, 8, 8]}
+                self.dynamic_shape.max_input_shape = {
+                    "input_data": [64, 128, 256]
+                }
+                self.dynamic_shape.opt_input_shape = {"input_data": [2, 16, 64]}
+            elif self.input_dim == 4:
+                self.dynamic_shape.min_input_shape = {
+                    "input_data": [1, 8, 8, 4]
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_data": [64, 64, 128, 128]
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_data": [2, 16, 64, 32]
+                }
 
         def clear_dynamic_shape():
             self.dynamic_shape.min_input_shape = {}