fix trt multiclass_nms3 (#45166) (#46034)

* Support dynamic shape in multiclass_nms3 Plugin for Paddle-TensorRT.

paddle/fluid/inference/tensorrt/convert/multiclass_nms3_op.cc

@@ -54,18 +54,34 @@ class MultiClassNMS3OpConverter : public OpConverter {
         PADDLE_GET_CONST(float, op_desc.GetAttr("nms_threshold"));
     int keep_top_k = PADDLE_GET_CONST(int, op_desc.GetAttr("keep_top_k"));
     bool normalized = PADDLE_GET_CONST(bool, op_desc.GetAttr("normalized"));
-    int num_classes = scores_tensor->getDimensions().d[0];
+    int class_index = engine_->with_dynamic_shape() ? 1 : 0;
+    int num_classes = scores_tensor->getDimensions().d[class_index];
 
     auto bboxes_dims = bboxes_tensor->getDimensions();
+    nvinfer1::IShuffleLayer* bboxes_expand_layer = nullptr;
+    nvinfer1::IShuffleLayer* scores_transpose_layer = nullptr;
+    if (engine_->with_dynamic_shape()) {
+      nvinfer1::Dims4 bboxes_expand_dims(
+          bboxes_dims.d[0], bboxes_dims.d[1], 1, bboxes_dims.d[2]);
+      bboxes_expand_layer =
+          TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *bboxes_tensor);
+      bboxes_expand_layer->setReshapeDimensions(bboxes_expand_dims);
+
+      nvinfer1::Permutation permutation{0, 2, 1};
+      scores_transpose_layer =
+          TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *scores_tensor);
+      scores_transpose_layer->setFirstTranspose(permutation);
+    } else {
       nvinfer1::Dims3 bboxes_expand_dims(bboxes_dims.d[0], 1, bboxes_dims.d[1]);
-    auto* bboxes_expand_layer =
+      bboxes_expand_layer =
           TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *bboxes_tensor);
       bboxes_expand_layer->setReshapeDimensions(bboxes_expand_dims);
 
       nvinfer1::Permutation permutation{1, 0};
-    auto* scores_transpose_layer =
+      scores_transpose_layer =
           TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *scores_tensor);
       scores_transpose_layer->setFirstTranspose(permutation);
+    }
 
     std::vector<nvinfer1::ITensor*> batch_nms_inputs;
     batch_nms_inputs.push_back(bboxes_expand_layer->getOutput(0));
@@ -101,27 +117,41 @@ class MultiClassNMS3OpConverter : public OpConverter {
                    fields.size() * sizeof(nvinfer1::PluginField)));
     plugin_collections->nbFields = static_cast<int>(fields.size());
     plugin_collections->fields = fields.data();
-
-    auto creator = GetPluginRegistry()->getPluginCreator("BatchedNMS_TRT", "1");
+    std::string nms_plugin_name = "BatchedNMS_TRT";
+    if (engine_->with_dynamic_shape()) {
+      nms_plugin_name = "BatchedNMSDynamic_TRT";
+    }
+    auto creator =
+        GetPluginRegistry()->getPluginCreator(nms_plugin_name.c_str(), "1");
     auto batch_nms_plugin =
-        creator->createPlugin("BatchNMSPlugin", plugin_collections);
+        creator->createPlugin(nms_plugin_name.c_str(), plugin_collections);
     free(plugin_collections);
 
     auto batch_nms_layer = engine_->network()->addPluginV2(
         batch_nms_inputs.data(), batch_nms_inputs.size(), *batch_nms_plugin);
+    // static shape: [keep_topk, 4], [keep_topk], [keep_topk]
+    // dynamic shape: [bs, keep_topk, 4], [bs, keep_topk], [bs, keep_topk]
     auto nmsed_boxes = batch_nms_layer->getOutput(1);
     auto nmsed_scores = batch_nms_layer->getOutput(2);
     auto nmsed_classes = batch_nms_layer->getOutput(3);
 
     auto nmsed_scores_transpose_layer =
         TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *nmsed_scores);
-    nmsed_scores_transpose_layer->setReshapeDimensions(
-        nvinfer1::Dims2(keep_top_k, 1));
     auto nmsed_classes_reshape_layer =
         TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *nmsed_classes);
+    if (engine_->with_dynamic_shape()) {
+      nmsed_scores_transpose_layer->setReshapeDimensions(
+          nvinfer1::Dims3(bboxes_dims.d[0], keep_top_k, 1));
+
       nmsed_classes_reshape_layer->setReshapeDimensions(
+          nvinfer1::Dims3(bboxes_dims.d[0], keep_top_k, 1));
+    } else {
+      nmsed_scores_transpose_layer->setReshapeDimensions(
           nvinfer1::Dims2(keep_top_k, 1));
 
+      nmsed_classes_reshape_layer->setReshapeDimensions(
+          nvinfer1::Dims2(keep_top_k, 1));
+    }
     std::vector<nvinfer1::ITensor*> concat_inputs;
     concat_inputs.push_back(nmsed_classes_reshape_layer->getOutput(0));
     concat_inputs.push_back(nmsed_scores_transpose_layer->getOutput(0));
@@ -129,7 +159,8 @@ class MultiClassNMS3OpConverter : public OpConverter {
 
     auto nms_concat_layer = TRT_ENGINE_ADD_LAYER(
         engine_, Concatenation, concat_inputs.data(), concat_inputs.size());
-    nms_concat_layer->setAxis(1);
+    int axis_index = engine_->with_dynamic_shape() ? 1 : 0;
+    nms_concat_layer->setAxis(axis_index + 1);
 
     // add fake index as output to be consistent with the outputs of
     // multiclass_nms3

paddle/fluid/inference/tensorrt/convert/multiclass_nms_op.cc

@@ -52,18 +52,34 @@ class MultiClassNMSOpConverter : public OpConverter {
         PADDLE_GET_CONST(float, op_desc.GetAttr("nms_threshold"));
     int keep_top_k = PADDLE_GET_CONST(int, op_desc.GetAttr("keep_top_k"));
     bool normalized = PADDLE_GET_CONST(bool, op_desc.GetAttr("normalized"));
-    int num_classes = scores_tensor->getDimensions().d[0];
+    int class_index = engine_->with_dynamic_shape() ? 1 : 0;
+    int num_classes = scores_tensor->getDimensions().d[class_index];
 
     auto bboxes_dims = bboxes_tensor->getDimensions();
+    nvinfer1::IShuffleLayer* bboxes_expand_layer = nullptr;
+    nvinfer1::IShuffleLayer* scores_transpose_layer = nullptr;
+    if (engine_->with_dynamic_shape()) {
+      nvinfer1::Dims4 bboxes_expand_dims(
+          bboxes_dims.d[0], bboxes_dims.d[1], 1, bboxes_dims.d[2]);
+      bboxes_expand_layer =
+          TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *bboxes_tensor);
+      bboxes_expand_layer->setReshapeDimensions(bboxes_expand_dims);
+
+      nvinfer1::Permutation permutation{0, 2, 1};
+      scores_transpose_layer =
+          TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *scores_tensor);
+      scores_transpose_layer->setFirstTranspose(permutation);
+    } else {
       nvinfer1::Dims3 bboxes_expand_dims(bboxes_dims.d[0], 1, bboxes_dims.d[1]);
-    auto* bboxes_expand_layer =
+      bboxes_expand_layer =
           TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *bboxes_tensor);
       bboxes_expand_layer->setReshapeDimensions(bboxes_expand_dims);
 
       nvinfer1::Permutation permutation{1, 0};
-    auto* scores_transpose_layer =
+      scores_transpose_layer =
           TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *scores_tensor);
       scores_transpose_layer->setFirstTranspose(permutation);
+    }
 
     std::vector<nvinfer1::ITensor*> batch_nms_inputs;
     batch_nms_inputs.push_back(bboxes_expand_layer->getOutput(0));
@@ -100,9 +116,14 @@ class MultiClassNMSOpConverter : public OpConverter {
     plugin_collections->nbFields = static_cast<int>(fields.size());
     plugin_collections->fields = fields.data();
 
-    auto creator = GetPluginRegistry()->getPluginCreator("BatchedNMS_TRT", "1");
+    std::string nms_plugin_name = "BatchedNMS_TRT";
+    if (engine_->with_dynamic_shape()) {
+      nms_plugin_name = "BatchedNMSDynamic_TRT";
+    }
+    auto creator =
+        GetPluginRegistry()->getPluginCreator(nms_plugin_name.c_str(), "1");
     auto batch_nms_plugin =
-        creator->createPlugin("BatchNMSPlugin", plugin_collections);
+        creator->createPlugin(nms_plugin_name.c_str(), plugin_collections);
     free(plugin_collections);
 
     auto batch_nms_layer = engine_->network()->addPluginV2(
@@ -113,12 +134,21 @@ class MultiClassNMSOpConverter : public OpConverter {
 
     auto nmsed_scores_transpose_layer =
         TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *nmsed_scores);
-    nmsed_scores_transpose_layer->setReshapeDimensions(
-        nvinfer1::Dims2(keep_top_k, 1));
     auto nmsed_classes_reshape_layer =
         TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *nmsed_classes);
+    if (engine_->with_dynamic_shape()) {
+      nmsed_scores_transpose_layer->setReshapeDimensions(
+          nvinfer1::Dims3(bboxes_dims.d[0], keep_top_k, 1));
+
+      nmsed_classes_reshape_layer->setReshapeDimensions(
+          nvinfer1::Dims3(bboxes_dims.d[0], keep_top_k, 1));
+    } else {
+      nmsed_scores_transpose_layer->setReshapeDimensions(
+          nvinfer1::Dims2(keep_top_k, 1));
+
       nmsed_classes_reshape_layer->setReshapeDimensions(
           nvinfer1::Dims2(keep_top_k, 1));
+    }
 
     std::vector<nvinfer1::ITensor*> concat_inputs;
     concat_inputs.push_back(nmsed_classes_reshape_layer->getOutput(0));
@@ -127,7 +157,8 @@ class MultiClassNMSOpConverter : public OpConverter {
 
     auto nms_concat_layer = TRT_ENGINE_ADD_LAYER(
         engine_, Concatenation, concat_inputs.data(), concat_inputs.size());
-    nms_concat_layer->setAxis(1);
+    int axis_index = engine_->with_dynamic_shape() ? 1 : 0;
+    nms_concat_layer->setAxis(axis_index + 1);
 
     RreplenishLayerAndOutput(
         nms_concat_layer, "multiclass_nms", {output_name}, test_mode);

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -33,7 +33,10 @@ namespace tensorrt {
 struct SimpleOpTypeSetTeller : public Teller {
   SimpleOpTypeSetTeller() {
 #if IS_TRT_VERSION_GE(7130)
+    // use TensorRT plugin
     teller_set.insert("group_norm");
+    teller_set.insert("multiclass_nms3");
+    teller_set.insert("multiclass_nms");
 #endif
 #if IS_TRT_VERSION_GE(7000)
     teller_set.insert("tile");
@@ -278,7 +281,6 @@ struct SimpleOpTypeSetTeller : public Teller {
       "c_allreduce_prod",
       "roll",
       "cast",
-      "multiclass_nms3",
       "transformer_input_convert",
       "recover_padding",
       "remove_padding",
@@ -853,7 +855,6 @@ bool OpTeller::Tell(const framework::ir::Node* node,
     }
 
     if (op_type == "multiclass_nms" || op_type == "multiclass_nms3") {
-      if (with_dynamic_shape) return false;
       auto* block = desc.Block();
       if (block == nullptr) {
         VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "

paddle/fluid/inference/tests/infer_ut/test_ppyolo_mbv3.cc

@@ -73,7 +73,7 @@ TEST(tensorrt_tester_ppyolo_mbv3, multi_thread4_trt_fp32_bz2) {
                   FLAGS_modeldir + "/model.pdiparams");
   config.EnableUseGpu(100, 0);
   config.EnableTensorRtEngine(
-      1 << 20, 2, 3, paddle_infer::PrecisionType::kFloat32, false, false);
+      1 << 25, 2, 3, paddle_infer::PrecisionType::kFloat32, false, false);
   LOG(INFO) << config.Summary();
   // get groudtruth by disbale ir
   paddle_infer::services::PredictorPool pred_pool_no_ir(config_no_ir, 1);

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

+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from trt_layer_auto_scan_test import TrtLayerAutoScanTest, SkipReasons
+from program_config import TensorConfig, ProgramConfig
+import numpy as np
+import paddle.inference as paddle_infer
+from functools import partial
+from typing import Optional, List, Callable, Dict, Any, Set
+import unittest
+
+
+class TrtConvertMulticlassNMSTest(TrtLayerAutoScanTest):
+
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        return True
+
+    def create_inference_config(self, use_trt=True) -> paddle_infer.Config:
+        if use_trt:
+            config = paddle_infer.Config()
+            config.disable_glog_info()
+            config.enable_use_gpu(100, 0)
+            config.set_optim_cache_dir(self.cache_dir)
+            config.switch_ir_debug()
+            config.enable_tensorrt_engine(
+                max_batch_size=self.trt_param.max_batch_size,
+                workspace_size=self.trt_param.workspace_size,
+                min_subgraph_size=self.trt_param.min_subgraph_size,
+                precision_mode=self.trt_param.precision,
+                use_static=self.trt_param.use_static,
+                use_calib_mode=self.trt_param.use_calib_mode)
+            if len(self.dynamic_shape.min_input_shape
+                   ) != 0 and self.dynamic_shape.min_input_shape.keys(
+                   ) == self.dynamic_shape.max_input_shape.keys(
+                   ) and self.dynamic_shape.min_input_shape.keys(
+                   ) == self.dynamic_shape.opt_input_shape.keys():
+                config.set_trt_dynamic_shape_info(
+                    self.dynamic_shape.min_input_shape,
+                    self.dynamic_shape.max_input_shape,
+                    self.dynamic_shape.opt_input_shape,
+                    self.dynamic_shape.disable_trt_plugin_fp16)
+            return config
+        else:
+            config = paddle_infer.Config()
+            config.switch_ir_debug(True)
+            config.set_optim_cache_dir(self.cache_dir)
+            config.disable_glog_info()
+            return config
+
+    def sample_program_configs(self):
+
+        def generate_boxes(batch, num_boxes):
+            return np.arange(batch * num_boxes * 4,
+                             dtype=np.float32).reshape([batch, num_boxes, 4])
+
+        def generate_scores(batch, num_boxes, num_classes):
+            return np.arange(batch * num_classes * num_boxes,
+                             dtype=np.float32).reshape(
+                                 [batch, num_classes, num_boxes])
+            # return np.random.rand(batch, num_classes, num_boxes).astype(np.float32)
+
+        for batch in [1, 2]:
+            self.batch = batch
+            for nms_eta in [0.8, 1.1]:
+                for num_boxes, num_classes in [[80, 100], [40, 200], [20, 400]]:
+                    self.num_boxes, self.num_classes = num_boxes, num_classes
+                    for score_threshold in [
+                            0.01,
+                    ]:
+                        ops_config = [{
+                            "op_type": "multiclass_nms",
+                            "op_inputs": {
+                                "BBoxes": ["input_bboxes"],
+                                "Scores": ["input_scores"],
+                            },
+                            "op_outputs": {
+                                "Out": ["nms_output_boxes"],
+                            },
+                            "op_attrs": {
+                                "background_label": -1,
+                                "score_threshold": score_threshold,
+                                "nms_top_k": num_boxes,
+                                "keep_top_k": num_boxes,
+                                "nms_threshold": 0.3,
+                                "normalized": False,
+                                "nms_eta": nms_eta
+                            }
+                        }]
+                        ops = self.generate_op_config(ops_config)
+                        program_config = ProgramConfig(
+                            ops=ops,
+                            weights={},
+                            inputs={
+                                "input_bboxes":
+                                TensorConfig(data_gen=partial(
+                                    generate_boxes, batch, num_boxes)),
+                                "input_scores":
+                                TensorConfig(
+                                    data_gen=partial(generate_scores, batch,
+                                                     num_boxes, num_classes))
+                            },
+                            outputs=["nms_output_boxes"])
+                        yield program_config
+
+    def sample_predictor_configs(
+            self, program_config) -> (paddle_infer.Config, List[int], float):
+
+        def generate_dynamic_shape(attrs):
+            # The last dim of input_bboxes should be static.
+            self.dynamic_shape.min_input_shape = {
+                "input_bboxes": [1, self.num_boxes, 4],
+                "input_scores": [1, self.num_classes, self.num_boxes],
+            }
+            self.dynamic_shape.max_input_shape = {
+                "input_bboxes": [8, self.num_boxes, 4],
+                "input_scores": [8, self.num_classes, self.num_boxes],
+            }
+            self.dynamic_shape.opt_input_shape = {
+                "input_bboxes": [self.batch, self.num_boxes, 4],
+                "input_scores": [self.batch, self.num_classes, self.num_boxes],
+            }
+
+        def clear_dynamic_shape():
+            self.dynamic_shape.min_input_shape = {}
+            self.dynamic_shape.max_input_shape = {}
+            self.dynamic_shape.opt_input_shape = {}
+
+        def generate_trt_nodes_num(attrs, dynamic_shape):
+            return 1, 2
+
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+
+        # for static_shape
+        clear_dynamic_shape()
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            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
+
+        # for dynamic_shape
+        generate_dynamic_shape(attrs)
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, True), 1e-5
+        # self.trt_param.precision = paddle_infer.PrecisionType.Half
+        # yield self.create_inference_config(), generate_trt_nodes_num(
+        #     attrs, True), (1e-2, 1e-2)
+
+    def assert_tensors_near(self, atol: float, rtol: float,
+                            tensor: Dict[str, np.array],
+                            baseline: Dict[str, np.array]):
+        # the order of tensorrt outputs are not consistent with paddle
+        for key, arr in tensor.items():
+            if key == "nms_output_boxes":
+                basline_arr = np.array(
+                    sorted(baseline[key].reshape((-1, 6)),
+                           key=lambda i: [i[0], i[1]]))
+                arr = np.array(
+                    sorted(arr.reshape((-1, 6)), key=lambda i: [i[0], i[1]]))
+            else:
+                basline_arr = np.array(baseline[key].reshape((-1, 1)))
+                arr = np.array(arr.reshape((-1, 1)))
+
+            self.assertTrue(
+                basline_arr.shape == arr.shape,
+                "The output shapes are not equal, the baseline shape is " +
+                str(basline_arr.shape) + ', but got ' + str(arr.shape))
+            diff = abs(basline_arr - arr)
+            np.testing.assert_allclose(
+                basline_arr,
+                arr,
+                rtol=rtol,
+                atol=atol,
+                err_msg='Output has diff, Maximum absolute error: {}'.format(
+                    np.amax(diff)))
+
+    def assert_op_size(self, trt_engine_num, paddle_op_num):
+        # tensorrt op num is not consistent with paddle
+        return True
+
+    def test(self):
+        self.trt_param.workspace_size = 1 << 25
+        self.run_test()
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_multiclass_nms3.py

@@ -71,8 +71,10 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
             # return np.random.rand(batch, num_classes, num_boxes).astype(np.float32)
 
         for batch in [1, 2]:
-            for num_boxes in [4, 12]:
-                for num_classes in [2, 6]:
+            self.batch = batch
+            for nms_eta in [0.8, 1.1]:
+                for num_boxes, num_classes in [[80, 100], [40, 200], [20, 400]]:
+                    self.num_boxes, self.num_classes = num_boxes, num_classes
                     for score_threshold in [
                             0.01,
                     ]:
@@ -94,7 +96,7 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
                                 "keep_top_k": num_boxes,
                                 "nms_threshold": 0.3,
                                 "normalized": False,
-                                "nms_eta": 1.1
+                                "nms_eta": nms_eta
                             }
                         }]
                         ops = self.generate_op_config(ops_config)
@@ -114,12 +116,26 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
                                 "nms_output_boxes", "nms_output_num",
                                 "nms_output_index"
                             ])
-
                         yield program_config
 
     def sample_predictor_configs(
             self, program_config) -> (paddle_infer.Config, List[int], float):
 
+        def generate_dynamic_shape(attrs):
+            # The last dim of input_bboxes should be static.
+            self.dynamic_shape.min_input_shape = {
+                "input_bboxes": [1, self.num_boxes, 4],
+                "input_scores": [1, self.num_classes, self.num_boxes],
+            }
+            self.dynamic_shape.max_input_shape = {
+                "input_bboxes": [8, self.num_boxes, 4],
+                "input_scores": [8, self.num_classes, self.num_boxes],
+            }
+            self.dynamic_shape.opt_input_shape = {
+                "input_bboxes": [self.batch, self.num_boxes, 4],
+                "input_scores": [self.batch, self.num_classes, self.num_boxes],
+            }
+
         def clear_dynamic_shape():
             self.dynamic_shape.min_input_shape = {}
             self.dynamic_shape.max_input_shape = {}
@@ -141,6 +157,15 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
         yield self.create_inference_config(), generate_trt_nodes_num(
             attrs, False), 1e-2
 
+        # for dynamic_shape
+        generate_dynamic_shape(attrs)
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, True), 1e-5
+        # self.trt_param.precision = paddle_infer.PrecisionType.Half
+        # yield self.create_inference_config(), generate_trt_nodes_num(
+        #     attrs, True), (1e-2, 1e-2)
+
     def assert_tensors_near(self, atol: float, rtol: float,
                             tensor: Dict[str, np.array],
                             baseline: Dict[str, np.array]):
@@ -176,7 +201,7 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
         return True
 
     def test(self):
-        self.trt_param.workspace_size = 1 << 20
+        self.trt_param.workspace_size = 1 << 25
         self.run_test()