[Paddle Inference] Add add onehot trt converter (PaddlePaddle#48655)

* add onehot trt converter * add unitest * fix bug * opt code * fix bug * fix depth_tensor * fix unitest * fix bug * fix unitest * fix bug * fix bug * fix bug * fix bug
lxsbupt · Dec 8, 2022 · 3313074 · 3313074
1 parent b6329a0
commit 3313074
Show file tree

Hide file tree

Showing 5 changed files with 306 additions and 0 deletions.
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -2299,6 +2299,8 @@ USE_TRT_CONVERTER(conv2d_transpose);
 USE_TRT_CONVERTER(leaky_relu);
 USE_TRT_CONVERTER(shuffle_channel);
 USE_TRT_CONVERTER(where);
+USE_TRT_CONVERTER(one_hot);
+USE_TRT_CONVERTER(one_hot_v2);
 USE_TRT_CONVERTER(swish);
 USE_TRT_CONVERTER(silu);
 USE_TRT_CONVERTER(group_norm);

diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
@@ -27,6 +27,7 @@ list(
   shuffle_channel_op.cc
   fill_any_like_op.cc
   where_op.cc
+  one_hot_op.cc
   swish_op.cc
   silu_op.cc
   instance_norm_op.cc

diff --git a/paddle/fluid/inference/tensorrt/convert/one_hot_op.cc b/paddle/fluid/inference/tensorrt/convert/one_hot_op.cc
@@ -0,0 +1,92 @@
+/* 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. */
+
+#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
+
+namespace paddle {
+namespace framework {
+class Scope;
+
+namespace proto {
+class OpDesc;
+}  // namespace proto
+}  // namespace framework
+}  // namespace paddle
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+/*
+ * OneHot Op
+ */
+class OneHotOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope,
+                  bool test_mode) override {
+#if IS_TRT_VERSION_GE(8510)
+    VLOG(3) << "convert a fluid one_hot op to tensorrt one_hot layer";
+    framework::OpDesc op_desc(op, nullptr);
+
+    const auto indices_tensor = engine_->GetITensor(op_desc.Input("X").front());
+    nvinfer1::ITensor* values_tensor;
+    nvinfer1::ITensor* depth_tensor;
+    const int dtype = PADDLE_GET_CONST(int, op_desc.GetAttr("dtype"));
+    if (dtype == 2 || dtype == 3) {  // int, int64
+      const std::vector<int> values_data = {0, 1};
+      values_tensor = Add1DConstantLayer<int>(values_data, "values_tensor");
+      if (dtype == 3) {  // int64
+        VLOG(3) << "trt not support int64, so it is converted to int32.";
+      }
+    } else if (dtype == 5 || dtype == 6) {  // float
+      const std::vector<float> values_data = {0.0f, 1.0f};
+      values_tensor = Add1DConstantLayer<float>(values_data, "values_tensor");
+      if (dtype == 6) {  // int64
+        VLOG(3) << "trt not support float64, so it is converted to float32.";
+      }
+    }
+
+    auto depth_name = op_desc.Input("depth_tensor");
+    if (depth_name.size() == 0) {
+      const int depth = PADDLE_GET_CONST(int, op_desc.GetAttr("depth"));
+      depth_tensor = Add1DConstantLayer<int>(depth, "depth_tensor", true);
+    } else {
+      nvinfer1::Dims depth_dims;
+      depth_dims.nbDims = 0;
+      nvinfer1::ITensor* depth_tensor_paddle =
+          engine_->GetITensor(depth_name.front());
+      auto shuffle_layer =
+          TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *depth_tensor_paddle);
+      shuffle_layer->setReshapeDimensions(depth_dims);
+      shuffle_layer->getOutput(0)->setName(depth_tensor_paddle->getName());
+      depth_tensor = shuffle_layer->getOutput(0);
+    }
+    auto layer = TRT_ENGINE_ADD_LAYER(
+        engine_, OneHot, *indices_tensor, *values_tensor, *depth_tensor, -1);
+
+    auto output_name = op_desc.Output("Out").front();
+    RreplenishLayerAndOutput(layer, "one_hot", {output_name}, test_mode);
+#else
+    VLOG(3) << "one_hot is not supported when TensorRT < 8.5.1";
+#endif
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(one_hot, OneHotOpConverter);
+REGISTER_TRT_OP_CONVERTER(one_hot_v2, OneHotOpConverter);
diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc
@@ -1783,6 +1783,45 @@ struct SimpleOpTypeSetTeller : public Teller {
       }
     }
 
+    if (op_type == "one_hot" || op_type == "one_hot_v2") {
+#if IS_TRT_VERSION_LT(8510)
+      VLOG(3) << "one_hot/one_hot_v2 is not supported when TensorRT < 8.5.1";
+      return false;
+#endif
+      if (!with_dynamic_shape) {
+        VLOG(3)
+            << "the one_hot/one_hot_v2 op does not support static shape yet";
+        return false;
+      }
+      if (desc.HasAttr("allow_out_of_range")) {
+        VLOG(3)
+            << "allow_out_of_range one_hot/one_hot_v2 op is not supported now.";
+        if (PADDLE_GET_CONST(bool, desc.GetAttr("allow_out_of_range")))
+          return false;
+      }
+      if (desc.HasAttr("dtype")) {
+        const int dtype = PADDLE_GET_CONST(int, desc.GetAttr("dtype"));
+        if (dtype != 2 && dtype != 3 && dtype != 5) {
+          VLOG(3) << "one_hot/one_hot_v2 op only support int32, int64, float.";
+          return false;
+        }
+      }
+      auto one_hot_inputs = desc.Inputs();
+      if (one_hot_inputs.find("depth_tensor") != one_hot_inputs.end()) {
+        if (desc.Input("depth_tensor").size() != 0) {
+          return true;
+        }
+      }
+
+      if (desc.HasAttr("depth")) {
+        const int depth = PADDLE_GET_CONST(int, desc.GetAttr("depth"));
+        if (depth <= 0) {
+          VLOG(3) << "depth only support positive in one_hot/one_hot_v2 op.";
+          return false;
+        }
+      }
+    }
+
     if (op_type == "skip_layernorm") {
       if (!with_dynamic_shape) {
         VLOG(3) << "the skip_layernorm does not support static shape yet";
@@ -2447,6 +2486,8 @@ struct SimpleOpTypeSetTeller : public Teller {
       "fc",
       "shuffle_channel",
       "where",
+      "one_hot",
+      "one_hot_v2",
       "swish",
       "silu",
       "celu",
@@ -2588,6 +2629,8 @@ struct SimpleOpTypeSetTeller : public Teller {
       "fc",
       "shuffle_channel",
       "where",
+      "one_hot",
+      "one_hot_v2",
       "swish",
       "silu",
       "celu",

diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_one_hot.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_one_hot.py
@@ -0,0 +1,168 @@
+# 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.
+
+import unittest
+from functools import partial
+from typing import List
+
+import numpy as np
+from program_config import ProgramConfig, TensorConfig
+from trt_layer_auto_scan_test import TrtLayerAutoScanTest
+
+import paddle.inference as paddle_infer
+
+
+class TrtConvertOneHotTest(TrtLayerAutoScanTest):
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        ver = paddle_infer.get_trt_compile_version()
+        if ver[0] * 1000 + ver[1] * 100 + ver[2] * 10 < 8510:
+            return False
+        return True
+
+    def sample_program_configs(self):
+        self.trt_param.workspace_size = 1073741824
+
+        def generate_indices(dims, batch):
+            if dims == 2:
+                return np.random.randint(0, 10, (batch, 4), dtype=np.int32)
+            elif dims == 3:
+                return np.random.randint(0, 10, (batch, 4, 6), dtype=np.int32)
+            else:
+                return np.random.randint(
+                    0, 10, (batch, 4, 6, 8), dtype=np.int32
+                )
+
+        def generate_depth(dims, batch):
+            return np.ones((1,), dtype=np.int32) * 10
+
+        for dims in [2, 3, 4]:
+            for batch in [1, 2]:
+                self.dims = dims
+                dics = [{"dtype": 5, "depth": 10}, {}]
+                ops_config = [
+                    {
+                        "op_type": "one_hot",
+                        "op_inputs": {
+                            "X": ["input_x_data"],
+                            "depth_tensor": ["input_depth_data"],
+                        },
+                        "op_outputs": {"Out": ["output_data"]},
+                        "op_attrs": dics[0],
+                        "outputs_dtype": {"output_data": np.int},
+                    },
+                ]
+                ops = self.generate_op_config(ops_config)
+
+                program_config = ProgramConfig(
+                    ops=ops,
+                    weights={
+                        "depth_tensor": TensorConfig(
+                            data_gen=partial(generate_depth, dims, batch)
+                        ),
+                    },
+                    inputs={
+                        "indices_tensor": TensorConfig(
+                            data_gen=partial(generate_indices, dims, batch)
+                        ),
+                    },
+                    outputs=["output_data"],
+                )
+
+                yield program_config
+
+    def sample_predictor_configs(
+        self, program_config
+    ) -> (paddle_infer.Config, List[int], float):
+        def generate_dynamic_shape(attrs):
+            if self.dims == 1:
+                self.dynamic_shape.min_input_shape = {
+                    "input_x_data": [1],
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_x_data": [2],
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_x_data": [1],
+                }
+            elif self.dims == 2:
+                self.dynamic_shape.min_input_shape = {
+                    "input_x_data": [1, 4],
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_x_data": [2, 4],
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_x_data": [1, 4],
+                }
+            elif self.dims == 3:
+                self.dynamic_shape.min_input_shape = {
+                    "input_x_data": [1, 4, 6],
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_x_data": [2, 4, 6],
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_x_data": [1, 4, 6],
+                }
+            elif self.dims == 4:
+                self.dynamic_shape.min_input_shape = {
+                    "input_x_data": [1, 4, 6, 8],
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_x_data": [2, 4, 6, 8],
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_x_data": [1, 4, 6, 8],
+                }
+
+        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):
+            if not dynamic_shape:
+                return 0, 3
+            return 1, 2
+
+        attrs = [op.attrs for op in 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-5
+
+        # 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-5
+
+    def test(self):
+        self.run_test()
+
+
+if __name__ == "__main__":
+    unittest.main()