[Paddle-TRT] support output_padding in conv2d_transpose and conv3d_transpose (#45004)

paddle/fluid/inference/tensorrt/convert/conv2d_op.cc

@@ -77,6 +77,12 @@ void ConvertConv2d(TensorRTEngine* engine,
       PADDLE_GET_CONST(std::vector<int>, op_desc.GetAttr("strides"));
   std::vector<int> paddings =
       PADDLE_GET_CONST(std::vector<int>, op_desc.GetAttr("paddings"));
+  // for conv2d_transpose
+  std::vector<int> output_padding;
+  if (op_desc.HasAttr("output_padding")) {
+    output_padding =
+        PADDLE_GET_CONST(std::vector<int>, op_desc.GetAttr("output_padding"));
+  }
   std::string padding_algorithm = "EXPLICIT";
   if (op_desc.HasAttr("padding_algorithm"))
     padding_algorithm =
@@ -90,15 +96,14 @@ void ConvertConv2d(TensorRTEngine* engine,
   nvinfer1::DimsHW nv_ksize(filter_h, filter_w);
   nvinfer1::DimsHW nv_dilations(dilations[0], dilations[1]);
   nvinfer1::DimsHW nv_strides(strides[0], strides[1]);
-  nvinfer1::DimsHW nv_paddings;
-  nvinfer1::Dims nv_pre_paddings;
-  nvinfer1::Dims nv_post_paddings;
+  nvinfer1::DimsHW nv_pre_paddings;
+  nvinfer1::DimsHW nv_post_paddings;
   if (paddings.size() == 2) {
-    nv_paddings.d[0] = paddings[0];
-    nv_paddings.d[1] = paddings[1];
+    nv_pre_paddings.d[0] = paddings[0];
+    nv_pre_paddings.d[1] = paddings[1];
+    nv_post_paddings.d[0] = paddings[0];
+    nv_post_paddings.d[1] = paddings[1];
   } else {
-    nv_pre_paddings.nbDims = 2;
-    nv_post_paddings.nbDims = 2;
     nv_pre_paddings.d[0] = paddings[0];
     nv_pre_paddings.d[1] = paddings[2];
     nv_post_paddings.d[0] = paddings[1];
@@ -138,12 +143,16 @@ void ConvertConv2d(TensorRTEngine* engine,
       platform::errors::Fatal("TensorRT create conv2d/conv2d_transpose"
                               " layer failed."));
   layer->setStride(nv_strides);
-  if (paddings.size() == 2) {
-    layer->setPadding(nv_paddings);
-  } else {
   layer->setPrePadding(nv_pre_paddings);
-    layer->setPostPadding(nv_post_paddings);
+  if (output_padding.size() > 0) {
+    nv_post_paddings.d[0] -= output_padding[0];
+    nv_post_paddings.d[1] -= output_padding[1];
+  }
+  if (nv_post_paddings.d[0] < 0 || nv_post_paddings.d[1] < 0) {
+    PADDLE_THROW(platform::errors::Fatal(
+        "The value in conv2d_transpose's PostPadding should be >= 0."));
   }
+  layer->setPostPadding(nv_post_paddings);
 
   layer->setNbGroups(groups);
   if (padding_algorithm == "SAME") {

paddle/fluid/inference/tensorrt/convert/conv3d_op.cc

@@ -78,10 +78,17 @@ void ConvertConv3d(TensorRTEngine* engine,
     padding_algorithm =
         PADDLE_GET_CONST(std::string, op_desc.GetAttr("padding_algorithm"));
 
+  // for conv3d_transpose
+  std::vector<int> output_padding;
+  if (op_desc.HasAttr("output_padding")) {
+    output_padding =
+        PADDLE_GET_CONST(std::vector<int>, op_desc.GetAttr("output_padding"));
+  }
+
   nvinfer1::Dims3 nv_ksize(filter_d, filter_h, filter_w);
   nvinfer1::Dims3 nv_dilations(dilations[0], dilations[1], dilations[2]);
   nvinfer1::Dims3 nv_strides(strides[0], strides[1], strides[2]);
-  nvinfer1::Dims3 nv_paddings(paddings[0], paddings[1], paddings[2]);
+  nvinfer1::Dims3 nv_pre_paddings(paddings[0], paddings[1], paddings[2]);
 
   auto weight = engine->GetTrtWeight(op_desc.Input("Filter").front(), *Y_t);
   float* bias_data = nullptr;
@@ -99,7 +106,25 @@ void ConvertConv3d(TensorRTEngine* engine,
       platform::errors::Fatal("TensorRT create conv3d/conv3d_transpose"
                               " layer failed."));
   layer->setStrideNd(nv_strides);
-  layer->setPaddingNd(nv_paddings);
+  layer->setPrePadding(nv_pre_paddings);
+  nvinfer1::Dims3 nv_post_paddings = nv_pre_paddings;
+  if (output_padding.size() > 0) {
+// Here is consistent with op_teller.cc
+#if IS_TRT_VERSION_GE(8400)
+    nv_post_paddings.d[0] -= output_padding[0];
+    nv_post_paddings.d[1] -= output_padding[1];
+    nv_post_paddings.d[2] -= output_padding[2];
+
+    if (nv_post_paddings.d[0] < 0 || nv_post_paddings.d[1] < 0 ||
+        nv_post_paddings.d[2] < 0) {
+      PADDLE_THROW(platform::errors::Fatal(
+          "The value in conv3d_transpose's PostPadding should be >= 0."));
+    }
+
+#endif
+  }
+  layer->setPostPadding(nv_post_paddings);
+
   layer->setNbGroups(groups);
   if (padding_algorithm == "SAME") {
     layer->setPaddingMode(nvinfer1::PaddingMode::kSAME_UPPER);

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -2061,10 +2061,11 @@ bool OpTeller::Tell(const framework::ir::Node* node,
     }
 #endif
 
-    // conv2d_transpose, conv3d_transpose, depthwise_conv2d_transpose
-    if (op_type.find("d_transpose") > 0) {
-      // trt doen't support output_padding,
-      // output_padding is set when stride > 1
+    // conv3d_transpose
+    if (op_type == "conv3d_transpose") {
+      // trt doen't support output_padding when < 8406
+      // output_padding is usually set when stride > 1
+#if !IS_TRT_VERSION_GE(8400)
       if (desc.HasAttr("output_padding")) {
         const std::vector<int> output_padding =
             PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("output_padding"));
@@ -2074,6 +2075,7 @@ bool OpTeller::Tell(const framework::ir::Node* node,
           if (max_padding > 0) return false;
         }
       }
+#endif
     }
 
     if (op_type == "conv3d" || op_type == "conv3d_transpose") {

python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_conv2d_transpose.py

@@ -301,7 +301,7 @@ class TrtConvertConv2dTransposeTest2(TrtLayerAutoScanTest):
             self.dynamic_shape.opt_input_shape = {}
 
         def generate_trt_nodes_num(attrs, dynamic_shape):
-            return 0, 3
+            return 1, 2
 
         attrs = [
             program_config.ops[i].attrs for i in range(len(program_config.ops))
@@ -311,19 +311,19 @@ class TrtConvertConv2dTransposeTest2(TrtLayerAutoScanTest):
         clear_dynamic_shape()
         self.trt_param.precision = paddle_infer.PrecisionType.Float32
         yield self.create_inference_config(), generate_trt_nodes_num(
-            attrs, False), 1e-5
+            attrs, False), 1e-4
         self.trt_param.precision = paddle_infer.PrecisionType.Half
         yield self.create_inference_config(), generate_trt_nodes_num(
-            attrs, False), (1e-5, 1e-3)
+            attrs, False), (1e0, 1e-3)
 
         # 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
+            attrs, True), 1e-4
         self.trt_param.precision = paddle_infer.PrecisionType.Half
         yield self.create_inference_config(), generate_trt_nodes_num(
-            attrs, True), (1e-5, 1e-3)
+            attrs, True), (1e0, 1e-3)
 
     def add_skip_trt_case(self):
         pass

python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_conv3d_transpose.py

+# Copyright (c) 2021 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 unittest
+import paddle.inference as paddle_infer
+from functools import partial
+from typing import Optional, List, Callable, Dict, Any, Set
+
+
+# Special case
+class TrtConvertConv3dTransposeTest(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 < 8400:
+            return False
+        return True
+
+    def sample_program_configs(self):
+        self.trt_param.workspace_size = 1073741824
+
+        def generate_input1(batch, num_channels, attrs: List[Dict[str, Any]]):
+            return np.ones([batch, num_channels, 4, 20, 30]).astype(np.float32)
+
+        def generate_weight1(num_channels, attrs: List[Dict[str, Any]]):
+            return np.random.random([num_channels, 64, 3, 3,
+                                     3]).astype(np.float32)
+
+        num_channels = 128
+        batch = 1
+        # in_channels
+        self.num_channels = num_channels
+        dics = [{
+            "data_fromat": 'NCHW',
+            "dilations": [1, 1, 1],
+            "padding_algorithm": 'EXPLICIT',
+            "groups": 1,
+            "paddings": [1, 1, 1],
+            "strides": [2, 2, 2],
+            "output_padding": [1, 1, 1],
+            "output_size": [],
+        }]
+
+        ops_config = [{
+            "op_type": "conv3d_transpose",
+            "op_inputs": {
+                "Input": ["input_data"],
+                "Filter": ["conv3d_weight"]
+            },
+            "op_outputs": {
+                "Output": ["output_data"]
+            },
+            "op_attrs": dics[0]
+        }]
+        ops = self.generate_op_config(ops_config)
+
+        program_config = ProgramConfig(
+            ops=ops,
+            weights={
+                "conv3d_weight":
+                TensorConfig(
+                    data_gen=partial(generate_weight1, num_channels, dics))
+            },
+            inputs={
+                "input_data":
+                TensorConfig(data_gen=partial(generate_input1, batch,
+                                              num_channels, dics))
+            },
+            outputs=["output_data"])
+
+        yield program_config
+
+    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, 128, 4, 20, 30],
+            }
+            self.dynamic_shape.max_input_shape = {
+                "input_data": [1, 128, 4, 20, 30],
+            }
+            self.dynamic_shape.opt_input_shape = {
+                "input_data": [1, 128, 4, 20, 30],
+            }
+
+        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-3
+
+        # 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-3
+
+    def add_skip_trt_case(self):
+        pass
+
+    def test(self):
+        self.add_skip_trt_case()
+        self.run_test()
+
+    def test_quant(self):
+        self.add_skip_trt_case()
+        self.run_test(quant=True)
+
+
+if __name__ == "__main__":
+    unittest.main()