环境安装：

pycuda安装：

开源项目信息

fasterRCNN的训练

faster rcnn onnx实践

第3步测试结果：

第4步测试结果：

单张图片测试代码：

多张图片预测代码：

第5步做了修改：

第7步测试：

第8步测试

onnx转trt操作

c++ 转换trt：

Onnx转trt代码及操作：

c++ tensorrt推理部分

c++onnx转tensorrt

打印模型输入输出参数：

可以忽略的报错：

环境安装：

pip install onnxsim

pytorch安装：

cuda版本是11.0，没有cuda11.0对应的torchvision，所以安装了cpu版：
pip install torch==1.10.0+cpu torchvision==0.11.0+cpu torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html

pip install torch==1.10.0+cu113 torchvision==0.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html

tensorrt安装：

onnx转tensorrt 实战干货总结_AI视觉网奇的博客-CSDN博客_onnx转tensorrt

pycuda安装：

win10安装pycuda2022_AI视觉网奇的博客-CSDN博客

开源项目信息

根据开源项目进行部署测试：

GitHub - thb1314/tensorrt-onnx-fasterrcnn-fpn-roialign

看了代码，感觉主要是把faster rcnn拆分为两个部分，

第1部分：

AnchorGenerator

RPNHead

RegionProposalNetwork

GeneralizedRCNNTransform

第2部分：

MultiScaleRoIAlign

TwoMLPHead

FastRCNNPredictor
RoIHeads，这个又依赖前面三个。

本部分基于repo

GitHub - shouxieai/tensorRT_Pro: C++ library based on tensorrt integration

安装部分请看该项目的readme部分，在本项目文件下tensorrt_code下

如果可以给该项目点个star的话，麻烦顺手给俺也点一个吧，谢谢。

fasterRCNN的训练

demo: 使用pytorch训练自己的Faster-RCNN目标检测模型 - 野生鹅鹅 - 博客园

其他git code还请自行查找

导出onnx的时候，需要加载加载自己训练的权重。

faster rcnn onnx实践

重要的环节是第3步和 第5步，

作者提供了8个步骤：
 x01export_FasterRCNN_onnx.py
 x02test_FasterRCNN_onnx.py
 x03extract_RPN.py
 x04testRPNonnx.py
 x05extract_ROIHeader.py
 x06reduceRpnOnnx.py
 x07reduce_header_onnx.py
 x08test_header_onnx.py

第一个步骤转onnx警告和解决方法：

that if the size of dimension 1 of the input is not 1, the ONNX model will return an error_AI视觉网奇的博客-CSDN博客

第3步测试结果：

输出特征名：

['rpn_boxes', 'feature_0', 'feature_1', 'feature_2', 'feature_3', 'feature_pool']

输出特征维度：

1. list 1000*5

2. list，长度4

0： 1 256 200 264

1： 1 256 100 132

2： 1 256 50 66

3： 1 256 25 33

pool： 256 13 17

3是图片

4是维度

原本代码batch_size为1时可以正确预测，但是batch_size为2时，预测结果都是第一张图片，

稍微修改了代码：

import torch
import os
import sys
sys.path.insert(0, os.path.abspath('..'))
from model import fasterrpn_resnet50_fpn
import glob
from torchvision import transforms
import cv2

if __name__ == '__main__':
    model = fasterrpn_resnet50_fpn(pretrained=True)
    model.eval()

    img_tensor_list = list()

    transform_func = transforms.Compose([
        transforms.ToTensor(),
        # transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    input_height, input_width = (600 + 31) // 32 * 32, (800 + 31) // 32 * 32

    image_list = list()
    for item in glob.glob("./*.jpg"):
        image_list.append(cv2.resize(cv2.imread(item), dsize=(input_width, input_height)))
        img_tensor_list.append(
            transform_func(cv2.cvtColor(cv2.resize(cv2.imread(item), dsize=(input_width, input_height)), cv2.COLOR_BGR2RGB)))

    with torch.no_grad():
        results = model(img_tensor_list, is_show=True)
    result = results[0]
    for i, item in enumerate(result):
        image = image_list[i].copy()
        for score_box in item:
            box = score_box[1:]
            box = box.numpy()
            cv2.rectangle(image, tuple(map(int, box[0:2])), tuple(map(int, box[2:4])), (0, 255, 0))
        cv2.imshow("win", image)
        cv2.waitKey()
        cv2.destroyWindow("win")

    output_names = ["rpn_boxes", *tuple(['feature_'+item for item in results[1].keys()])]

    print(output_names)
    dynamic_axes = {'input':{0: "N"},'rpn_boxes': {0: "N"},'feature_0': {0: "N"}, 'feature_1': {0: "N"}, 'feature_2': {0: "N"}, 'feature_3': {0: "N"}, 'feature_pool': {0: "N"}}
    onnx_save_path = 'rpn_backbone_resnet50.onnx'
    torch.onnx.export(model, torch.rand(2, 3, input_height, input_width), onnx_save_path, verbose=False,
                      do_constant_folding=True,
                      input_names=["input"], output_names=output_names,
                      dynamic_axes=dynamic_axes,
                      opset_version=11)

    import onnxsim
    import onnx
    model = onnx.load(onnx_save_path)
    # convert model
    model_simp, check = onnxsim.simplify(model, check_n=0,input_shapes={'input':[-1,3,input_height,input_width]},
                                dynamic_input_shape=True)
                                # dynamic_input_shape=False)
    with open(onnx_save_path,'wb') as f:
    # with open(onnx_save_path.replace(".onnx","_simp.onnx"),'wb') as f:
        onnx.save(model_simp, f)

多图片可以批量预测了，但是预测结果都不对。

第4步测试结果：

单张图片测试代码：

import onnxruntime as rt
import numpy as  np
import torch
import torchvision
import cv2
from torchvision import transforms

def get_classes(filepath):
    with open(filepath, 'r', encoding='gbk') as f:
        return [item.strip() for item in f.readlines()]

if __name__ == '__main__':
    onnx_save_path = "rpn_backbone_resnet50.onnx"
    img = cv2.imread('./car.jpg')
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = cv2.resize(img, dsize=(800, 608))

    normalize = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    img_tensor = normalize(img).unsqueeze(dim=0)
    img_input = img_tensor.numpy().astype(np.float32)

    sess = rt.InferenceSession(onnx_save_path)
    input_name = sess.get_inputs()[0].name
    label_names = [sess.get_outputs()[i].name for i in range(1)]

    print("input_name",input_name)

    pred_onnx = sess.run(label_names, {input_name:img_input})

    print("label_names", label_names,"pred size",pred_onnx[0].shape)
    # output without nms
    pred_onnx = dict(zip(label_names, pred_onnx))

    image = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
    for box in pred_onnx['rpn_boxes'][0]:
        box = box[1:]
        cv2.rectangle(image, tuple(map(int,box[0:2])), tuple(map(int,box[2:4])), (0,255,0))
    cv2.imshow("img", img)
    cv2.imshow("win", image)
    cv2.waitKey()
    cv2.destroyWindow("win")

步骤4测试可视化结果，与3是不一样的， 43906，步骤3的结果是10006

第四步backbone输入输出维度：

input_name input
label_names ['rpn_boxes'] pred size (1, 4390, 6)

多张图片预测代码：

import onnxruntime as rt
import numpy as  np
import torch
import torchvision
import cv2
from torchvision import transforms

def get_classes(filepath):
    with open(filepath, 'r', encoding='gbk') as f:
        return [item.strip() for item in f.readlines()]

if __name__ == '__main__':
    onnx_save_path = "rpn_backbone_resnet50.onnx"
    img = cv2.imread('./car.jpg')
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = cv2.resize(img, dsize=(800, 608))
    print(img.shape)
    normalize = transforms.Compose([
        transforms.ToTensor(),
        # transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    img_tensor = normalize(img)#.unsqueeze(dim=0)

    img_input = img_tensor.numpy().astype(np.float32)

    img_input=np.array([img_input,img_input])

    sess = rt.InferenceSession(onnx_save_path)
    input_name = sess.get_inputs()[0].name

    for data in sess.get_outputs():

        print("outname",data.name)
    label_names = [sess.get_outputs()[i].name for i in range(1)]

    print("input_name",input_name)

    pred_onnx = sess.run(label_names, {input_name:img_input})

    print("label_names", label_names,"pred size",pred_onnx[0].shape)
    # output without nms
    pred_onnx = dict(zip(label_names, pred_onnx))

    image = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
    for box in pred_onnx['rpn_boxes'][0]:
        box = box[1:]
        cv2.rectangle(image, tuple(map(int,box[0:2])), tuple(map(int,box[2:4])), (0,255,0))
    cv2.imshow("img", img)
    cv2.imshow("win", image)
    cv2.waitKey()
    cv2.destroyWindow("win")

第5步做了修改：

改完支持多batch_size，但是多张图片预测，结果都是第一张图片的。

import torch
import os
import sys
sys.path.insert(0, os.path.abspath('..'))
from model import fasterrpn_resnet50_fpn, fasterroiheader_resnet50_fpn
import math
import glob
from torchvision import transforms
import cv2
import os

if __name__ == '__main__':
    model = fasterrpn_resnet50_fpn(pretrained=True)
    model_header = fasterroiheader_resnet50_fpn(pretrained=True, transform=model.transform, box_score_thresh=0.5,box_nms_thresh=0.3)
    model.eval()
    model_header.eval()
    img_tensor_list = list()

    transform_func = transforms.Compose([
        transforms.ToTensor(),
        # transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    input_height, input_width = (600 + 31) // 32 * 32, (800 + 31) // 32 * 32
    image_list = list()
    for item in glob.glob("./*.jpg"):
        image_list.append(cv2.resize(cv2.imread(item), dsize=(input_width, input_height)))
        img_tensor_list.append(
            transform_func(cv2.cvtColor(cv2.resize(cv2.imread(item), dsize=(input_width, input_height)), cv2.COLOR_BGR2RGB)))

    with torch.no_grad():
        proposals, features, images, original_image_sizes = model(img_tensor_list)
        if not os.path.exists('buffle.pkl'):
            with open('buffle.pkl', 'wb') as f:
                torch.save({
                    'proposals':proposals,
                    'features':features,
                    'image_sizes':images.image_sizes
                }, f)
        for feature in features.values():
            print('feature.shape',feature.shape)
        proposals = [item[:,1:] for item in proposals]
        if not os.path.exists('roi_result.pkl'):
            roi_result = model_header.roi_heads.box_roi_pool(features, proposals, images.image_sizes)

            with open('roi_result.pkl', 'wb') as f:
                torch.save({
                    'roi_result':roi_result
                }, f)

            dummy_image = torch.rand(1, 3, input_height, input_width)
            batch_size = int(dummy_image.size(0))
            dummy_proposals = [torch.rand((model.rpn.post_nms_top_n(), 4)) for _ in range(batch_size)]
            height, width = int(dummy_image.size(2)), int(dummy_image.size(3))
            dummy_features = {
                key: torch.rand(batch_size, model.backbone.out_channels, math.ceil(height / (2 ** (i + 2))),
                                math.ceil(width / (2 ** (i + 2)))) for i, key in enumerate(features.keys())}

            input_names = [*tuple(['feature_' + key for key in dummy_features.keys()]), 'proposals']

            dynamic_axes = {'proposals': {0: "N"}}
            dynamic_axes.update({'feature_'+key: {0: "B"} for key in dummy_features.keys()})
            dynamic_axes.update({name: {0: "N"} for name in ['outputs']})

            class Wrapper(torch.nn.Module):

                def __init__(self, image_sizes, model):
                    super(Wrapper, self).__init__()
                    self.image_sizes = image_sizes
                    self.model = model
                
                def forward(self, x, boxes):
                    return self.model(x, boxes, self.image_sizes)
            """
            torch.onnx.export(Wrapper(images.image_sizes, model_header.roi_heads.box_roi_pool), (features, dummy_proposals),
                              "roialign.onnx", verbose=True,
                              do_constant_folding=True,
                              input_names=input_names, output_names=["outputs"],
                              dynamic_axes=dynamic_axes,
                              opset_version=11)

            print(roi_result.shape)
            """

        result = model_header(features, proposals, images, original_image_sizes)

    for i, item in enumerate(result):
        image = image_list[i].copy()
        for score_box in item['boxes']:
            box = score_box
            box = box.numpy()
            cv2.rectangle(image, tuple(map(int, box[0:2])), tuple(map(int, box[2:4])), (0, 255, 0))
        cv2.imshow("win", image)
        cv2.waitKey()
        cv2.destroyWindow("win")
    output_names = ["boxes", "labels", "scores"]

    dummy_image = torch.rand(1, 3, input_height, input_width)
    batch_size = int(dummy_image.size(0))
    dummy_proposals = [torch.rand((model.rpn.post_nms_top_n(), 4)) for _ in range(batch_size)]
    height,width = int(dummy_image.size(2)),int(dummy_image.size(3))
    dummy_features = {key:torch.rand(batch_size, model.backbone.out_channels, math.ceil(height / (2 ** (i + 2))), math.ceil(width / (2 ** (i + 2)))) for i,key in enumerate(features.keys())}
    print(dummy_features.keys())
    input_names = [*tuple(['feature_'+key for key in dummy_features.keys()]), 'proposals']
    dynamic_axes = {'proposals': {0: "N"}}
    dynamic_axes.update({name: {0: "N"} for name in output_names})
    onnx_save_path = "header.onnx"
    torch.onnx.export(model_header, (dummy_features, dummy_proposals, dummy_image), onnx_save_path, verbose=True,
                      do_constant_folding=True,
                      input_names=input_names, output_names=output_names,
                      dynamic_axes=dynamic_axes,
                      opset_version=11)

第6步测试backbone输出维度：

import onnx_graphsurgeon as gs
import onnx

def cutOnnx():
    onnx_save_path = "rpn_backbone_resnet50.onnx"
    graph = gs.import_onnx(onnx.load(onnx_save_path))

    for output in graph.outputs:

        print(0,output)
    graph.outputs = graph.outputs[0:-1]
    for output in graph.outputs:
        print(1,output)
    graph.cleanup()
    # remove feature pool
    onnx.save(gs.export_onnx(graph), onnx_save_path)

if __name__ == '__main__':
    cutOnnx()

结果：

#0 Variable (rpn_boxes): (shape=[1, 4390, 6], dtype=float32)
# 0 Variable (feature_0): (shape=[1, 256, 152, 200], dtype=float32)
# 0 Variable (feature_1): (shape=[1, 256, 76, 100], dtype=float32)
# 0 Variable (feature_2): (shape=[1, 256, 38, 50], dtype=float32)
# 0 Variable (feature_3): (shape=[1, 256, 19, 25], dtype=float32)
# 1 Variable (rpn_boxes): (shape=[1, 4390, 6], dtype=float32)
# 1 Variable (feature_0): (shape=[1, 256, 152, 200], dtype=float32)
# 1 Variable (feature_1): (shape=[1, 256, 76, 100], dtype=float32)
# 1 Variable (feature_2): (shape=[1, 256, 38, 50], dtype=float32)

第7步测试：

x07reduce_header_onnx.py：精简header onnx，仅保留全连接层部分。具体细节：

1.将网络的输入更改为roialigned_feature和proposals，去掉roi align和fpn_level的计算部分。

tensor = tensors["218"] Line11 这段代码即对该reshape操作输入的替换。

218哪里来的？

netron软件查看的，右下角INPUTS的 data name218

这个层在整个结构的大概位置：

这里可以根据自己的生成的onnx改变名字，下面是局部放大点的图：

1. 对输出的box和score节点的前面的reshape操作进行处理

对应一下几行代码

这里需要根据自己的生成的onnx修改，切记

shape_score = gs.Constant(name="shape_score", values=np.array((-1, 90), dtype=np.int64))
    shape_boxes = gs.Constant(name="shape_boxes", values=np.array((-1, 90, 4), dtype=np.int64))
    shape_boxes_last_node = gs.Constant(name="shape_boxes_last_node", values=np.array((-1, 91, 4), dtype=np.int64))

# 这里的Reshape_320和Reshape_322是box和score的上一个reshape节点
# 这里填写上面的框选的部分分
for node in graph.nodes:
    if node.name == "Reshape_320":
        node.inputs[-1] = shape_boxes
    elif node.name == "Reshape_322":
        node.inputs[-1] = shape_score
    # the last second reshape node relative to box output
    elif node.name == "Reshape_308":
        node.inputs[-1] = shape_boxes_last_node

好像报错了：
[W] colored module is not installed, will not use colors when logging. To enable colors, please install the colored module: python3 -m pip install colored
[E] No function: shape registered for opset: 11
[W] colored module is not installed, will not use colors when logging. To enable colors, please install the colored module: python3 -m pip install colored
[E] No function: len registered for opset: 11

但是没有红色显示

清华园好像不能安装了，用豆瓣的源可以安装：

pip install colored -i https://pypi.doubanio.com/simple

第7步按照作者的转换脚本，开始的时候报错，代码：

import onnx_graphsurgeon as gs
import onnx
import numpy as np

def cutOnnx():
    onnx_save_path = "header.onnx"
    graph = gs.import_onnx(onnx.load(onnx_save_path))

    tensors = graph.tensors()
    tensor = tensors["218"]
    graph.inputs = [graph.inputs[-1], tensor.to_variable(dtype=np.float32, shape=('N', 256, 7, 7))]
    graph.inputs[-1].name = "roialigned_feature"

    graph.outputs = [graph.outputs[0], graph.outputs[-1]]

    shape_score = gs.Constant(name="shape_score", values=np.array((-1, 90), dtype=np.int64))
    shape_boxes = gs.Constant(name="shape_boxes", values=np.array((-1, 90, 4), dtype=np.int64))
    shape_boxes_last_node = gs.Constant(name="shape_boxes_last_node", values=np.array((-1, 91, 4), dtype=np.int64))

    # 这里的Reshape_320和Reshape_322是box和score的上一个reshape节点
    for node in graph.nodes:
        if node.name == "Reshape_320":
            node.inputs[-1] = shape_boxes
        elif node.name == "Reshape_322":
            node.inputs[-1] = shape_score
        # the last second reshape node relative to box output
        elif node.name == "Reshape_308":
            node.inputs[-1] = shape_boxes_last_node
    
    # 添加N,90,4 和 N,90,1的结点
    for item in graph.outputs:
        item.shape.insert(1, 90)
        # print(item.shape)
    for graph_output in graph.outputs:
        graph_output.shape[0] = 'N'
    graph.cleanup()
    new_onnx_filepath = 'new_'+onnx_save_path
    onnx.save(gs.export_onnx(graph), new_onnx_filepath)

    import onnxsim
    model = onnx.load(new_onnx_filepath)
    # convert model
    model_simp, check = onnxsim.simplify(model, check_n=0,input_shapes={'roialigned_feature':[1,256, 7, 7],'proposals':[1,4]}, 
                                dynamic_input_shape=True)

    onnx.save(model_simp, new_onnx_filepath)


if __name__ == '__main__':
    cutOnnx()

报错： 

tensor = tensors["218"]
KeyError: '218'

torch换到作者的版本1.10，这个报错没有了。

修改了一下，支持批量预测：

import onnx_graphsurgeon as gs
import onnx
import numpy as np

def cutOnnx():
    onnx_save_path = "header.onnx"
    graph = gs.import_onnx(onnx.load(onnx_save_path))

    tensors = graph.tensors()

    for key, value in tensors.items():
        print(key , value)

    # tensor = tensors["onnx::Reshape_325"]
    tensor = tensors["218"]
    graph.inputs = [graph.inputs[-1], tensor.to_variable(dtype=np.float32, shape=('batch_size', 256, 7, 7))]
    graph.inputs[-1].name = "roialigned_feature"

    graph.outputs = [graph.outputs[0], graph.outputs[-1]]

    shape_score = gs.Constant(name="shape_score", values=np.array((-1, 90), dtype=np.int64))
    shape_boxes = gs.Constant(name="shape_boxes", values=np.array((-1, 90, 4), dtype=np.int64))
    shape_boxes_last_node = gs.Constant(name="shape_boxes_last_node", values=np.array((-1, 91, 4), dtype=np.int64))

    # 这里的Reshape_320和Reshape_322是box和score的上一个reshape节点
    for node in graph.nodes:
        if node.name == "Reshape_320":
            node.inputs[-1] = shape_boxes
        elif node.name == "Reshape_322":
            node.inputs[-1] = shape_score
        # the last second reshape node relative to box output
        elif node.name == "Reshape_308":
            node.inputs[-1] = shape_boxes_last_node
    
    # 添加N,90,4 和 N,90,1的结点
    for item in graph.outputs:
        item.shape.insert(1, 90)
        # print(item.shape)
    for graph_output in graph.outputs:
        graph_output.shape[0] = 'N'
    graph.cleanup()
    new_onnx_filepath = 'new_'+onnx_save_path
    onnx.save(gs.export_onnx(graph), new_onnx_filepath)

    import onnxsim
    model = onnx.load(new_onnx_filepath)
    # convert model
    model_simp, check = onnxsim.simplify(model, check_n=0,input_shapes={'roialigned_feature':[-1,256, 7, 7],'proposals':[-1,4]},
                                dynamic_input_shape=True)

    onnx.save(model_simp, new_onnx_filepath)


if __name__ == '__main__':
    cutOnnx()

第8步测试

import onnxruntime as rt
import numpy as  np

if __name__ == '__main__':

    sess = rt.InferenceSession('new_header.onnx')
    input_names = [item.name for item in sess.get_inputs()]
    output_names = [item.name for item in sess.get_outputs()]

    # proposal = np.array([1,1,10,10], dtype=np.float32).reshape(-1, 4)
    batch_size = 1
    input_dict = dict(
        proposals = np.random.randn(batch_size, 4).astype(dtype=np.float32),
        roialigned_feature = np.random.randn(batch_size, 256, 7, 7).astype(dtype=np.float32)
    )
    pred_onnx = sess.run(output_names, input_dict)
    pred_onnx = dict(zip(output_names, pred_onnx))
    print(pred_onnx['boxes'].shape)
    # print(pred_onnx['boxes'])
    print(pred_onnx['scores'].shape)
    # print(pred_onnx['scores'])

结果bath_size为1时正确，大于1就报错。

输入两个参数：proposals 和roialigned_feature

维度为 batch_size,4

和batch_size,256, 7, 7

输出：

(1, 90, 4)
(1, 90)

onnx转trt操作

c++ 转换trt：

builder\trt_builder.cpp

报错代码：

 void set_layer_hook_reshape(const LayerHookFuncReshape& func){
        //register_layerhook_reshape(func);
    }

register_layerhook_reshape函数是在NvOnnxParser.cpp中，

OnnxParser代码是在生成的时候用，需要protobuf，版本未知，

protobuf-cpp-3.11.4

pytorch转onnx

开源项目给了转换代码：

test/x01export_FasterRCNN_onnx.py

onnx测试代码：

test/x02test_FasterRCNN_onnx.py

Onnx转trt代码及操作：

onnx转tensorrt 实战干货总结_AI视觉网奇的博客-CSDN博客_onnx转tensorrt

博客的目录：onnx转tensorrt 分类成功

转trt报错了

[08/06/2022-11:35:29] [TRT] [E] [graphShapeAnalyzer.cpp::nvinfer1::builder::`anonymous-namespace'::ShapeNodeRemover::analyzeShapes::1285] Error Code 4: Miscellaneous (IShuffleLayer Reshape_1226: reshape changes volume. Reshaping [720588174] to [1,4507].)
Completed parsing of ONNX file
Building an engine from file F:\project\jushi\tensorrt-onnx-fasterrcnn-fpn-roialign-master\test\fasterrcnn_backbone_resnet50_fpn_roialign.onnx; this may take a while...
[08/06/2022-11:35:29] [TRT] [E] 4: [network.cpp::nvinfer1::Network::validate::2633] Error Code 4: Internal Error (Network must have at least one output)

解决方法，手动设置最后一层：

      last_layer = network.get_layer(network.num_layers - 1)
        network.mark_output(last_layer.get_output(0))

代码：

def ONNX_build_engine(onnx_file_path, write_engine=True):
    # 通过加载onnx文件，构建engine
    # :param onnx_file_path: onnx文件路径
    # :return: engine

    G_LOGGER = trt.Logger(trt.Logger.WARNING)
    # 1、动态输入第一点必须要写的
    explicit_batch = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
    batch_size = 8  # trt推理时最大支持的batchsize
    with trt.Builder(G_LOGGER) as builder, builder.create_network(explicit_batch) as network, trt.OnnxParser(network, G_LOGGER) as parser:
        builder.max_batch_size = batch_size
        config = builder.create_builder_config()
        config.max_workspace_size = GiB(2)
        config.set_flag(trt.BuilderFlag.FP16)
        print('Loading ONNX file from path {}...'.format(onnx_file_path))
        with open(onnx_file_path, 'rb') as model:
            print('Beginning ONNX file parsing')
            parser.parse(model.read())
        print('Completed parsing of ONNX file')
        print('Building an engine from file {}; this may take a while...'.format(onnx_file_path))
        # 重点
        profile = builder.create_optimization_profile()  # 动态输入时候需要 分别为最小输入、常规输入、最大输入
        # 有几个输入就要写几个profile.set_shape 名字和转onnx的时候要对应
        # tensorrt6以后的版本是支持动态输入的，需要给每个动态输入绑定一个profile，用于指定最小值，常规值和最大值，如果超出这个范围会报异常。
        profile.set_shape("inputs", (1, 3, 600, 600), (8, 3, 600, 600), (16, 3, 600, 600))
        config.add_optimization_profile(profile)
        last_layer = network.get_layer(network.num_layers - 1)
        network.mark_output(last_layer.get_output(0))
        engine = builder.build_engine(network, config)
        print("Completed creating Engine")
        # 保存engine文件
        if write_engine:
            engine_file_path = 'efficientnet_b1.trt'
            with open(engine_file_path, "wb") as f:
                f.write(engine.serialize())
        return engine

但是维度不能对齐：

ShapeNodeRemover::analyzeShapes::1285] Error Code 4: Miscellaneous (IShuffleLayer Reshape_1226: reshape changes volume. Reshaping [720588174] to [1,4507].)

torch升级版本后，转trt报错变成了：
Error Code 9: Internal Error (Floor_45: IUnaryLayer cannot be used to compute a shape tensor)

backbone部分可以转trt：
rpn_backbone_resnet50.onnx

c++ tensorrt推理部分

c++onnx转tensorrt

也可以python转，

onnx生成tensorrt的时候，用的自带的代码，自带的，需要protobuf，

“google/protobuf/port_def.inc”:

官方的faster rcnn导出trt报错：

reshape changes volume. Reshaping [720588174] to [1,4507].)

fasterrcnn.cpp：

缩放尺寸，归一化:

virtual bool preprocess(Job& job, const Mat& image) override{

打印模型输入输出参数：

void InferImpl::print(){

xxxxx

}

宽800，高608，宽高比和car图片相反。

推理报错：

sub_model推理报错：
[][error][trt_builder.cpp:30]:NVInfer: 3: [executionContext.cpp::nvinfer1::rt::ExecutionContext::setBindingDimensions::944] Error Code 3: API Usage Error (Parameter check failed at: executionContext.cpp::nvinfer1::rt::ExecutionContext::setBindingDimensions::944, condition: profileMaxDims.d[i] >= dimensions.d[i]. Supplied binding dimension [787,256,7,7] for bindings[1] exceed min ~ max range at index 0, maximum dimension in profile is 1, minimum dimension in profile is 1, but supplied dimension is 787.
)

787是anchors的数量：new_header需要的为1。

int number_anchors = roi_align_inputs_index / 6; 

感觉就是backbone和new_header的维度没对上。

new_header.onnx

tensorrt推理：

app_fasterrcnn\fasterrcnn.cpp

//forward
engine->forward(false);
engine->synchronize();

infer\trt_infer.cpp

bool execute_result = context->context_->enqueueV2

N和bach_size是一样的。

可以忽略的报错：

WARNING: The shape inference of prim::Constant type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
WARNING: The shape inference of prim::Constant type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.

TracerWarning: Iterating over a tensor might cause the trace to be incorrect. Passing a tensor of different shape won't change the number of iterations executed (and might lead to errors or silently give incorrect results).