基于efficientDet在jetson NX上实现目标检测

简介

EfficientNet从三个维度均扩大了，但是扩大多少，就是通过作者提出来的复合模型扩张方法结合神经结构搜索技术获得的。        

复合模型扩张方法

实现最优化目标。 优化目标就是在资源有限的情况下，要最大化 Accuracy, 优化目标的公式表达如下：  作者发现，更大的网络具有更大的宽度、深度或分辨率，往往可以获得更高的精度，但精度增益在达到80%后会迅速饱和，这表明了只对单一维度进行扩张的局限性，实验结果如下图：        

作者指出，模型扩张的各个维度之间并不是完全独立的，比如说，对于更大的分辨率图像，应该使用更深、更宽的网络，这就意味着需要平衡各个扩张维度，而不是在单一维度张扩张。

所以本文提出了复合扩张方法，这也是文章核心的地方。         

先固定公式中的一个参数，然后求别参数，从而得到基础的d0网络，然后固定其他参数，根据基础网络扩展d1以上的网络

网络结构

efficientDet

文章动机

１、如何高效的进行多尺度特征融合(efficient multi-scale feature fusion)：提到多尺度融合，在融合不同的输入特征时，以往的研究（FPN以及一些对FPN的改进工作）大多只是没有区别的将特征相加；然而，由于这些不同的输入特征具有不同的分辨率，我们观察到它们对融合输出特征的贡献往往是不平等的，为了解决这一问题，作者提出了一种简单而高效的加权（类似与attention）双向特征金字塔网络（BiFPN），它引入可学习的权值来学习不同输入特征的重要性，同时反复应用自顶向下和自下而上的多尺度特征融合。

2、如何对模型进行扩张（参考上文 EfficientNet ，同时考虑depth、width和resolution）

作者基于EfficientNet, 提出对检测器的backbone等网络进行模型缩放，并且结合提出的BiFPN提出了新的检测器家族，叫做EfficientDet。本文提出的检测器的主要遵循one-stage的设计思想，通过优化网络结构，可以达到更高的效率和精度。

BiFPN结构

如下图所示，BiFPN在图e的基础上增加了shortcut，这些都是在现有的一些工作的基础上添砖加瓦。        

伪softmax

作者提出了快速的限制方法         

efficientDet静态图导出

之前的写好的是动态图模式的复现，我们首先要将其转换成静态图模式

%cd EfficientDet/

/home/aistudio/EfficientDet

安装两个库

银河易创

一站式AIGC创作平台，集成GPT-3.5、GPT-4、文心一言等对话模型、Midjourney、DallE等绘画工具、AI音乐、AI视频和AI PPT等功能！

下载

!pip install pycocotools webcolors

# 先检查动态图能不能跑通!python b.py

接下来导出静态图

注意，自定义算子的apply无法在静态图里面使用，因此如果要导出静态图，必须要先设置好自定义算子在静态静态图的运动模式

#自定义算子class SwishImplementation(PyLayer):    @staticmethod
    def forward(ctx, i):
        result = i * F.sigmoid(i)
        ctx.save_for_backward(i)        return result    @staticmethod
    def backward(ctx, grad_output):
        i = ctx.saved_tensor()[0]
        sigmoid_i = F.sigmoid(i)        return grad_output * (sigmoid_i * (1 + i * (1 - sigmoid_i)))

# 动态图模式class MemoryEfficientSwish(nn.Layer):
    def forward(self, x):
        return SwishImplementation.apply(x)

# 静态图的自定义算子的运用class Swish(nn.Layer):
    def forward(self, x):
        return x * F.sigmoid(x)

# 导出静态图import paddleimport cv2import numpy as npfrom backbone import EfficientDetBackbonefrom efficientdet.utils import BBoxTransform, ClipBoxesfrom utils.utils import invert_affine, postprocess, preprocess_videoimport os

compound_coef = 0# 更改compound可以选择efficientDet的版本force_input_size = None  # set None to use default sizethreshold = 0.2iou_threshold = 0.2obj_list = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',            'fire hydrant', '', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep',            'cow', 'elephant', 'bear', 'zebra', 'giraffe', '', 'backpack', 'umbrella', '', '', 'handbag', 'tie',            'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove',            'skateboard', 'surfboard', 'tennis racket', 'bottle', '', 'wine glass', 'cup', 'fork', 'knife', 'spoon',            'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut',            'cake', 'chair', 'couch', 'potted plant', 'bed', '', 'dining table', '', '', 'toilet', '', 'tv',            'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',            'refrigerator', '', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier',            'toothbrush']# tf bilinear interpolation is different from any other's, just make doinput_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[compound_coef] if force_input_size is None else force_input_size# load modelmodel = EfficientDetBackbone(compound_coef=compound_coef, num_classes=len(obj_list)
,onnx_export=True)
model.set_state_dict(paddle.load(f'weights/efficientdet-d{compound_coef}.pdparams'))
model.eval()
input_spec = paddle.static.InputSpec([1, 3, 512, 512], 'float32', 'image')# input_spec = paddle.static.InputSpec([None, 3, 512, 512], 'float32', 'image')# 填写底下那句，到后面None会被表示成-1，这样如果动态图里面有-1就不行了model = paddle.jit.to_static(model, input_spec=[input_spec])
paddle.jit.save(model, 'inference_models/efficientDet')# paddle.onnx.export(model, "inference", input_spec=[input_spec],opset_version=11)# 打印保存的模型文件名print(os.listdir('inference_models'))

注意在设置模型输入类型时，如果动态图模式里，维度大小表示时使用了-1.则输入类型可能不能有-1

模型可视化

点击项目右边的数据可视化，然后选择模型文件为inference_models/efficientDet.pdmodel，然后进入可视化即可

这个模型有点大！

测试静态图

# Core Author: Zylo117# Script's Author: winter2897 """
Simple Inference Script of EfficientDet-Pytorch for detecting objects on webcam
"""import timeimport paddleimport cv2import numpy as npfrom efficientdet.utils import BBoxTransform, ClipBoxesfrom utils.utils import invert_affine, postprocess, preprocess_video# Video's pathvideo_src = 'video.mp4'  # set int to use webcam, set str to read from a video filecompound_coef = 0force_input_size = None  # set None to use default sizethreshold = 0.2iou_threshold = 0.2obj_list = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',            'fire hydrant', '', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep',            'cow', 'elephant', 'bear', 'zebra', 'giraffe', '', 'backpack', 'umbrella', '', '', 'handbag', 'tie',            'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove',            'skateboard', 'surfboard', 'tennis racket', 'bottle', '', 'wine glass', 'cup', 'fork', 'knife', 'spoon',            'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut',            'cake', 'chair', 'couch', 'potted plant', 'bed', '', 'dining table', '', '', 'toilet', '', 'tv',            'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',            'refrigerator', '', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier',            'toothbrush']# tf bilinear interpolation is different from any other's, just make doinput_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[compound_coef] if force_input_size is None else force_input_sizedef display(preds, imgs):
    for i in range(len(imgs)):        if len(preds[i]['rois']) == 0:            return imgs[i]        for j in range(len(preds[i]['rois'])):            # print("x1 y1 x2 y2",preds[i]['rois'][j])
            (x1, y1, x2, y2) = preds[i]['rois'][j].astype(np.int)
            
            cv2.rectangle(imgs[i], (x1, y1), (x2, y2), (255, 255, 0), 2)
            obj = obj_list[preds[i]['class_ids'][j]]
            score = float(preds[i]['scores'][j])
            cv2.putText(imgs[i], '{}, {:.3f}'.format(obj, score),
                        (x1, y1 + 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5,
                        (255, 255, 0), 1)        
        return imgs[i]# BoxregressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()# Video capturecap = cv2.VideoCapture(video_src)
fourcc = cv2.VideoWriter_fourcc(*'MP4V')
fps = cap.get(cv2.CAP_PROP_FPS)
width, height = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)), int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
outvideo = cv2.VideoWriter('result.mp4', fourcc, fps, (width, height))

model = paddle.jit.load('inference_models/efficientDet')# predictor=predict_config("inference_models/u2netp.pdmodel","inference_models/u2netp.pdiparams")while True:
    ret, frame = cap.read()    if not ret:        break
    # print("frame shape is",frame.shape)
    # frame preprocessing
    ori_imgs, framed_imgs, framed_metas = preprocess_video(frame, max_size=input_size)

    x = paddle.stack([paddle.to_tensor(fi) for fi in framed_imgs], 0)

    x = x.astype(paddle.float32).transpose([0, 3, 1, 2])    # model predict
    with paddle.no_grad():        
        # y=x.numpy().astype(np.float32)
        # l=predict(predictor,[y])
        time_start = time.time()
        l=model(x)        print('Time Cost：{}'.format(time.time()-time_start) , "s")
        regression=paddle.to_tensor(l[-3]).astype(paddle.float32)
        classification=paddle.to_tensor(l[-2]).astype(paddle.float32)
        anchors=paddle.to_tensor(l[-1]).astype(paddle.float32)       # print(l)
        out = postprocess(x,
                        anchors, regression, classification,
                        regressBoxes, clipBoxes,
                        threshold, iou_threshold)    # result
    out = invert_affine(framed_metas, out)
    img_show = display(out, ori_imgs)    # show frame by frame
    # cv2.imshow('frame',img_show)
    outvideo.write(img_show)    if cv2.waitKey(1) & 0xFF == ord('q'): 
        breakcap.release()
cv2.destroyAllWindows()

cpu版的aistudio推理较慢，GPU版大约在25帧左右

部署到jetson NX

jetson NX上配置paddleGPU的环境

具体参考我之前写的 paddle在jetson NX的配置

efficientDet的部署

把efficientdet，utils和inference_models三个文件夹下载下来（里面包含了模型和目标检测所用的所有工具） 然后在板子上用以下代码即可运行efficientDet。 或者使用work文件下，我已经打包好了的

import paddleimport cv2import numpy as npfrom efficientdet.utils import BBoxTransform, ClipBoxesfrom utils.utils import invert_affine, postprocess, preprocess_video# Video's pathvideo_src = 'video.mp4'  # set int to use webcam, set str to read from a video filecompound_coef = 0force_input_size = None  # set None to use default sizethreshold = 0.2iou_threshold = 0.2obj_list = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',            'fire hydrant', '', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep',            'cow', 'elephant', 'bear', 'zebra', 'giraffe', '', 'backpack', 'umbrella', '', '', 'handbag', 'tie',            'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove',            'skateboard', 'surfboard', 'tennis racket', 'bottle', '', 'wine glass', 'cup', 'fork', 'knife', 'spoon',            'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut',            'cake', 'chair', 'couch', 'potted plant', 'bed', '', 'dining table', '', '', 'toilet', '', 'tv',            'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',            'refrigerator', '', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier',            'toothbrush']# tf bilinear interpolation is different from any other's, just make doinput_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[compound_coef] if force_input_size is None else force_input_size# function for displaydef gstreamer_pipeline(
    capture_width=1280,
    capture_height=720,
    display_width=1280,
    display_height=720,
    framerate=60,
    flip_method=0,):
    return (        "nvarguscamerasrc ! "
        "video/x-raw(memory:NVMM), "
        "width=(int)%d, height=(int)%d, "
        "format=(string)NV12, framerate=(fraction)%d/1 ! "
        "nvvidconv flip-method=%d ! "
        "video/x-raw, width=(int)%d, height=(int)%d, format=(string)BGRx ! "
        "videoconvert ! "
        "video/x-raw, format=(string)BGR ! appsink"
        % (
            capture_width,
            capture_height,
            framerate,
            flip_method,
            display_width,
            display_height,
        )
    ) 

    



def display(preds, imgs):
    for i in range(len(imgs)):        if len(preds[i]['rois']) == 0:            return imgs[i]        for j in range(len(preds[i]['rois'])):            # print("x1 y1 x2 y2",preds[i]['rois'][j])
            (x1, y1, x2, y2) = preds[i]['rois'][j].astype(np.int)
            
            cv2.rectangle(imgs[i], (x1, y1), (x2, y2), (255, 255, 0), 2)
            obj = obj_list[preds[i]['class_ids'][j]]
            score = float(preds[i]['scores'][j])
            cv2.putText(imgs[i], '{}, {:.3f}'.format(obj, score),
                        (x1, y1 + 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5,
                        (255, 255, 0), 1)        
        return imgs[i]# BoxregressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()# Video capturecap = cv2.VideoCapture(gstreamer_pipeline(flip_method=0), cv2.CAP_GSTREAMER)# cap = cv2.VideoCapture(video_src)fps = cap.get(cv2.CAP_PROP_FPS)
width, height = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)), int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

model = paddle.jit.load('inference_models/efficientDet')while True:
    ret, frame = cap.read()    if not ret:        break
    # print("frame shape is",frame.shape)
    # frame preprocessing
    ori_imgs, framed_imgs, framed_metas = preprocess_video(frame, max_size=input_size)

    x = paddle.stack([paddle.to_tensor(fi) for fi in framed_imgs], 0)

    x = x.astype(paddle.float32).transpose([0, 3, 1, 2])    # model predict
    with paddle.no_grad():        
        # y=x.numpy().astype(np.float32)
        # l=predict(predictor,[y])
        l=model(x)
        regression=paddle.to_tensor(l[-3]).astype(paddle.float32)
        classification=paddle.to_tensor(l[-2]).astype(paddle.float32)
        anchors=paddle.to_tensor(l[-1]).astype(paddle.float32)        # print(l)
        out = postprocess(x,
                        anchors, regression, classification,
                        regressBoxes, clipBoxes,
                        threshold, iou_threshold)    # result
    out = invert_affine(framed_metas, out)
    img_show = display(out, ori_imgs)    # show frame by frame
    cv2.imshow('frame',img_show)    
    if cv2.waitKey(1) & 0xFF == ord('q'): 
        breakcap.release()
cv2.destroyAllWindows()

期望FPS为4（很慢） YOLOv4在NX上的FPS也差不多