MTCNN算法的理解介绍及代码分析（2）

最新推荐文章于 2021-08-03 16:53:33 发布

LanceHang

最新推荐文章于 2021-08-03 16:53:33 发布

阅读量718

点赞数 3

分类专栏：深度学习

本文链接：https://blog.csdn.net/LanceHang/article/details/110952809

版权

深度学习专栏收录该内容

14 篇文章 1 订阅

订阅专栏

代码地址
这里走的是一个github上的代码mtcnn-pytorch

运行效果
如下是效果图，我们可以看到对一张图片内的每一张人脸进行检测，每个人脸有一个边界框与五个标记点。

这里我们使用cv来实现图片显示，以下是TEST.py文件，
首先从src目录下的detector文件中导入detect_faces函数，从visualization——utils中导入show_bboxes函数.
用Image打开图片，这时图片的输出类型是<class’PIL.JpegImagePlugin.JpegImageFile’>，经过detect_faces的处理得到边界框与关键点。
用cv2打开处理后的图片，这里的输出是<class ‘numpy.ndarray’> 。这里我们复制一份图片。
接着遍历边界框和关键点，这里使关键点为红色，边界框为白色

from src.detector import detect_faces
from src.visualization_utils import show_bboxes
from PIL import Image

import matplotlib.pyplot as plt

import cv2
#
image = Image.open('images/office2.jpg')
bounding_boxes, landmarks = detect_faces(image)
#
image = cv2.imread('images/office2.jpg')

draw = image.copy()
for b in bounding_boxes:
    cv2.rectangle(draw, (int(b[0]), int(b[1])), (int(b[2]), int(b[3])), (255, 255, 255))
#
for p in landmarks:
    for i in range(5):
        cv2.circle(draw, (p[i], p[i + 5]), 1, (0, 0, 255), 2)

cv2.imshow("detection result", draw)
cv2.waitKey(0)

一、上述测试涉及到了一个函数，detect_faces()，它属于detector.py文件。
（1）构建图像金字塔。

.其中min_face_size=20.0代表最小人脸， thresholds=[0.6, 0.7, 0.8]代表了三个模型的阀值，他们判断是否是人脸的置信越来越高。
nms_thresholds=[0.7, 0.7, 0.7])用于作为去重函数nums的参数（后面分析nums函数）
接着加载这三个模型，三个模型会有不同的输出，前两个包括分类输出score（人脸的概率）和回归结果bbox（N*4个坐标值）。最后一个比前面多了landmark位置的回归（5个关键点坐标）。
接着开始构建图像金字塔。获取图片的长和宽，取最小边。看到最小检测尺寸 min_detection_size = 12，缩小因子 factor = 0.707（根号0.5），通过原图像最小的边来不断使其缩小，直至满足最小检测尺寸，并通过scale保存下来其对应缩小因子。
（2）运行P-Net，对应 run_first_stage()，得到一些框，首先把为空的去除，剩下的不为空的框通过np.vstack()将其由list按照列叠加成一个新的为numpy array，便于后面处理操作。接着用nms(bounding_boxes[:, 0:5], nms_thresholds[0])进行去重操作，并把去重后的框放在边界框列表中。然后使用calibrate_box()通过offset来调整b-box坐标，用convert_to_square()将框调整为正方形，得到的正方形的中心点还是原来矩形的中心点，边长是矩阵宽高的最大值，最后进行一个四舍五入操作。
到了R-Net，先把P-Net的输出放到Variable框架内，经过rnet得到两个输出，分别是预测得到的偏移回归信息，以及置信分类信息。有了置信值，就能根据阈值来判断是不是为人脸，从而进行筛选。接着继续进行去重，调整坐标和形状，进行四舍五入。
接着使STAGE3，这里是O-Net部分，这里还是先把R-Net的输出放到Varible内，经过onet得到三个输出，除了R-Net的两个输出，这里多了一个关键点landmarks = output[0].data.numpy() # shape [n_boxes, 10]，接着还是判断是否为人脸，进行一系列操作。
（3）接着是compute landmark points阶段，即计算关键点位置，因为O-Net得到的关键点是缩放后的关键点位置，所以并不是原本图像的，这里就需要根据width和height去计算原本的关键点位置，最后在进行一次去重操作，得到最终的边界框位置以及关键点位置。

import numpy as np
import torch
from torch.autograd import Variable
from .get_nets import PNet, RNet, ONet
from .box_utils import nms, calibrate_box, get_image_boxes, convert_to_square
from .first_stage import run_first_stage


def detect_faces(image, min_face_size=20.0, # 其中20代表最小的人脸
                 thresholds=[0.6, 0.7, 0.8],#代表了三个模型的阀值，其判断是人脸的置信越来越高
                 nms_thresholds=[0.7, 0.7, 0.7]):
    """
    Arguments:
        image: an instance of PIL.Image.
        min_face_size: a float number.
        thresholds: a list of length 3.
        nms_thresholds: a list of length 3.

    Returns:
        two float numpy arrays of shapes [n_boxes, 4] and [n_boxes, 10],
        bounding boxes and facial landmarks.
    """

    # LOAD MODELS
    pnet = PNet()
    rnet = RNet()
    onet = ONet()
    onet.eval()

    # BUILD AN IMAGE PYRAMID
    width, height = image.size#获取图片的长和宽
    min_length = min(height, width)#取最小

    min_detection_size = 12#最小检测尺寸
    factor = 0.707  # sqrt(0.5) 缩小因子

    # scales for scaling the image
    scales = []

    # scales the image so that
    # minimum size that we can detect equals to
    # minimum face size that we want to detect
    m = min_detection_size/min_face_size#
    min_length *= m

    factor_count = 0
    while min_length > min_detection_size:
        scales.append(m*factor**factor_count)
        min_length *= factor
        factor_count += 1

    # STAGE 1

    # it will be returned
    bounding_boxes = []

    # run P-Net on different scales
    for s in scales:
        boxes = run_first_stage(image, pnet, scale=s, threshold=thresholds[0])
        bounding_boxes.append(boxes)

    # collect boxes (and offsets, and scores) from different scales
    bounding_boxes = [i for i in bounding_boxes if i is not None]
    bounding_boxes = np.vstack(bounding_boxes)

    keep = nms(bounding_boxes[:, 0:5], nms_thresholds[0])
    bounding_boxes = bounding_boxes[keep]

    # use offsets predicted by pnet to transform bounding boxes
    bounding_boxes = calibrate_box(bounding_boxes[:, 0:5], bounding_boxes[:, 5:])
    # shape [n_boxes, 5]

    bounding_boxes = convert_to_square(bounding_boxes)
    bounding_boxes[:, 0:4] = np.round(bounding_boxes[:, 0:4])

    # STAGE 2

    img_boxes = get_image_boxes(bounding_boxes, image, size=24)
    with torch.no_grad():
        img_boxes = Variable(torch.FloatTensor(img_boxes))
    output = rnet(img_boxes)
    offsets = output[0].data.numpy()  # shape [n_boxes, 4] 预测得到的偏移回归信息
    probs = output[1].data.numpy()  # shape [n_boxes, 2] 置信度分类信息

    keep = np.where(probs[:, 1] > thresholds[1])[0]  #置信度是否大于阈值，进一步筛选人脸框
    bounding_boxes = bounding_boxes[keep]
    bounding_boxes[:, 4] = probs[keep, 1].reshape((-1,))
    offsets = offsets[keep]

    keep = nms(bounding_boxes, nms_thresholds[1])
    bounding_boxes = bounding_boxes[keep]
    bounding_boxes = calibrate_box(bounding_boxes, offsets[keep])
    bounding_boxes = convert_to_square(bounding_boxes)
    bounding_boxes[:, 0:4] = np.round(bounding_boxes[:, 0:4])

    # STAGE 3

    img_boxes = get_image_boxes(bounding_boxes, image, size=48)
    if len(img_boxes) == 0: 
        return [], []
    with torch.no_grad():
        img_boxes = Variable(torch.FloatTensor(img_boxes))
    output = onet(img_boxes)
    landmarks = output[0].data.numpy()  # shape [n_boxes, 10]
    offsets = output[1].data.numpy()  # shape [n_boxes, 4]
    probs = output[2].data.numpy()  # shape [n_boxes, 2]

    keep = np.where(probs[:, 1] > thresholds[2])[0]
    bounding_boxes = bounding_boxes[keep]
    bounding_boxes[:, 4] = probs[keep, 1].reshape((-1,))
    offsets = offsets[keep]
    landmarks = landmarks[keep]

    # compute landmark points
    width = bounding_boxes[:, 2] - bounding_boxes[:, 0] + 1.0
    height = bounding_boxes[:, 3] - bounding_boxes[:, 1] + 1.0
    xmin, ymin = bounding_boxes[:, 0], bounding_boxes[:, 1]
    landmarks[:, 0:5] = np.expand_dims(xmin, 1) + np.expand_dims(width, 1)*landmarks[:, 0:5]
    landmarks[:, 5:10] = np.expand_dims(ymin, 1) + np.expand_dims(height, 1)*landmarks[:, 5:10]

    bounding_boxes = calibrate_box(bounding_boxes, offsets)
    keep = nms(bounding_boxes, nms_thresholds[2], mode='min')
    bounding_boxes = bounding_boxes[keep]
    landmarks = landmarks[keep]

    return bounding_boxes, landmarks

二、接下来说detector.py中STAGE1中用到的函数run_first_stage（），它属于first_stage.py文件，这个文件还有一个函数generate_bboxes()。

run_first_stage()这个可以看注释，生成边界框，并进行去重操作。其中有四个参数，一个PIL图片，一个神经网络P-Net，图像的宽高规模scale，一个置信度threshold。返回 a float numpy array of shape [n_boxes, 9],bounding boxes with scores and offsets (4 + 1 + 4).
_generate_bboxes生成可能是人脸的边界框位置，调整下offset的形式,同时计算得到每个scale对应的bbox坐标值。

import torch
from torch.autograd import Variable
import math
from PIL import Image
import numpy as np
from .box_utils import nms, _preprocess


def run_first_stage(image, net, scale, threshold):
    """Run P-Net, generate bounding boxes, and do NMS.

    Arguments:
        image: an instance of PIL.Image.
        net: an instance of pytorch's nn.Module, P-Net.
        scale: a float number,
            scale width and height of the image by this number.
        threshold: a float number,
            threshold on the probability of a face when generating
            bounding boxes from predictions of the net.

    Returns:
        a float numpy array of shape [n_boxes, 9],
            bounding boxes with scores and offsets (4 + 1 + 4).
    """

    # scale the image and convert it to a float array
    width, height = image.size
    sw, sh = math.ceil(width*scale), math.ceil(height*scale)
    img = image.resize((sw, sh), Image.BILINEAR)
    img = np.asarray(img, 'float32')

    with torch.no_grad():
        img = Variable(torch.FloatTensor(_preprocess(img)))
    output = net(img)
    probs = output[1].data.numpy()[0, 1, :, :]
    offsets = output[0].data.numpy()
    # probs: probability of a face at each sliding window
    # offsets: transformations to true bounding boxes

    boxes = _generate_bboxes(probs, offsets, scale, threshold)
    if len(boxes) == 0:
        return None

    keep = nms(boxes[:, 0:5], overlap_threshold=0.5)
    return boxes[keep]


def _generate_bboxes(probs, offsets, scale, threshold):
    """Generate bounding boxes at places
    where there is probably a face.

    Arguments:
        probs: a float numpy array of shape [n, m].
        offsets: a float numpy array of shape [1, 4, n, m].
        scale: a float number,
            width and height of the image were scaled by this number.
        threshold: a float number.

    Returns:
        a float numpy array of shape [n_boxes, 9]
    """

    # applying P-Net is equivalent, in some sense, to
    # moving 12x12 window with stride 2
    stride = 2
    cell_size = 12

    # indices of boxes where there is probably a face
    inds = np.where(probs > threshold)

    if inds[0].size == 0:
        return np.array([])

    # transformations of bounding boxes
    tx1, ty1, tx2, ty2 = [offsets[0, i, inds[0], inds[1]] for i in range(4)]
    # they are defined as:
    # w = x2 - x1 + 1
    # h = y2 - y1 + 1
    # x1_true = x1 + tx1*w
    # x2_true = x2 + tx2*w
    # y1_true = y1 + ty1*h
    # y2_true = y2 + ty2*h

    offsets = np.array([tx1, ty1, tx2, ty2])
    score = probs[inds[0], inds[1]]

    # P-Net is applied to scaled images
    # so we need to rescale bounding boxes back
    bounding_boxes = np.vstack([
        np.round((stride*inds[1] + 1.0)/scale),
        np.round((stride*inds[0] + 1.0)/scale),
        np.round((stride*inds[1] + 1.0 + cell_size)/scale),
        np.round((stride*inds[0] + 1.0 + cell_size)/scale),
        score, offsets
    ])
    # why one is added?

    return bounding_boxes.T

三、之前一直用到了一个去重函数nms，它属于box_utils.py文件，这个主要有五个函数以及一个_preprocess（）对图片进行预处理。

其中nms函数，用来非最大值抑制，也就是通过对这些边界框进行判断，选出极大值进行保留，从而达到去重的操作。
convert_to_square函数，是吧边界框转换为方形
calibrate_box，通过offset使边界框更可能是最终的目的边界框，返回调整后的bbox。
get_image_boxes，从框出的图片像素裁剪到指定大小，便于计算。
correct_bboxes，通过检查避免框的尺寸超过图像或者为负值，在之前通过offset整合时，尺寸大小容易过大或过小。它会返回十个参数，前四个中dx,dy多设为0，出现正数，此时x,y为负数了，取了下相反数，edx，edy又进行了调整。而 y, x, ey, ex则为原本的始末坐标，h，w为宽和高。
最后就是图片的预处理了
Returns:
dy, dx, edy, edx: a int numpy arrays of shape [n],
coordinates of the boxes with respect to the cutouts.
y, x, ey, ex: a int numpy arrays of shape [n],
corrected ymin, xmin, ymax, xmax.
h, w: a int numpy arrays of shape [n],
just heights and widths of boxes.
```
 in the following order:
     [dy, edy, dx, edx, y, ey, x, ex, w, h].
```

import numpy as np
from PIL import Image


def nms(boxes, overlap_threshold=0.5, mode='union'):
    """Non-maximum suppression.

    Arguments:
        boxes: a float numpy array of shape [n, 5],
            where each row is (xmin, ymin, xmax, ymax, score).
        overlap_threshold: a float number.
        mode: 'union' or 'min'.

    Returns:
        list with indices of the selected boxes
    """

    # if there are no boxes, return the empty list
    if len(boxes) == 0:
        return []

    # list of picked indices
    pick = []

    # grab the coordinates of the bounding boxes
    x1, y1, x2, y2, score = [boxes[:, i] for i in range(5)]

    area = (x2 - x1 + 1.0)*(y2 - y1 + 1.0)
    ids = np.argsort(score)  # in increasing order

    while len(ids) > 0:

        # grab index of the largest value
        last = len(ids) - 1
        i = ids[last]
        pick.append(i)

        # compute intersections
        # of the box with the largest score
        # with the rest of boxes

        # left top corner of intersection boxes
        ix1 = np.maximum(x1[i], x1[ids[:last]])
        iy1 = np.maximum(y1[i], y1[ids[:last]])

        # right bottom corner of intersection boxes
        ix2 = np.minimum(x2[i], x2[ids[:last]])
        iy2 = np.minimum(y2[i], y2[ids[:last]])

        # width and height of intersection boxes
        w = np.maximum(0.0, ix2 - ix1 + 1.0)
        h = np.maximum(0.0, iy2 - iy1 + 1.0)

        # intersections' areas
        inter = w * h
        if mode == 'min':
            overlap = inter/np.minimum(area[i], area[ids[:last]])
        elif mode == 'union':
            # intersection over union (IoU)
            overlap = inter/(area[i] + area[ids[:last]] - inter)

        # delete all boxes where overlap is too big
        ids = np.delete(
            ids,
            np.concatenate([[last], np.where(overlap > overlap_threshold)[0]])
        )

    return pick


def convert_to_square(bboxes):
    """Convert bounding boxes to a square form.

    Arguments:
        bboxes: a float numpy array of shape [n, 5].

    Returns:
        a float numpy array of shape [n, 5],
            squared bounding boxes.
    """

    square_bboxes = np.zeros_like(bboxes)
    x1, y1, x2, y2 = [bboxes[:, i] for i in range(4)]
    h = y2 - y1 + 1.0
    w = x2 - x1 + 1.0
    max_side = np.maximum(h, w)
    square_bboxes[:, 0] = x1 + w*0.5 - max_side*0.5
    square_bboxes[:, 1] = y1 + h*0.5 - max_side*0.5
    square_bboxes[:, 2] = square_bboxes[:, 0] + max_side - 1.0
    square_bboxes[:, 3] = square_bboxes[:, 1] + max_side - 1.0
    return square_bboxes


def calibrate_box(bboxes, offsets):
    """Transform bounding boxes to be more like true bounding boxes.
    'offsets' is one of the outputs of the nets.

    Arguments:
        bboxes: a float numpy array of shape [n, 5].
        offsets: a float numpy array of shape [n, 4].

    Returns:
        a float numpy array of shape [n, 5].
    """
    x1, y1, x2, y2 = [bboxes[:, i] for i in range(4)]
    w = x2 - x1 + 1.0
    h = y2 - y1 + 1.0
    w = np.expand_dims(w, 1)
    h = np.expand_dims(h, 1)

    # this is what happening here:
    # tx1, ty1, tx2, ty2 = [offsets[:, i] for i in range(4)]
    # x1_true = x1 + tx1*w
    # y1_true = y1 + ty1*h
    # x2_true = x2 + tx2*w
    # y2_true = y2 + ty2*h
    # below is just more compact form of this

    # are offsets always such that
    # x1 < x2 and y1 < y2 ?

    translation = np.hstack([w, h, w, h])*offsets
    bboxes[:, 0:4] = bboxes[:, 0:4] + translation
    return bboxes


def get_image_boxes(bounding_boxes, img, size=24):
    """Cut out boxes from the image.

    Arguments:
        bounding_boxes: a float numpy array of shape [n, 5].
        img: an instance of PIL.Image.
        size: an integer, size of cutouts.

    Returns:
        a float numpy array of shape [n, 3, size, size].
    """

    num_boxes = len(bounding_boxes)
    width, height = img.size

    [dy, edy, dx, edx, y, ey, x, ex, w, h] = correct_bboxes(bounding_boxes, width, height)
    img_boxes = np.zeros((num_boxes, 3, size, size), 'float32')

    for i in range(num_boxes):
        img_box = np.zeros((h[i], w[i], 3), 'uint8')

        img_array = np.asarray(img, 'uint8')
        img_box[dy[i]:(edy[i] + 1), dx[i]:(edx[i] + 1), :] =\
            img_array[y[i]:(ey[i] + 1), x[i]:(ex[i] + 1), :]

        # resize
        img_box = Image.fromarray(img_box)
        img_box = img_box.resize((size, size), Image.BILINEAR)
        img_box = np.asarray(img_box, 'float32')

        img_boxes[i, :, :, :] = _preprocess(img_box)

    return img_boxes


def correct_bboxes(bboxes, width, height):
    """Crop boxes that are too big and get coordinates
    with respect to cutouts.

    Arguments:
        bboxes: a float numpy array of shape [n, 5],
            where each row is (xmin, ymin, xmax, ymax, score).
        width: a float number.
        height: a float number.

    Returns:
        dy, dx, edy, edx: a int numpy arrays of shape [n],
            coordinates of the boxes with respect to the cutouts.
        y, x, ey, ex: a int numpy arrays of shape [n],
            corrected ymin, xmin, ymax, xmax.
        h, w: a int numpy arrays of shape [n],
            just heights and widths of boxes.

        in the following order:
            [dy, edy, dx, edx, y, ey, x, ex, w, h].
    """

    x1, y1, x2, y2 = [bboxes[:, i] for i in range(4)]
    w, h = x2 - x1 + 1.0,  y2 - y1 + 1.0
    num_boxes = bboxes.shape[0]

    # 'e' stands for end
    # (x, y) -> (ex, ey)
    x, y, ex, ey = x1, y1, x2, y2

    # we need to cut out a box from the image.
    # (x, y, ex, ey) are corrected coordinates of the box
    # in the image.
    # (dx, dy, edx, edy) are coordinates of the box in the cutout
    # from the image.
    dx, dy = np.zeros((num_boxes,)), np.zeros((num_boxes,))
    edx, edy = w.copy() - 1.0, h.copy() - 1.0

    # if box's bottom right corner is too far right
    ind = np.where(ex > width - 1.0)[0]
    edx[ind] = w[ind] + width - 2.0 - ex[ind]
    ex[ind] = width - 1.0

    # if box's bottom right corner is too low
    ind = np.where(ey > height - 1.0)[0]
    edy[ind] = h[ind] + height - 2.0 - ey[ind]
    ey[ind] = height - 1.0

    # if box's top left corner is too far left
    ind = np.where(x < 0.0)[0]
    dx[ind] = 0.0 - x[ind]
    x[ind] = 0.0

    # if box's top left corner is too high
    ind = np.where(y < 0.0)[0]
    dy[ind] = 0.0 - y[ind]
    y[ind] = 0.0

    return_list = [dy, edy, dx, edx, y, ey, x, ex, w, h]
    return_list = [i.astype('int32') for i in return_list]

    return return_list


def _preprocess(img):
    """Preprocessing step before feeding the network.

    Arguments:
        img: a float numpy array of shape [h, w, c].

    Returns:
        a float numpy array of shape [1, c, h, w].
    """
    img = img.transpose((2, 0, 1))
    img = np.expand_dims(img, 0)
    img = (img - 127.5)*0.0078125
    return img

四、存放网络模型的文件get_nets.py，里面有供我们使用的P,R,O三种网络模型。其原理与mtcnn论文一致。
在这里插入图片描述

import torch
import torch.nn as nn
import torch.nn.functional as F
from collections import OrderedDict
import numpy as np


class Flatten(nn.Module):

    def __init__(self):
        super(Flatten, self).__init__()

    def forward(self, x):
        """
        Arguments:
            x: a float tensor with shape [batch_size, c, h, w].
        Returns:
            a float tensor with shape [batch_size, c*h*w].
        """

        # without this pretrained model isn't working
        x = x.transpose(3, 2).contiguous()

        return x.view(x.size(0), -1)


class PNet(nn.Module):

    def __init__(self):

        super(PNet, self).__init__()

        # suppose we have input with size HxW, then
        # after first layer: H - 2,
        # after pool: ceil((H - 2)/2),
        # after second conv: ceil((H - 2)/2) - 2,
        # after last conv: ceil((H - 2)/2) - 4,
        # and the same for W

        self.features = nn.Sequential(OrderedDict([
            ('conv1', nn.Conv2d(3, 10, 3, 1)),
            ('prelu1', nn.PReLU(10)),
            ('pool1', nn.MaxPool2d(2, 2, ceil_mode=True)),

            ('conv2', nn.Conv2d(10, 16, 3, 1)),
            ('prelu2', nn.PReLU(16)),

            ('conv3', nn.Conv2d(16, 32, 3, 1)),
            ('prelu3', nn.PReLU(32))
        ]))

        self.conv4_1 = nn.Conv2d(32, 2, 1, 1)
        self.conv4_2 = nn.Conv2d(32, 4, 1, 1)

        weights = np.load('src/weights/pnet.npy')[()]
        for n, p in self.named_parameters():
            p.data = torch.FloatTensor(weights[n])

    def forward(self, x):
        """
        Arguments:
            x: a float tensor with shape [batch_size, 3, h, w].
        Returns:
            b: a float tensor with shape [batch_size, 4, h', w'].
            a: a float tensor with shape [batch_size, 2, h', w'].
        """
        x = self.features(x)
        a = self.conv4_1(x)
        b = self.conv4_2(x)
        a = F.softmax(a,dim=1)
        return b, a


class RNet(nn.Module):

    def __init__(self):

        super(RNet, self).__init__()

        self.features = nn.Sequential(OrderedDict([
            ('conv1', nn.Conv2d(3, 28, 3, 1)),
            ('prelu1', nn.PReLU(28)),
            ('pool1', nn.MaxPool2d(3, 2, ceil_mode=True)),

            ('conv2', nn.Conv2d(28, 48, 3, 1)),
            ('prelu2', nn.PReLU(48)),
            ('pool2', nn.MaxPool2d(3, 2, ceil_mode=True)),

            ('conv3', nn.Conv2d(48, 64, 2, 1)),
            ('prelu3', nn.PReLU(64)),

            ('flatten', Flatten()),
            ('conv4', nn.Linear(576, 128)),
            ('prelu4', nn.PReLU(128))
        ]))

        self.conv5_1 = nn.Linear(128, 2)
        self.conv5_2 = nn.Linear(128, 4)

        weights = np.load('src/weights/rnet.npy')[()]
        for n, p in self.named_parameters():
            p.data = torch.FloatTensor(weights[n])

    def forward(self, x):
        """
        Arguments:
            x: a float tensor with shape [batch_size, 3, h, w].
        Returns:
            b: a float tensor with shape [batch_size, 4].
            a: a float tensor with shape [batch_size, 2].
        """
        x = self.features(x)
        a = self.conv5_1(x)
        b = self.conv5_2(x)
        a = F.softmax(a,dim=1)
        return b, a


class ONet(nn.Module):

    def __init__(self):

        super(ONet, self).__init__()

        self.features = nn.Sequential(OrderedDict([
            ('conv1', nn.Conv2d(3, 32, 3, 1)),
            ('prelu1', nn.PReLU(32)),
            ('pool1', nn.MaxPool2d(3, 2, ceil_mode=True)),

            ('conv2', nn.Conv2d(32, 64, 3, 1)),
            ('prelu2', nn.PReLU(64)),
            ('pool2', nn.MaxPool2d(3, 2, ceil_mode=True)),

            ('conv3', nn.Conv2d(64, 64, 3, 1)),
            ('prelu3', nn.PReLU(64)),
            ('pool3', nn.MaxPool2d(2, 2, ceil_mode=True)),

            ('conv4', nn.Conv2d(64, 128, 2, 1)),
            ('prelu4', nn.PReLU(128)),

            ('flatten', Flatten()),
            ('conv5', nn.Linear(1152, 256)),
            ('drop5', nn.Dropout(0.25)),
            ('prelu5', nn.PReLU(256)),
        ]))

        self.conv6_1 = nn.Linear(256, 2)
        self.conv6_2 = nn.Linear(256, 4)
        self.conv6_3 = nn.Linear(256, 10)

        weights = np.load('src/weights/onet.npy')[()]
        for n, p in self.named_parameters():
            p.data = torch.FloatTensor(weights[n])

    def forward(self, x):
        """
        Arguments:
            x: a float tensor with shape [batch_size, 3, h, w].
        Returns:
            c: a float tensor with shape [batch_size, 10].
            b: a float tensor with shape [batch_size, 4].
            a: a float tensor with shape [batch_size, 2].
        """
        x = self.features(x)
        a = self.conv6_1(x)
        b = self.conv6_2(x)
        c = self.conv6_3(x)
        a = F.softmax(a,dim=1)
        return c, b, a

源于百度：
另外，这里还用到了一种激活函数PRelu，是一种带参数的Relu激活函数，其中如果ai=0，那么就是Relu激活函数。如果参数ai很小，则会退化为LRelu。有实验证明，与ReLU相比，LReLU对最终的结果几乎没什么影响。（我觉得这像是一个极限微分问题…）
在这里插入图片描述
五、这个项目中还有一个文件visualzation.py是用来显示结果的，因为这里使用了cv2，并且这个文件不涉及算法内容，不做分析。