【图像信号处理】1 视频与图像情绪分类

1. 视频分类

FER(Facial Expression Recognition)，人脸表情识别

情绪标注：包含恐惧…
情感标注：积极，消极，中性

MELD：来源于老友记，多人对话形式，是EmotionLines老友记部分的多模态扩充（文本+视频）。1433段对话，共13708句。标注了7类情绪：Neutral, Happiness, Surprise, Sadness, Anger, Disgust, Fear和3类情感：Positive, Negative, Neutral，非中性情绪占比53%。 MELD是对话情绪识别中常用的数据集之一，优点是数据集质量较高并且有多模态信息，缺点是数据集中的对话涉及到的剧情背景太多，情绪识别难度很大，Github在此

EULA：带注释的知识共享情感数据库，用于情感视频内容分析，需要申请，官网

FER-2013：表征学习的挑战，面部表情识别的挑战，官网，数据怎么得出来不太清楚

MMI Facial Expression Database：MMI面部表情数据库是一个正在进行的项目，其目的是向面部表情分析界提供大量的面部表情视觉数据。阻碍人类行为自动分析领域的新发展，特别是情感识别的一个主要问题是缺乏显示行为和情感的数据库。为了解决这个问题，MMI-面部表情数据库在2002年被设想为建立和评估面部表情识别算法的资源。该数据库解决了其他面部表情数据库中的一些关键遗漏。特别是，它包含了一个面部表情的全部时间模式的记录，从中性开始，经过一系列的起始、顶点和偏移阶段，再返回到中性脸。需要申请，官网

ExpW：丰富的人脸表征来捕捉性别、表情、头部姿势和年龄相关的属性，官网，这是图像与情绪的对应

1.1 MTCNN

• Tensorflow Github

pip install mtcnn

• Pytorch Github

关于cv2.VideoWriter_fourcc函数用法可见此

视频逐帧保存可见此文和此文

MTCNN也会出错：

解决方法：每次都取第一个人脸检测的结果，人脸识别代码：

from src import detect_faces
from PIL import Image
import cv2      # 4.5.1


image_path = 'D:/Photos/ysc.jpg'
path = '... your path/MMSA/Datasets/SIMS/Raw/video/video_0001/0001.mp4'
output_path = '... your path/output.avi'


def show_bboxes(image, bounding_boxes, facial_landmarks=[]):
    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), 10)
        # print(int(b[2]) - int(b[0]))
        # print(int(b[3]) - int(b[1]))
    for p in facial_landmarks:
        for i in range(5):
            cv2.circle(draw, (p[i], p[i + 5]), 1, (0, 0, 255), 10)
    return draw


def readtest():
    videoname = path
    capture = cv2.VideoCapture(videoname)
    # frame_width = int(capture.get(cv2.CAP_PROP_FRAME_WIDTH))  # 获取视频宽度
    # frame_height = int(capture.get(cv2.CAP_PROP_FRAME_HEIGHT))  # 获取视频高度
    fps = capture.get(cv2.CAP_PROP_FPS)  # 视频平均帧率
    n_fps = capture.get(cv2.CAP_PROP_FRAME_COUNT)
    fourcc = cv2.VideoWriter_fourcc('X','V','I','D')  # 'X','2','6','4', 'M', 'P', '4', 'V', 'X','V','I','D'
    videoWriter = cv2.VideoWriter(output_path, fourcc, fps, (150, 96))     # 150*96(188, 252)， , False灰度图
    cnt = 0
    if capture.isOpened():
        while True:
            print('\r%d / %d' % (cnt, n_fps), end='')
            ret, img = capture.read()  # img 就是一帧图片
            if not ret: break  # 当获取完最后一帧就结束
            img_PIL = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))     #  OpenCV => PIL
            bounding_boxes, landmarks = detect_faces(img_PIL)

            if len(bounding_boxes) == 0:
                print(videoname)
                # draw = show_bboxes(img, bounding_boxes, landmarks)
                # cv2.imshow('Window', draw)
                # cv2.waitKey(0)
            else:
                bounding_boxes = [[int(i) for i in bounding_boxes[0]]]
                # landmarks = [[int(i) for i in landmarks[0]]]
                # img = show_bboxes(img, bounding_boxes, landmarks)

                for b in bounding_boxes:
                    draw = img[b[1]:b[3], b[0]:b[2]]
                draw = cv2.resize(draw, (150, 96))     # 缩放至固定尺寸
			#	draw = cv2.cvtColor(draw, cv2.COLOR_BGR2GRAY)  # 灰度图
                videoWriter.write(draw)
            cnt += 1
        videoWriter.release()
    else:
        print('视频打开失败！')

readtest()

1.2 安装OpenFace 2.2.0

官网，需要引用以下：

OpenFace 2.0: Facial Behavior Analysis Toolkit
Tadas Baltrušaitis, Amir Zadeh, Yao Chong Lim, and Louis-Philippe Morency,
IEEE International Conference on Automatic Face and Gesture Recognition, 2018

运行报错：

OpenCV: terminate handler is called! The last OpenCV error is:
OpenCV(4.1.0) Error: Assertion failed (type == CV_64FC2) in cv::opt_AVX2::gemmImpl, file c:\build\master_winpack-build-win64-vc15\opencv\modules\core\src\matmul.simd.hpp, line 1066

安装opencv 4.1.0试试，官网下载地址，但试了很久并不能产生示例的output，所以放弃OpenFace，其运行过程

在ubuntu上安装一下试试，但是git很慢，而且可能还需要权限，我可能安装不了

官网应该是这个吧，安装见此，一个示例

要求如下，但是我dlib安不上，再试试服务器吧：

numpy >= 1.1, < 2.0		# 1.19.2
scipy >= 0.13, < 0.17		# 1.6.0
pandas >= 0.13, < 0.18		# 1.2.3
scikit-learn >= 0.17, < 0.18
nose >= 1.3.1, < 1.4		# 1.3.7
nolearn == 0.5b1

1.3 LSTM

视频转灰度，对于MELD数据集，有’sadness’, ‘disgust’, ‘fear’, ‘joy’, ‘anger’, ‘neutral’, 'surprise’这些情绪，认为：

emo2num = {
    'sadness':0,
    'disgust':0,
    'fear':0,
    'joy':1,
    'anger':0,
    'surprise':1
}

原来总共有9989个视频，转换后有5279个视频，其中标签为0数量：2331，标签为1数量：2948

将总共有2281个视频（Neutral：345），转换后标签为0数量：1238，标签为1的数量：698

LSTM的dim为150×96=14400，95%作为训练集（5015），5%作为测试集（264），最大迭代次数为10，可以发现Loss降得很慢，训练时间很长，结果如下：

当前代数:0,cnt:5007,平均训练集Loss:22.2459,验证集精确度:0.5379,Total time:4m 37s
当前代数:1,cnt:10014,平均训练集Loss:22.0783,验证集精确度:0.5000,Total time:9m 23s
当前代数:2,cnt:15021,平均训练集Loss:22.0122,验证集精确度:0.5909,Total time:13m 35s
当前代数:3,cnt:20028,平均训练集Loss:21.9838,验证集精确度:0.5758,Total time:17m 53s
当前代数:4,cnt:25035,平均训练集Loss:21.9460,验证集精确度:0.5114,Total time:22m 14s
当前代数:5,cnt:30042,平均训练集Loss:21.9223,验证集精确度:0.5530,Total time:26m 29s
当前代数:6,cnt:35049,平均训练集Loss:21.9133,验证集精确度:0.4242,Total time:30m 49s
当前代数:7,cnt:40056,平均训练集Loss:21.9022,验证集精确度:0.5909,Total time:35m 14s
当前代数:8,cnt:45063,平均训练集Loss:21.8852,验证集精确度:0.5909,Total time:39m 37s
当前代数:9,cnt:50070,平均训练集Loss:21.8807,验证集精确度:0.5417,Total time:44m 43s

最大迭代次数100，学习率设为0.005，90%作为训练集（4751）:

当前代数:26,平均训练集Loss:22.2553,验证集精确度:0.5844,Total time:121m 1s
Save model!
...
当前代数:99,平均训练集Loss:22.2550,验证集精确度:0.5370,Total time:447m 57s

最大迭代次数100，学习率还是设为0.001，90%作为训练集（4751）:

...
当前代数:99,平均训练集Loss:21.7791,验证集精确度:0.5285,Total time:446m 25s

精确度没有超过0.5844的，可能与参数太少有关，也可能与数据集太少且没做特征提取有关，也可能是batch=1有关…所以还是放弃做视频部分的实验吧~

1.4 卷积神经网络

torch.nn.conv3d官方文档

MaxPool3d官方文档

Loss降得太慢了…

Cuda is available!
当前代数:0,平均训练集Loss:22.0244,验证集精确度:0.4867,Total time:31m 9s
当前代数:1,平均训练集Loss:22.0167,验证集精确度:0.5209,Total time:62m 18s
当前代数:2,平均训练集Loss:22.0122,验证集精确度:0.5209,Total time:93m 22s
当前代数:3,平均训练集Loss:21.9955,验证集精确度:0.5209,Total time:124m 44s
当前代数:4,平均训练集Loss:21.9869,验证集精确度:0.5209,Total time:156m 2s
当前代数:5,平均训练集Loss:21.9884,验证集精确度:0.5209,Total time:187m 23s
当前代数:6,平均训练集Loss:21.9869,验证集精确度:0.5209,Total time:218m 30s
当前代数:7,平均训练集Loss:21.9848,验证集精确度:0.5209,Total time:250m 2s
当前代数:8,平均训练集Loss:21.9827,验证集精确度:0.5209,Total time:281m 41s
当前代数:9,平均训练集Loss:21.9814,验证集精确度:0.5209,Total time:313m 20s

emo2num = {
    'sadness':0,
    'disgust':0,
    'fear':0,
    'joy':1,
    'anger':0,
    'surprise':1
}

import csv
list = []
emo_path = '/mnt/Data1/ysc/MELD.Raw/train/train_sent_emo.csv'
with open(emo_path, 'r', encoding='utf-8') as f:
    csv_file = csv.reader(f)
    for line in csv_file:
        if line[3]=='Emotion':continue
        if line[3]=='neutral':continue
        list.append([line[5], line[6], emo2num[line[3]]])

# print(list)

# print(path_video)
# print(label)
# print(len(path_video))
# print(len(label))




import torch
import numpy as np
import torch.nn as nn
from torch.utils.data import Dataset        # Dataset的抽象类,所有其他数据集都应该进行子类化,所有子类应该override__len__和__getitem__，前者提供了数据集的大小，后者支持整数索引，范围从0到len(self)
import cv2
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
from PIL import Image
import torch.nn.functional as F

path = '...your path'

IMAGE_H = 150       # 默认输入网络的图片大小
IMAGE_W = 96

data_transform = transforms.Compose([transforms.ToTensor()])        # H*W*C to C*H*W

class VideoDataset(Dataset):
    def __init__(self, mode):
        self.mode = mode
        self.length = 0

        self.path_video = []
        self.label = []
        path = '/mnt/Data1/ysc/MELD.Raw/train/face/'

        for i in list:
            if i[0]==811 and i[1]==10: continue
            if i[0] == 813 and i[1] == 1: continue
            if i[0] == 517 and i[1] == 1: continue
            if i[0] == 608: continue
            if i[0] == 967 and i[1] == 7: continue
            if i[0] == 810 and i[1] == 0: continue
            self.path_video.append(path + 'dia{}_utt{}.avi'.format(i[0], i[1]))
            self.label.append(i[2])
            self.length += 1

    def __getitem__(self, item):
        if self.mode == 'train':


            capture = cv2.VideoCapture(self.path_video[item])
            video_tensor = torch.tensor([])
            if capture.isOpened():
                while True:
                    ret, img = capture.read()  # img 就是一帧图片
                    if not ret: break  # 当获取完最后一帧就结束
                    img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
                    img_PIL = Image.fromarray(img)
                    img_tensor = data_transform(img_PIL)
                    video_tensor = torch.cat((video_tensor, img_tensor))
            n_fps = capture.get(cv2.CAP_PROP_FRAME_COUNT)
            if n_fps!=video_tensor.size()[0]: print('Error!')
            label_tensor = torch.tensor(self.label[item])

            return video_tensor.unsqueeze(0), label_tensor

    def __len__(self):
        return self.length         # 5279


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv3d(1, 2, kernel_size=[1,3,3], padding=[0,1,1])
        self.conv2 = nn.Conv3d(2, 4, kernel_size=[1,4,3], padding=[0,1,1])
        self.conv3 = nn.Conv3d(4, 8, kernel_size=[1, 4, 3], padding=[0, 1, 1])
        self.conv4 = nn.Conv3d(8, 16, kernel_size=[1, 3, 3], padding=[0, 1, 1])
        self.conv5 = nn.Conv3d(16, 32, kernel_size=[1, 4, 3], padding=[0, 1, 1])
        # self.gru = nn.GRU(input_size=384, hidden_size=HIDDEN_SIZE, num_layers=2, bidirectional=True)
        self.gru = nn.GRU(input_size=150*96, hidden_size=HIDDEN_SIZE, num_layers=2, bidirectional=True)
        self.liner = nn.Linear(128, 2)

    # def forward(self, x):       # 1 * 32 * 150 * 96
    #
    #     x = self.conv1(x)       # 2 * 32 * 150 * 96
    #     x = F.relu(x)
    #     x = F.max_pool3d(x, [1,2,2])        # 2 * 32 * 75 * 48
    #     x = self.conv2(x)       # 4 * 32 * 74 * 48
    #     x = F.relu(x)
    #     x = F.max_pool3d(x, [1,2,2])        # 4 * 32 * 37 * 24
    #     x = self.conv3(x)       # 8 * 32 * 36 * 24
    #     x = F.relu(x)
    #     x = F.max_pool3d(x, [1,2,2])        # 8 * 32 * 18 * 12
    #     x = self.conv4(x)       # 16 * 32 * 18 * 12
    #     x = F.relu(x)
    #     x = F.max_pool3d(x, [1, 2, 2])  # 16 * 32 * 9 * 6
    #     x = self.conv5(x)   # 32 * 32 * 8 * 6
    #     x = F.relu(x)
    #     x = F.max_pool3d(x, [1, 2, 2])  # 32 * 32 * 4 * 3
    #     x = x.permute(2,0,1,3,4).contiguous()
    #     x = x.view(x.size()[0],x.size()[1],-1)      # seq_len * batch * 384
    #     x, hidden = self.gru(x)     # seq_len * batch * (64*2), 4 * batch * 64
    #     y = hidden.permute(1,0,2).contiguous()      # batch * 4 * hidden
    #     y = y.view(y.size()[0], -1)     # batch * 256
    #     y = self.liner(y)
    #
    #     return y


    def forward(self, x):       # 1 * 32 * 150 * 96
        x = x.permute(2,0,1,3,4).contiguous()
        x = x.view(x.size()[0],x.size()[1],-1)      # seq_len * batch * 384
        x, hidden = self.gru(x)     # seq_len * batch * (64*2), 4 * batch * 64
        y = hidden.permute(1,0,2).contiguous()      # batch * 4 * hidden
        y = y.view(y.size()[0], -1)     # batch * 256
        y = self.liner(y)
        return y



def epoch_time(start_time, end_time):
    elapsed_time = end_time - start_time
    elapsed_mins = int(elapsed_time / 60)
    elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
    return elapsed_mins, elapsed_secs


def evaluate(dataloader):
    model.eval()
    total_acc, total_count = 0, 0
    with torch.no_grad():
        for wave, label in dataloader:
            if len(wave.size()) != 5: continue
            wave, label = wave.cuda(), label.cuda()
            predicted_label = model(wave)

            # print(predicted_label.argmax(1))
            total_acc += (predicted_label.argmax(1) == label).sum().item()
            total_count += 1
    model.train()
    return total_acc / total_count


HIDDEN_SIZE = 32
BATCH_SIZE = 1

from torch.utils.data.dataset import random_split
import time
import matplotlib.pyplot as plt

datafile = VideoDataset('train')
dataloader = DataLoader(datafile)
n_train = int(len(datafile) * 0.9)
split_train, split_valid = random_split(dataset=datafile, lengths=[n_train, len(datafile) - n_train])
train_dataloader = DataLoader(split_train, batch_size=BATCH_SIZE, shuffle=True)
valid_dataloader = DataLoader(split_valid, batch_size=BATCH_SIZE, shuffle=True)
print(len(train_dataloader))
print(len(valid_dataloader))
model = Net()
acc_min = 0.5844
cnt = 0
if torch.cuda.is_available() == True:
    print('Cuda is available!')
    model = model.cuda()
    criterion = nn.CrossEntropyLoss()
    optimizer = torch.optim.Adam(model.parameters(), lr=0.001)  # 107186
    start_time = time.time()
    losses = 0
    loss_plot = []
    for epoch in range(100):  # 最大迭代次数
        for video, label in train_dataloader:

            if len(video.size()) != 5: continue

            video, label = video.cuda(), label.cuda()
            # print(cnt)

            out = model(video)
            loss = criterion(out, label)
            losses += loss


            if (cnt+1) % 32 == 0:
                losses.backward()
                loss_plot.append(losses)
                # print(losses)
                torch.nn.utils.clip_grad_norm_(model.parameters(), 0.25)
                optimizer.step()
                optimizer.zero_grad()
                losses = 0

            # if (cnt+1) % 500 == 0:
                # print(losses)
                # print((cnt+1)/500)
                # acc_valid = evaluate(valid_dataloader)
            cnt += 1
        end_time = time.time()
        epoch_mins, epoch_secs = epoch_time(start_time, end_time)
        acc_valid = evaluate(valid_dataloader)
        print('当前代数:{},平均训练集Loss:{:.4f},验证集精确度:{:.4f},Total time:{}m {}s'.format(epoch,
                                                                                                  sum(loss_plot) / len(loss_plot),
                                                                                               acc_valid, epoch_mins, epoch_secs))

        if acc_valid > acc_min:
            acc_min = acc_valid
            torch.save(model.state_dict(), 'model_video.pth')
            print('Save model!')

    plt.plot(loss_plot)
    plt.show()

1.5 Gabor小波变化

1. Cv2.GetGaborKernel官方文档
2. scikit-image官方文档
3. 提取最大值并进行可视化，此和此
4. 底层代码
5. 网格提取特征
6. 详细思路

2. 图像分类

还是基于图像做好了，对每一帧来判断

2.1 百度paddle

百度paddle表情识别示例，图像分类模型库

2.2

利用图片数据集，处理一下数据后大概是这样：

"0" "angry"
"1" "disgust"
"2" "fear"
"3" "happy"
"4" "sad"
"5" "surprise"
"6" "neutral"

剔除人脸识别的置信度在10%以下的结果，忽略neutral，将以上分别对应为0、1，最终的得到的n_pos,n_neg数量分别为：25775、16518

85%作为训练集：Length of train set is 35949, Length of valid set is 6344，最大迭代次数为10，batch=32，每50次迭代（1600）张图片打印一次结果：

当前代数:9,batch:0.425,平均训练集Loss:0.6041,验证集精确度:0.7070,Total time:158m 52s
Save model!
...
当前代数:9,batch:0.55,平均训练集Loss:0.6028,验证集精确度:0.7015,Total time:161m 49s

最大迭代次数改为100，增加学习率修剪，并没有很好的样子，LOSS降不下去，可能是学习率衰减的太小了，那我们就不要修剪学习率：

...
当前代数:46,batch:13.0,平均训练集Loss:0.6693,验证集精确度:0.6125,Total time:618m 53s
当前代数:46,batch:14.0,平均训练集Loss:0.6693,验证集精确度:0.6125,Total time:619m 33s

每200次迭代（6400）判断一次结果：

当前代数:14,batch:1.0,平均训练集Loss:0.5851,验证集精确度:0.7071,Total time:109m 39s
Save model!
...
当前代数:18,batch:2.0,平均训练集Loss:0.5701,验证集精确度:0.7112,Total time:130m 23s
Save model!
...
当前代数:99,batch:5.0,平均训练集Loss:0.4519,验证集精确度:0.6734,Total time:540m 50s

可见效果还是一般，剔除人脸识别的置信度在25%以下的结果，再次运行：
得到的n_pos,n_neg数量分别为：23008，14383

剔除人脸识别的置信度在40%以下的结果，19048，11229，Length of train set is 25735, Length of valid set is 4542，每100次迭代（3200）判断一次结果：

当前代数:8,batch:8.0,平均训练集Loss:0.6128,验证集精确度:0.7197,Total time:48m 4s
Save model!
...
当前代数:19,batch:7.0,平均训练集Loss:0.5486,验证集精确度:0.7472,Total time:100m 19s
Save model!
...
前代数:99,batch:8.0,平均训练集Loss:0.3421,验证集精确度:0.7078,Total time:552m 51s

剔除人脸识别的置信度在50%以下的结果，15787 8752，Length of train set is 20858, Length of valid set is 3681，每100次迭代（3200）判断一次结果：

当前代数:29,batch:6.0,平均训练集Loss:0.4824,验证集精确度:0.7691,Total time:108m 51s
Save model!
...
当前代数:99,batch:6.0,平均训练集Loss:0.2892,验证集精确度:0.7205,Total time:343m 55s

剔除人脸识别的置信度在60%以下的结果，12329 6177，Length of train set is 15730, Length of valid set is 2776，结果如下：

当前代数:37,batch:4.0,平均训练集Loss:0.5150,验证集精确度:0.7777,Total time:80m 35s
Save model!
...
当前代数:99,batch:4.0,平均训练集Loss:0.3353,验证集精确度:0.7504,Total time:207m 50s

剔除人脸识别的置信度在70%以下的结果，8647 3919，Length of train set is 10681, Length of valid set is 1885，结果如下：

...
当前代数:99,batch:3.0,平均训练集Loss:0.2360,验证集精确度:0.7353,Total time:128m 58s

batch=64，每30代（1920张）图片判断一下验证集精确度：

当前代数:99,batch:5.0,平均训练集Loss:0.2479,验证集精确度:0.7480,Total time:160m 3s

剔除人脸识别的置信度在75%以下的结果，6899 2961 Length of train set is 8381, Length of valid set is 1479 batch=32，每30代（960张）图片判断一下验证集精确度，但是训练三代都没有变化

这里进行数据增强，把NEG的直接扩展一遍：6899 5922 Length of train set is 10897, Length of valid set is 1924，结果如下：

当前代数:90,batch:8.0,平均训练集Loss:0.1941,验证集精确度:0.8784,Total time:219m 38s
Save model!
...
当前代数:99,batch:11.0,平均训练集Loss:0.1819,验证集精确度:0.8555,Total time:242m 34s

图片分类代码

import torch
import numpy as np
import torch.nn as nn
from torch.utils.data import Dataset        # Dataset的抽象类,所有其他数据集都应该进行子类化,所有子类应该override__len__和__getitem__，前者提供了数据集的大小，后者支持整数索引，范围从0到len(self)
import cv2
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
import torch.nn.functional as F
from PIL import Image
import time
from torch.utils.data.dataset import random_split
import matplotlib.pyplot as plt

path_image = '/mnt/Data1/ysc/image/origin/'
path_label = '/mnt/Data1/ysc/image/label/label.lst'

def show_bboxes(path, bounging_boxes = None):
    b = bounging_boxes
    # print(b)

    # print(path)
    image = cv2.imread(path)
    # image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    # draw = image.copy()
    # cv2.rectangle(draw, (int(b[1]), int(b[0])), (int(b[2]), int(b[3])), (255, 255, 255), 10)
    draw = image[b[0]:b[3], b[1]:b[2]]

    draw = cv2.resize(draw, (96, 150))  # 缩放至固定尺寸
    # cv2.namedWindow('image')
    # cv2.imshow('image', draw)
    # cv2.waitKey()
    return draw


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 2, 3, padding=1)
        self.conv2 = nn.Conv2d(2, 4, [4,3], padding=1)
        self.conv3 = nn.Conv2d(4, 8, [4, 3], padding=1)
        self.conv4 = nn.Conv2d(8, 16, 3, padding=1)
        self.conv5 = nn.Conv2d(16, 32, [4,3], padding=1)
        self.fc1 = nn.Linear(384, 64)
        self.fc2 = nn.Linear(64, 16)
        self.fc3 = nn.Linear(16, 2)

    def forward(self, x):       #
        x = self.conv1(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)  # batch * 2, 75, 48
        x = self.conv2(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)  # batch * 4, 37, 24
        x = self.conv3(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)  # batch * 8, 18, 12
        x = self.conv4(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)  # batch * 16, 9, 6
        x = self.conv5(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)  # batch * 32, 4, 3
        x = x.view(x.size()[0], -1)     # batch * 384
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)

        return x


emo2num = {
    '4':'0',
    '1':'0',
    '2':'0',
    '3':'1',
    '0':'0',
    '5':'1'
}

label_list={}
with open(path_label, 'r', encoding='utf-8') as f:
    for line in f.readlines():
        tmp = line.strip().split(' ')
        if eval(tmp[-2]) < 10: continue
        if tmp[-1] == '6': continue
        tmp[-1] = emo2num[tmp[-1]]
        label_list[tmp[0]] = [eval(i) for i in tmp[2:]]



data_transform = transforms.Compose([transforms.ToTensor()])


class ImageDataset(Dataset):
    def __init__(self):
        self.image = []
        self.label = []
        self.n_pos = 0
        self.n_neg = 0

        cnt = 0

        for key in label_list:
            # if cnt == 2000:break
            # else:cnt += 1

            self.image.append(key)
            # img = show_bboxes(path_image + 'shocked_expression_190.jpg', [47, 111, 206, 142, 70.0982, 1])
            if label_list[key][-1] == 1: self.n_pos += 1
            elif label_list[key][-1] == 0:self.n_neg += 1

            self.label.append(label_list[key][-1])
            # print('\r{}'.format(cnt), end='')
            # cnt += 1
        print(self.n_pos, self.n_neg)

    def __getitem__(self, item):
        key = self.image[item]
        img = show_bboxes(path_image + key, label_list[key])  # numpy.ndarray, (150, 96)
        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
        img_PIL = Image.fromarray(img)
        img_tensor = data_transform(img_PIL)
        lab = torch.LongTensor([self.label[item]])
        return img_tensor, lab

    def __len__(self):
        return self.n_pos+self.n_neg


def evaluate(dataloader):
    model.eval()
    total_acc, total_count = 0, 0
    with torch.no_grad():
        for i, (wave, label) in enumerate(dataloader):
            wave, label = wave.cuda(), label.cuda()
            predicted_label = model(wave)
            # print(predicted_label.argmax(1))
            # print(predicted_label.argmax(1))
            # print(label)
            total_acc += (predicted_label.argmax(1) == label.squeeze()).sum().item()
            # print(total_acc)
            total_count += label.size(0)
            # print(total_count)
    model.train()
    return total_acc / total_count



def epoch_time(start_time, end_time):
    elapsed_time = end_time - start_time
    elapsed_mins = int(elapsed_time / 60)
    elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
    return elapsed_mins, elapsed_secs


model = Net()
BATCH_SIZE = 32
datafile = ImageDataset()
n_train = int(len(datafile) * 0.85)
split_train, split_valid = random_split(dataset=datafile, lengths=[n_train, len(datafile) - n_train])
train_dataloader = DataLoader(split_train, batch_size=BATCH_SIZE, shuffle=True)
valid_dataloader = DataLoader(split_valid, batch_size=BATCH_SIZE, shuffle=True)
print('Length of train set is {}, Length of valid set is {}'.format(len(split_train), len(split_valid)))

if __name__ == '__main__':
    acc_min = 0.7070
    if torch.cuda.is_available() == True:
        print('Cuda is available!')
        model = model.cuda()
        optimizer = torch.optim.Adam(model.parameters(), lr=0.001)  # 学习率
        criterion = nn.CrossEntropyLoss()
        # scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 1, gamma=0.9)
        losses = []
        start_time = time.time()
        for epoch in range(100):  # 最大迭代次数
            cnt = 0
            for img, label in train_dataloader:
                img, label = img.cuda(), label.cuda()
                # print(img.size(), label.size())
                out = model(img)
                loss = criterion(out, label.squeeze())
                losses.append(loss)
                loss.backward()
                torch.nn.utils.clip_grad_norm_(model.parameters(), 0.25)
                optimizer.step()
                optimizer.zero_grad()
                if (cnt+1) % 200 == 0:
                    end_time = time.time()
                    epoch_mins, epoch_secs = epoch_time(start_time, end_time)
                    acc_valid = evaluate(valid_dataloader)
                    print('当前代数:{},batch:{},平均训练集Loss:{:.4f},验证集精确度:{:.4f},Total time:{}m {}s'.format(epoch,
                                                                                                      (cnt+1)/200,
                                                                                             sum(losses) / len(losses),
                                                                                             acc_valid, epoch_mins,
                                                                                             epoch_secs))
                    if acc_valid > acc_min:
                        acc_min = acc_valid
                        torch.save(model.state_dict(), 'model_image.pth')
                        print('Save model!')
                    # else:
                        # scheduler.step()

                cnt += 1




        plt.plot(losses)
        plt.show()

小结

最近没时间做视频的情绪分类了，所以应该就这样吧