本 Task 是『基于人脸的常见表情识别』训练营的第 3 课,如果你未学习前面的课程,请从 Task1 开始学习,本 Task 需要使用到数据集,在 Task2 中提供下载的渠道。
在完成了数据准备后,接下来就是卷积神经网络一展身手的时候了,通过本次任务,您将学到如何使用 PyTorch 深度学习框架,实现卷积神经网络的定义、训练和预测。
模型搭建与训练
得到了数据之后,接下来咱们使用 PyTorch 这个框架来进行模型的训练。整个训练流程包括数据接口准备、模型定义、结果保存与分析。
1. 数据接口准备
PyTorch 图像分类直接利用文件夹作为输入,只需要把不同类的数据放到不同的文件夹中。数据读取的完整代码如下:
data_transforms = {
'train': transforms.Compose([
transforms.RandomSizedCrop(48),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5])
]),
'val': transforms.Compose([
transforms.Scale(64),
transforms.CenterCrop(48),
transforms.ToTensor(),
transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5])
]),
}
data_dir = './train_val_data/'
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),
data_transforms[x]) for x in ['train', 'val']}
dataloders = {x: torch.utils.data.DataLoader(image_datasets[x],
batch_size=16,
shuffle=True,
num_workers=4) for x in ['train', 'val']}
上面脚本中的函数,输入一个文件夹,输出图片路径以及标签,在开始训练之前需要将数据集进行拆分,拆分成训练集(train)和验证集(val),训练集和测试集的比例为9:1
,train_val_data
文件结构如下所示,其中 0 代表 none、 1 代表pouting、2 代表 smile、3 代表 openmouth:
- train_val_data
- train
- 0
- 1
- 2
- 3
- val
- 0
- 1
- 2
- 3
到此,数据接口就定义完毕了,接下来在训练代码中看如何使用迭代器进行数据读取就可以了。
2. 模型定义
创建数据接⼝后,我们开始定义⼀个⽹络 simpleconv3
import torch.nn as nn
import torch.nn.functional as F
class simpleconv3(nn.Module):
def __init__(self):
super(simpleconv3,self).__init__()
self.conv1 = nn.Conv2d(3, 12, 3, 2)
self.bn1 = nn.BatchNorm2d(12)
self.conv2 = nn.Conv2d(12, 24, 3, 2)
self.bn2 = nn.BatchNorm2d(24)
self.conv3 = nn.Conv2d(24, 48, 3, 2)
self.bn3 = nn.BatchNorm2d(48)
self.fc1 = nn.Linear(48 * 5 * 5 , 1200)
self.fc2 = nn.Linear(1200 , 128)
self.fc3 = nn.Linear(128 , 4)
def forward(self , x):
x = F.relu(self.bn1(self.conv1(x)))
#print "bn1 shape",x.shape
x = F.relu(self.bn2(self.conv2(x)))
x = F.relu(self.bn3(self.conv3(x)))
x = x.view(-1 , 48 * 5 * 5)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
上面就是我们定义的网络,是一个简单的 3 层卷积。在 torch.nn 下,有各种网络层,这里就用到了 nn.Conv2d,nn.BatchNorm2d 和 nn.Linear,分别是卷积层,BN 层和全连接层。我们以一个卷积层为例:
conv1 = nn.Conv2d(in_channels=3, out_channels=12, kernel_size=3, stride=2)
bn1 = nn.BatchNorm2d(num_features=12)
- in_channels:输入通道数
- out_channels:输出通道数
- kernel_size:卷积核的大小
- stride:卷积核的移动步长
更全面的参数,请自查 API:https://pytorch.org/
3. 模型训练
这里准备好了数据集,可以直接使用
深度学习一般使用 GPU 进行训练,大家可以根据下图的操作,将设备切换至 GPU状态,再运行下面的程序,切换过程需要一定时间,请耐心等待
#coding:utf8
from __future__ import print_function, division
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.autograd import Variable
import torchvision
from torchvision import datasets, models, transforms
import time
import os
from tensorboardX import SummaryWriter
import torch.nn.functional as F
import numpy as np
import warnings
warnings.filterwarnings('ignore')
writer = SummaryWriter()
def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
for epoch in range(num_epochs):
print('Epoch {}/{}'.format(epoch, num_epochs - 1))
for phase in ['train', 'val']:
if phase == 'train':
scheduler.step()
model.train(True) # Set model to training mode
else:
model.train(False) # Set model to evaluate mode
running_loss = 0.0
running_corrects = 0.0
for data in dataloders[phase]:
inputs, labels = data
if use_gpu:
inputs = Variable(inputs.cuda())
labels = Variable(labels.cuda())
else:
inputs, labels = Variable(inputs), Variable(labels)
optimizer.zero_grad()
outputs = model(inputs)
_, preds = torch.max(outputs.data, 1)
loss = criterion(outputs, labels)
if phase == 'train':
loss.backward()
optimizer.step()
running_loss += loss.data.item()
running_corrects += torch.sum(preds == labels).item()
epoch_loss = running_loss / dataset_sizes[phase]
epoch_acc = running_corrects / dataset_sizes[phase]
if phase == 'train':
writer.add_scalar('data/trainloss', epoch_loss, epoch)
writer.add_scalar('data/trainacc', epoch_acc, epoch)
else:
writer.add_scalar('data/valloss', epoch_loss, epoch)
writer.add_scalar('data/valacc', epoch_acc, epoch)
print('{} Loss: {:.4f} Acc: {:.4f}'.format(
phase, epoch_loss, epoch_acc))
writer.export_scalars_to_json("./all_scalars.json")
writer.close()
return model
if __name__ == '__main__':
data_transforms = {
'train': transforms.Compose([
transforms.RandomSizedCrop(48),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5])
]),
'val': transforms.Compose([
transforms.Scale(64),
transforms.CenterCrop(48),
transforms.ToTensor(),
transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5])
]),
}
data_dir = './Emotion_Recognition_File/train_val_data/' # 数据集所在的位置
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),
data_transforms[x]) for x in ['train', 'val']}
dataloders = {x: torch.utils.data.DataLoader(image_datasets[x],
batch_size=64,
shuffle=True if x=="train" else False,
num_workers=8) for x in ['train', 'val']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
use_gpu = torch.cuda.is_available()
print("是否使用 GPU", use_gpu)
modelclc = simpleconv3()
print(modelclc)
if use_gpu:
modelclc = modelclc.cuda()
criterion = nn.CrossEntropyLoss()
optimizer_ft = optim.SGD(modelclc.parameters(), lr=0.1, momentum=0.9)
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=100, gamma=0.1)
modelclc = train_model(model=modelclc,
criterion=criterion,
optimizer=optimizer_ft,
scheduler=exp_lr_scheduler,
num_epochs=10) # 这里可以调节训练的轮次
if not os.path.exists("models"):
os.mkdir('models')
torch.save(modelclc.state_dict(),'models/model.ckpt')
训练的过程需要注意几个参数,第一个是数据加载器(dataloders)中的 batch_size
,这个代表的含义是每次送入模型训练的图片数量,这个需要根据GPU的显存来设置,显存越大,可以设置越大,这个数一般设置为 2 的整数次幂(如 4、8、16、32 等)
dataloders = {x: torch.utils.data.DataLoader(image_datasets[x],
batch_size=64,
shuffle=True if x=="train" else False,
num_workers=8) for x in ['train', 'val']}
第二个需要注意的参数是训练函数的 num_epochs
,这个参数代表的意义是,模型训练的轮次。
modelclc = train_model(model=modelclc,
criterion=criterion,
optimizer=optimizer_ft,
scheduler=exp_lr_scheduler,
num_epochs=10) # 这里可以调节训练的轮次
模型测试
上⾯已经训练好了模型,我们接下来的⽬标,就是要⽤它来做推理,真正把模型⽤起来,下⾯我们载⼊⼀个图⽚,⽤模型进⾏测试。 结果在 results
文件夹中
# coding:utf8
import sys
import numpy as np
import cv2
import os
import dlib
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.autograd import Variable
import torchvision
from torchvision import datasets, models, transforms
import time
from PIL import Image
import torch.nn.functional as F
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')
PREDICTOR_PATH = "./Emotion_Recognition_File/face_detect_model/shape_predictor_68_face_landmarks.dat"
predictor = dlib.shape_predictor(PREDICTOR_PATH)
cascade_path = './Emotion_Recognition_File/face_detect_model/haarcascade_frontalface_default.xml'
cascade = cv2.CascadeClassifier(cascade_path)
if not os.path.exists("results"):
os.mkdir("results")
def standardization(data):
mu = np.mean(data, axis=0)
sigma = np.std(data, axis=0)
return (data - mu) / sigma
def get_landmarks(im):
rects = cascade.detectMultiScale(im, 1.3, 5)
x, y, w, h = rects[0]
rect = dlib.rectangle(int(x), int(y), int(x + w), int(y + h))
return np.matrix([[p.x, p.y] for p in predictor(im, rect).parts()])
def annotate_landmarks(im, landmarks):
im = im.copy()
for idx, point in enumerate(landmarks):
pos = (point[0, 0], point[0, 1])
cv2.putText(im,
str(idx),
pos,
fontFace=cv2.FONT_HERSHEY_SCRIPT_SIMPLEX,
fontScale=0.4,
color=(0, 0, 255))
cv2.circle(im, pos, 3, color=(0, 255, 255))
return im
testsize = 48 # 测试图大小
data_transforms = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
])
net = simpleconv3()
net.eval()
modelpath = "./models/model.ckpt" # 模型路径
net.load_state_dict(
torch.load(modelpath, map_location=lambda storage, loc: storage))
# 一次测试一个文件
img_path = "./Emotion_Recognition_File/find_face_img/"
imagepaths = os.listdir(img_path) # 图像文件夹
for imagepath in imagepaths:
im = cv2.imread(os.path.join(img_path, imagepath), 1)
try:
rects = cascade.detectMultiScale(im, 1.3, 5)
x, y, w, h = rects[0]
rect = dlib.rectangle(int(x), int(y), int(x + w), int(y + h))
landmarks = np.matrix([[p.x, p.y]
for p in predictor(im, rect).parts()])
except:
# print("没有检测到人脸")
continue # 没有检测到人脸
xmin = 10000
xmax = 0
ymin = 10000
ymax = 0
for i in range(48, 67):
x = landmarks[i, 0]
y = landmarks[i, 1]
if x < xmin:
xmin = x
if x > xmax:
xmax = x
if y < ymin:
ymin = y
if y > ymax:
ymax = y
roiwidth = xmax - xmin
roiheight = ymax - ymin
roi = im[ymin:ymax, xmin:xmax, 0:3]
if roiwidth > roiheight:
dstlen = 1.5 * roiwidth
else:
dstlen = 1.5 * roiheight
diff_xlen = dstlen - roiwidth
diff_ylen = dstlen - roiheight
newx = xmin
newy = ymin
imagerows, imagecols, channel = im.shape
if newx >= diff_xlen / 2 and newx + roiwidth + diff_xlen / 2 < imagecols:
newx = newx - diff_xlen / 2
elif newx < diff_xlen / 2:
newx = 0
else:
newx = imagecols - dstlen
if newy >= diff_ylen / 2 and newy + roiheight + diff_ylen / 2 < imagerows:
newy = newy - diff_ylen / 2
elif newy < diff_ylen / 2:
newy = 0
else:
newy = imagecols - dstlen
roi = im[int(newy):int(newy + dstlen), int(newx):int(newx + dstlen), 0:3]
roi = cv2.cvtColor(roi, cv2.COLOR_BGR2RGB)
roiresized = cv2.resize(roi,
(testsize, testsize)).astype(np.float32) / 255.0
imgblob = data_transforms(roiresized).unsqueeze(0)
imgblob.requires_grad = False
imgblob = Variable(imgblob)
torch.no_grad()
predict = F.softmax(net(imgblob))
print(predict)
index = np.argmax(predict.detach().numpy())
im_show = cv2.imread(os.path.join(img_path, imagepath), 1)
im_h, im_w, im_c = im_show.shape
pos_x = int(newx + dstlen)
pos_y = int(newy + dstlen)
font = cv2.FONT_HERSHEY_SIMPLEX
cv2.rectangle(im_show, (int(newx), int(newy)),
(int(newx + dstlen), int(newy + dstlen)), (0, 255, 255), 2)
if index == 0:
cv2.putText(im_show, 'none', (pos_x, pos_y), font, 1.5, (0, 0, 255), 2)
if index == 1:
cv2.putText(im_show, 'pout', (pos_x, pos_y), font, 1.5, (0, 0, 255), 2)
if index == 2:
cv2.putText(im_show, 'smile', (pos_x, pos_y), font, 1.5, (0, 0, 255), 2)
if index == 3:
cv2.putText(im_show, 'open', (pos_x, pos_y), font, 1.5, (0, 0, 255), 2)
# cv2.namedWindow('result', 0)
# cv2.imshow('result', im_show)
cv2.imwrite(os.path.join('results', imagepath), im_show)
# print(os.path.join('results', imagepath))
plt.imshow(im_show[:, :, ::-1]) # 这里需要交换通道,因为 matplotlib 保存图片的通道顺序是 RGB,而在 OpenCV 中是 BGR
plt.show()
# cv2.waitKey(0)
# cv2.destroyAllWindows()
tensor([[8.1330e-03, 6.7033e-04, 9.8497e-01, 6.2311e-03]],
grad_fn=)
再次说明:0 代表 none、 1 代表pouting、2 代表 smile、3 代表 openmouth
上面展示的图片上方会有一个输出,如:tensor([[8.1330e-03, 6.7033e-04, 9.8497e-01, 6.2311e-03]])
这个代表的含义是,该图片在这个模型预测下,是该类别的可能性,比如上面这个例子
9.8497e-01
是四个值最大的,它的索引是 2(从 0 开始算),所以预测该图片为smile
以上就是本训练营的全部内容了,希望通过本次训练营,可以帮助大家搭建起一个深度学习项目的框架,然后将所学的知识应用到自己的项目、工作中。谢谢!