该博客介绍了如何使用PaddlePaddle框架，结合PaddleHub的EfficientNet_B1模型进行人脸情绪识别的训练和预测。通过加载数据、设置数据集、定义模型、训练和验证过程，展示了Keras、PyTorch和Paddle三种框架的Baseline。代码中还包含了数据预处理、交叉验证以及模型保存和预测的实现。
各框架Baseline地址
- Kears and Pytorch：https://gitee.com/coggle/competition-baseline/tree/master/competition/科大讯飞AI开发者大赛2021/人脸情绪识别挑战赛
- Paddle: https://aistudio.baidu.com/aistudio/projectdetail/2194725?shared=1
本项目给出Paddle在组网时应用Paddlehub模型。目前网上均未找到此种方法，所以这里给出本人所采用的一种方法。
本项目基于阿水大佬的代码，采用Paddle实现。
Paddle代码解析
加载数据，处理
%cd /home/aistudio
!unzip ./data/data100388/Datawhale_人脸情绪识别_数据集.zip
%cd Datawhale_人脸情绪识别_数据集
!nohup unzip train.zip
!nohup unzip test.zip
导入模块
import paddle
import paddle.nn as nn
import cv2
from PIL import Image
import numpy as np
import pandas as pd
import glob
from sklearn.model_selection import train_test_split, StratifiedKFold, KFold
import time
import paddle.vision.transforms as transforms
import warnings
import paddlehub
# 忽略警告输出
warnings.filterwarnings("ignore")
paddle.set_device('gpu')
设置Dataset，自定义读取数据集
train_jpg = glob.glob('/home/aistudio/Datawhale_人脸情绪识别_数据集/train/*/*')
np.random.shuffle(train_jpg)
train_jpg = np.array(train_jpg)
class QRDataset(paddle.io.Dataset):
def __init__(self, img_path, transform=None):
self.img_path = img_path
if transform is not None:
self.transform = transform
else:
self.transform = None
def __getitem__(self, index):
start_time = time.time()
img = Image.open(self.img_path[index]).convert('RGB')
lbl_dict = {'angry': 0,
'disgusted': 1,
'fearful': 2,
'happy': 3,
'neutral': 4,
'sad': 5,
'surprised': 6}
if self.transform is not None:
img = self.transform(img)
if 'test' in self.img_path[index]:
return img, paddle.to_tensor(np.array(0))
else:
lbl_int = lbl_dict[self.img_path[index].split('/')[-2]]
return img, paddle.to_tensor(np.array(lbl_int))
def __len__(self):
return len(self.img_path)
模型组网
from EfficientNet_B1 import EfficientNet_B1
class XunFeiNet(nn.Layer):
def __init__(self):
super(XunFeiNet, self).__init__()
#efmodel = paddlehub.Module(name = "efficientnetb1_imagenet")
self.model = EfficientNet_B1()
self.fc = nn.Linear(1000, 7)
def forward(self, img):
out = self.model(img)
out = self.fc(out)
return out
对于EfficientNet_B1模块，是在本地建一个Python文件，具体需要的到下列Github路径中寻找：https://github.com/PaddlePaddle/PaddleHub/blob/release/v2.1/modules/image/classification/efficientnetb1_imagenet/module.py
设置验证、预测、训练代码，利用基础API
def validate(val_loader, model, criterion):
model.eval()
acc1 = []
with paddle.no_grad():
end = time.time()
for i, (input_, target) in enumerate(val_loader):
output = model(input_)
target = paddle.unsqueeze(paddle.to_tensor(target), 1)
loss = criterion(output, target)
#loss = nn.functional.sigmoid_focal_loss(output, target)
acc1.append(paddle.metric.accuracy(output, target))
print(' * Val Acc@1 {0}'.format(np.mean(acc1)))
return np.mean(acc1)
def predict(test_loader, model, tta=10):
model.eval()
test_pred_tta = None
for _ in range(tta):
test_pred = []
with paddle.no_grad():
end = time.time()
for i, (input_, target) in enumerate(test_loader):
output = model(input_)
target = paddle.unsqueeze(paddle.to_tensor(target), 1)
output = output.numpy()
test_pred.append(output)
test_pred = np.vstack(test_pred)
if test_pred_tta is None:
test_pred_tta = test_pred
else:
test_pred_tta += test_pred
return test_pred_tta
def train(train_loader, model, criterion, optimizer, epoch):
model.train()
end = time.time()
acc1 = []
for i, (input_, target) in enumerate(train_loader):
output = model(input_)
target = paddle.unsqueeze(paddle.to_tensor(target), 1)
loss = criterion(output, target)
#loss = nn.functional.sigmoid_focal_loss(output, target)
acc1.append(paddle.metric.accuracy(output, target))
loss.backward()
optimizer.step()
optimizer.clear_grad()
if i % 100 == 0:
print('Train: {0}'.format(np.mean(acc1)))
设置交叉验证训练，此处只用一折
skf = KFold(n_splits=5, random_state=233, shuffle=True)
for flod_idx, (train_idx, val_idx) in enumerate(skf.split(train_jpg, train_jpg)):
train_loader = paddle.io.DataLoader(
QRDataset(train_jpg[train_idx][:],
transforms.Compose([
transforms.ColorJitter(hue=.05, saturation=.05),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.Resize((196, 196)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
), batch_size=128, shuffle=True, num_workers=0, use_buffer_reader=True
)
val_loader = paddle.io.DataLoader(
QRDataset(train_jpg[val_idx][:1000],
transforms.Compose([
transforms.Resize((196, 196)),
# transforms.Resize((124, 124)),
# transforms.RandomCrop((88, 88)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
), batch_size=48, shuffle=False, num_workers=0,use_buffer_reader=True
)
model = XunFeiNet()
'''
model = paddlehub.Module(
name='resnet50_vd_imagenet_ssld',
label_list=["fearful", "surprised", "happy", "disgusted", "angry", "sad", "neutral"],
load_checkpoint=None)
'''
criterion = nn.CrossEntropyLoss()
optimizer = paddle.optimizer.SGD(0.01, parameters = model.parameters())
best_acc = 0.0
for epoch in range(100):
print('\nEpoch: ', epoch)
train(train_loader, model, criterion, optimizer, epoch)
val_acc = validate(val_loader, model, criterion)
if val_acc > best_acc:
best_acc = val_acc
paddle.save(model.state_dict(), './resnet18_fold{0}.pdparams'.format(flod_idx))
break
预测生成结果
import glob
import numpy as np
import paddle
import time
from PIL import Image
import paddle.vision.transforms as transforms
import paddlehub
import pandas as pd
class QRDataset(paddle.io.Dataset):
def __init__(self, img_path, transform=None):
self.img_path = img_path
if transform is not None:
self.transform = transform
else:
self.transform = None
def __getitem__(self, index):
start_time = time.time()
img = Image.open(self.img_path[index]).convert('RGB')
lbl_dict = {'angry': 0,
'disgusted': 1,
'fearful': 2,
'happy': 3,
'neutral': 4,
'sad': 5,
'surprised': 6}
if self.transform is not None:
img = self.transform(img)
if 'test' in self.img_path[index]:
return img, paddle.to_tensor(np.array(0))
else:
lbl_int = lbl_dict[self.img_path[index].split('/')[-2]]
return img, paddle.to_tensor(np.array(lbl_int))
def __len__(self):
return len(self.img_path)
def predict(test_loader, model, tta=10):
model.eval()
test_pred_tta = None
for _ in range(tta):
test_pred = []
with paddle.no_grad():
end = time.time()
for i, (input_, target) in enumerate(test_loader):
output = model(input_)[0]
target = paddle.unsqueeze(paddle.to_tensor(target), 1)
output = output.numpy()
test_pred.append(output)
test_pred = np.vstack(test_pred)
if test_pred_tta is None:
test_pred_tta = test_pred
else:
test_pred_tta += test_pred
return test_pred_tta
test_jpg = glob.glob('./Datawhale_人脸情绪识别_数据集/test/*')
test_jpg = np.array(test_jpg)
test_jpg.sort()
test_loader = paddle.io.DataLoader(
QRDataset(test_jpg,
transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
), batch_size=50, shuffle=False, num_workers=5
)
model = paddlehub.Module(
name='resnet50_vd_imagenet_ssld',
label_list=["fearful", "surprised", "happy", "disgusted", "angry", "sad", "neutral"],
load_checkpoint='resnet18_fold0.pdparams')
test_pred = predict(test_loader, model, 5)
cls_name = np.array(['angry', 'disgusted', 'fearful', 'happy','neutral', 'sad', 'surprised'])
submit_df = pd.DataFrame({'name': test_jpg, 'label': cls_name[test_pred.argmax(1)]})
submit_df['name'] = submit_df['name'].apply(lambda x: x.split('/')[-1])
submit_df = submit_df.sort_values(by='name')
submit_df.to_csv('paddle_submit.csv', index=None)
常见问题：
- GPU和显存利用率过低，但是还提示显存爆炸。解决：设置DataLoader，num_works=0。
- 。。。。暂无
扫一扫
3227
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
