对baseline的模型进行了数据扩增和loss设计的改进,但不知为啥效果不好。
1.增加albumentations数据扩增方式:
def __getitem__(self, index):
img = Image.open(self.img_path[index]).convert('RGB')#PIL读取数据为‘RGBA’格式
img = np.asarray(img)
#img = cv2.imread(self.img_path[index])
#img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = Compose([
Blur(blur_limit=3),#模糊变换
OpticalDistortion(),#视觉扭曲变换
GaussNoise(), #高斯噪声
],p=1.0)(image=img)['image']
#img = cv2.GaussianBlur(img,(5,5),0) 高斯模糊变换,高斯矩阵长宽为(5,5),标准差为0
img = Image.fromarray(img)
if self.transform is not None:
img = self.transform(img)
2.对loss添加权重系数:
def train(train_loader, model, criterion, optimizer):
# 切换模型为训练模式
model.train()
train_loss = []
for i, (input, target) in enumerate(train_loader):
if use_cuda:
input = input.cuda()
target = target.cuda()
c0, c1, c2, c3, c4 = model(input)
loss = 0.1*criterion(c0, target[:, 0]) + \
0.1*criterion(c1, target[:, 1]) + \
0.1*criterion(c2, target[:, 2]) + \
0.2*criterion(c3, target[:, 3]) + \
0.5*criterion(c4, target[:, 4])
# loss /= 6
optimizer.zero_grad()
loss.backward()
optimizer.step()
#if i % 100 == 0:
# print(loss.item())
train_loss.append(loss.item())
return np.mean(train_loss)
3.cv2读取数据默认是BGR格式,转换成RGB格式为:
img = cv2.imread(‘cat.jpg’)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
4.cv2格式与PIL格式互转:
PIL转cv2
pil_image = PIL.Image.open(‘image.jpg’)
cv_image = cv2.cvtColor(numpy.asarray(pil_image), cv2.COLOR_RGB2BGR)
cv2转PIL
pil_im2 = Image.fromarray(img)