简单处理API
读取图像:
image.imdecode(open('../img/cat1.jpg', 'rb').read())
图像类型转换:
img.astype('float32')
图像增强流程
具体增强方式教程有很详细的示意,不再赘述
辅助函数,用于将增强函数应用于单张图片:
def apply_aug_list(img, augs):
for f in augs:
img = f(img)
return img
对于训练图片我们随机水平翻转和剪裁。对于测试图片仅仅就是中心剪裁。我们假设剪裁成28×28×3用于输入网络:
train_augs = [
image.HorizontalFlipAug(.5),
image.RandomCropAug((28,28))
] test_augs = [
image.CenterCropAug((28,28))
]
使用如下闭包来增强:
def get_transform(augs):
def transform(data, label):
# data: sample x height x width x channel
# label: sample
data = data.astype('float32')
if augs is not None:
# apply to each sample one-by-one and then stack
data = nd.stack(*[
apply_aug_list(d, augs) for d in data])
data = nd.transpose(data, (0,3,1,2))
return data, label.astype('float32')
return transform
基本逻辑就是这样。