趣味体验馆踩坑指南

import reimport base64import numpy as npimport paddle.fluid as fluidfrom PIL import Imagefrom io import BytesIO# 将 Base64 转成 PIL Image def base64_to_pil(image_base64):
    """read image from memory"""
    image_base64 = re.sub('^data:image/.+;base64,', '', image_base64)  # 需要去除头部格式信息
    image_mem = BytesIO(base64.b64decode(image_base64))  # python3
    image_pil = Image.open(image_mem).convert('RGB')    return image_pil# 预处理代码def preprocess(img):
    # 图像缩放
    max_size, min_size = 720, 32
    w, h = img.size    if max(h,w)>max_size:
        img = img.resize((max_size, int(h/w*max_size))) if h<w else img.resize((int(w/h*max_size), max_size))    elif min(h,w)<min_size:
        img = img.resize((min_size, int(h/w*min_size))) if h>w else img.resize((int(w/h*min_size), min_size))    # 裁剪图片
    w, h = img.size
    img = np.array(img)
    img = img[:h-(h%32), :w-(w%32), :]    # 归一化
    img = img/127.5 - 1.0
    
    return imgdef reader_infer(data_args):
    """
    reader inter
    :param data_args: 接口请求参数
    :return [[]], feeder
    """

    def reader():
        """
        reader
        :return:
        """
        # image <type 'image'> default value:None 
        image = data_args['image']        
        # 格式转换
        image = base64_to_pil(image)        
        # 预处理
        image = preprocess(image)        
        # 根据输入 Tensor 的名称和 Shape 构建 DataFeeder
        h, w, c = image.shape
        img = fluid.layers.data(name='x2paddle_generator_input', shape=[h, w, c], dtype='float32')
        feeder = fluid.DataFeeder([img], fluid.CPUPlace())        return [[image]], feeder    return reader

import reimport base64import numpy as npfrom PIL import Imagefrom io import BytesIO 


# PIL Image 转换为 Base64def pil_to_base64(image_pil):
    """save PIL image in memory, and convert to base64"""
    image_buffer = BytesIO()    # image_buffer = StringIO()  # python2
    image_pil.save(image_buffer, format='JPEG')
    image_bytes = image_buffer.getvalue()
    image_base64 = base64.b64encode(image_bytes).decode()    return image_base64# 后处理def postprocess(output):
    # 反归一化
    image = (output.squeeze() + 1.) / 2 * 255

    # 限幅
    image = np.clip(image, 0, 255).astype(np.uint8)    
    # 格式转换
    image = Image.fromarray(image)    
    return imagedef output(results, data_args):
    """
    模型评测结果输出转换
    :param results
    :param data_args 请求参数
    :return dict
    """
    # todo
    # 后处理
    image = postprocess(results[0])    
    # Base64 编码
    code = pil_to_base64(image)    
    # 加上文件头进行结果返回
    return {'image': 'data:image/jpeg;base64,'+code}