飞桨常规赛：黄斑中央凹定位（GAMMA挑战赛任务二） - 11月第3名方案

class MyDataset(paddle.io.Dataset):
    def __init__(self, img_folder, label_file=None, idx=None, argument=False):
        '''设置图片目录、标签文件、样本编号（用于训练集和验证集划分）、数据增强等'''
        img_paths = get_all_paths(img_folder, '.jpg')        if idx is None:
            idx = np.arange(len(img_paths))
        imgs = [read_image(img_paths[i]) for i in idx]        if label_file:
            label_df = pd.read_excel(label_file)
            label_X = label_df['Fovea_X'].values
            label_Y = label_df['Fovea_Y'].values
            labels = [(label_X[i], label_Y[i]) for i in idx]        else:
            labels = [None,]*len(idx)
        datas = [(img, label) for img, label in zip(imgs, labels)]
        self.datas = datas
        self.argument = argument
        self.color_jitter = paddle.vision.transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1)    def __getitem__(self, idx):
        '''返回的param用于预测时恢复原始图像坐标'''
        img, loc = self.datas[idx]
        img = img.copy()        if self.argument:
            img, loc = self.do_arg(img, loc)
        img, label, param = self.pre_process(img, loc)        return img, label, param    def __len__(self):
        return len(self.datas)    def do_arg(self, img, loc):
        '''
            数据增强方式包括：颜色抖动、水平翻转、垂直翻转。
            数据增强后对变换参数作相应改动。
        '''
        h, w, _ = img.shape
        img = self.color_jitter(img)        if np.random.uniform(0,1)>0.5:            # horizon flip
            img[:,::-1,:] = img
            loc = (w-loc[0], loc[1])        if np.random.uniform(0,1)>0.5:            # vertical flip
            img[::-1,:,:] = img
            loc = (loc[0], h-loc[1])        return img, loc    def pre_process(self, img, loc):
        '''图像归一化到[-0.5, 0.5]区间，标签归一化到[0, 1]区间'''
        img, loc, param = crop_center(img, loc)
        img = img.transpose([2,0,1]).astype(np.float32)/255-0.5
        label = np.zeros([2], dtype=np.float32)        if loc is not None:
            fx, fy = loc
            r0, r1 = OUTPUT_RANGE
            w = (fx-r0)/(r1-r0)
            h = (fy-r0)/(r1-r0)
            label[:] = (h, w)        return img, label, param

def crop_center(img, loc=None):
    h, w, d = img.shape
    dx = (w-h)//2
    img = img[:, dx:(dx+h), :]
    img = cv2.resize(img, (INPUT_IMAGE_SHAPE, INPUT_IMAGE_SHAPE))
    scale = h/INPUT_IMAGE_SHAPE    if loc is not None:
        x, y = loc
        loc = ((x-dx)/h, y/h)    return img, loc, (scale, dx)def restore_xy(loc, param):
    x, y = loc
    scale, dx = param
    loc = (x*INPUT_IMAGE_SHAPE*scale+dx, y*INPUT_IMAGE_SHAPE*scale)    return loc

self.backbone = paddleseg.models.backbones.ResNet50_vd(pretrained=pretrain_url)

feats = self.backbone(x)
feats = [feats[i] for i in self.feat_indices]

class MyLoss(paddle.nn.Layer):
    def __init__(self, loss_w = LOSS_WEIGHTS):
        super(MyLoss, self).__init__()
        self.loss_w = loss_w    def forward(self, preds, label):
        loss = 0
        for pred, w in zip(preds, self.loss_w):
            loss += w*paddle.nn.functional.smooth_l1_loss(pred, label)        return loss

vdl_writer = visualdl.LogWriter(vdl_path)

vdl_writer.add_scalar('train/loss', loss_scalar, iters)

# 安装paddleseg!pip install paddleseg

# 下载比赛数据，报名比赛后在比赛页面可查看下载地址!wget https://xxxxxxxxxxxxxxx/task2_Fovea_localization.zip# 解压至competition_data文件夹!unzip task2_Fovea_localization.zip -d competition_data

!cd ~/work && python -W ignore train.py

!cd ~/work && python -W ignore predict.py