智慧农业赛题——牛只图像分割竞赛

!unzip -qoa data/data127515/train_dataset.zip -d train

!unzip -qoa data/data127515/test_dataset_A.zip -d test

from PIL import Image
img=Image.open('train/200/images/60ac8ef2b7f3a3763b8340af.jpg')
img

!pip install paddlex

!mv train/200/data.json train/200/annotations.json

!paddlex --split_dataset --format COCO --dataset_dir train/200 --val_value 0.2

!ls train/200

# 统计mean和stdimport osimport cv2import numpy as npfrom tqdm import tqdm


img_list_1440 = []
img_list_576 = []
means = np.zeros(3)
stdevs = np.zeros(3)
imgs_path = "train/200/images"imgs_name = os.listdir(imgs_path)for idx, name in tqdm(enumerate(imgs_name)):    if name == ".ipynb_checkpoints":        continue
    img = cv2.cvtColor(cv2.imread(os.path.join(imgs_path, name)), cv2.COLOR_BGR2RGB)
    img = img[:, :, :, np.newaxis]    if img.shape[0] == 1440:
        img_list_1440.append(img)    else:
        img_list_576.append(img)for img_list in [img_list_1440, img_list_576]:
    imgs = np.concatenate(img_list, axis=-1)
    imgs = imgs.astype(np.float32) / 255.
    for i in tqdm(range(3)):
        pixels = imgs[:, :, i, :].ravel()
        means[i] += float(np.mean(pixels))
        stdevs[i] += float(np.std(pixels))print(means / 2, stdevs / 2)

# 准备数据集import paddlex as pdxfrom paddlex import transforms as T


train_transforms = T.Compose([
    T.MixupImage(mixup_epoch=-1), 
    T.RandomDistort(),
    T.RandomExpand(im_padding_value=[123.675, 116.28, 103.53]), 
    T.RandomCrop(),
    T.RandomHorizontalFlip(), 
    T.BatchRandomResize(
        target_sizes=[320, 352, 384, 416, 448, 480, 512, 544, 576, 608],
        interp="RANDOM"), 
    T.Normalize(
        mean=[0.46157165, 0.46698335, 0.46580717], std=[0.28069802, 0.28096266, 0.28323689])
])

eval_transforms = T.Compose([
    T.Resize(
        target_size=608, interp="CUBIC"), 
    T.Normalize(
        mean=[0.46157165, 0.46698335, 0.46580717], std=[0.28069802, 0.28096266, 0.28323689])
])

train_dataset = pdx.datasets.CocoDetection(
    data_dir="train/200",
    ann_file="train/200/train.json",
    transforms=train_transforms,
    shuffle=True)
eval_dataset = pdx.datasets.CocoDetection(
    data_dir="train/200",
    ann_file="train/200/val.json",
    transforms=eval_transforms)

import numpy as npimport paddlex as pdxfrom paddlex import transforms as T# 初始化模型并进行训练num_classes = len(train_dataset.labels)
model = pdx.det.MaskRCNN(
    num_classes=num_classes, 
    backbone="ResNet50_vd")
    
    

model.train(
    num_epochs=200,
    train_dataset=train_dataset,
    train_batch_size=20,
    eval_dataset=eval_dataset,
    pretrain_weights="COCO",
    learning_rate=0.005 / 12,
    warmup_steps=500,
    warmup_start_lr=0.0,
    save_interval_epochs=5,
    lr_decay_epochs=[85, 135],
    save_dir="output/maskrcnn_r50vd",
    use_vdl=True)

# 预测import osimport os.path as ospimport jsonimport numpy as npimport paddlex as pdxfrom tqdm import tqdmfrom pycocotools import mask as maskUtilsdef binary_mask_to_rle(binary_mask):
    binary_mask = np.asfortranarray(binary_mask)
    rle = maskUtils.encode(binary_mask)
    rle["counts"] = str(rle["counts"], encoding="utf-8")    return rle

model = pdx.load_model("output/maskrcnn_r50vd/best_model")
test_path = "test/images/"test_name = os.listdir(test_path)
res_json = []for name in tqdm(test_name):
    path = osp.join(test_path, name)    if path.split(".")[-1] != "png":        continue
    result = model.predict(path)    for res in result:        # print(res)
        # print(50*'*')
        res_json.append({            "image_id": osp.join("images", name),            "category_id": 1,            "segmentation": binary_mask_to_rle(res["mask"]),            "score": res["score"]
        })    # pdx.det.visualize(image_name, result, threshold=0.5, save_dir="./output_picture/maskrcnn_r50vd")with open("test.json", "w", encoding="utf-8") as f:
    json.dump(res_json, f)