KO
|
EN
gitlite — search
Search
#python
#java
#python3
#arduino
#golang
#machine-learning
#rust
#html
#flask
#javascript
#seismology
#nodejs
Changeable
★ 18
Open GitHub ↗
基于pytorch的目标检测数据增强工具包。
Download README (.md)
Explore Similar Repositories
song_of_soul
:
No description available.
guiju
:
simple web project
cassie
:
Sketch complex curves and surfaces in VR
keestalkstech-code-gallery
:
This repository contains code samples for articles on my blog. It is easier to keep the code up to date when all of the items are grouped.
GB28181MProxy
:
支持国家标准 GB/T 28181-2016和拉取视频流实现缩放、码流控制、RTP推送
// repository documentation
Was this content helpful?
★ 0
(0 ratings)
Select Rating:
★
★
★
★
★
Submit Feedback
Recent Feedback
×
Download README
Do you want to download the
README.md
file for
Changeable
?
Download (.md)
# changeable 基于pytorch的目标检测数据增强工具包。 **VOC格式数据集** -> **transforms** -> **无限dotaloader** -> **mosaic数据增强** -> **anchors匹配** => **输出** : tuple(image[B, C, H, W], boxes[B, num_anchors * num_classes, 4], labels[B, num_anchors*num_classes], image_name[B]) | | | | ------------------------------------------------------------ | ------------------------------------------------------------ | |  |  | # install ```bash pip install changeable ``` 欢迎大家多提建议、bug。如果有想实现的数据增强方式，也可以提出来。 一有时间就会更新。 # Documentation ## dataloader 定义了类似于torch.data.Dataloader的 dataloader。 实现的功能： - 无限采样。（训练流程更好控制） - mosaic图片生成。 实现过程，采用先裁剪，后合成mosaic图片的方式。较直接覆盖，变化更加丰富。 可先对单张图片进行数据增强，然后合成mosaic图片，使用更加方便。 - anchor分配。（将boxes与label，和anchors对齐，对于不同目标数量的图片，boxes格式均相同，方便批次处理） ```python def dataloader(dataset, batch_size=1, resize: Tuple[int, int]=None, anchors_assigner=None, shuffle=True, num_workers=0, use_mosaic=False): """ 数据加载器。 通过指定的anchor分配器，将 可指定使用mosaic数据增强，将四张图组合成mosaic图片。 :param dataset: 数据集，torch.data.dataset，输出为(image:ndarry, boxes:ndarry, labels:ndarry, id:str) :param batch_size: 批次 :param resize: 归一化后的尺寸(使用mosaic时，mosaic图片尺寸；不使用mosaic时，resize填充灰边，保持宽高比) :param anchors_assigner: anchor分配器, AnchorsAssigner, :param shuffle: 是否打乱数据集 :param num_workers: 线程数 :param use_mosaic: 是否使用mosaic数据增强 :return: 提供anchor分配器: images: [B, C, W, H], boxes: [B, num_anchors, 4], labels: [B, num_anchors], 不提供anchor分配器: images: [1, C, W, H], boxes: [1, num_object, 4], labels: [1, num_object], """ ``` ## mosaic 在合成mosaic时，读入四张图片。先随机生成图片交汇点；然后计算各位置图片所需尺寸；对四张图片分别进行尺寸裁剪，并去除中心在裁剪区域外的标注框；合并生成mosaic图片。 mosaic图片生成，在dataloader中进行，因而可以在合成前，通过dataset中transforms对单张图片进行数据增强。 | | | | | | | :--------------------: | :--------------------------------------: | :--------------------------------------: | :--------------------------------------: | :--------------------------------------: | | mosaic |  |  |  |  | | mosaic with trasnforms |  |  |  |  | ## AnchorsAssigner （anchors分配） **每个标注框都最少含有一个与之最接近的anchor**。 ```python class AnchorsAssignerIOU(AnchorsAssigner): def __init__(self, anchors: AnchorsGenerator, threshold: float=0.6): # 按照iou阈值分配anchors。参数分别为Anchors、iou阈值（iou大于该阈值为匹配成功）。 def __call__(self, gt_boxes: Union[ndarray, Tensor], gt_labels: Union[ndarray, Tensor]): # 标注框与标签。标注框格式'xyxy' ``` ```python class AnchorsAssignerWH(AnchorsAssigner): def __init__(self, anchors: AnchorsGenerator, threshold: float=3): # 按照宽高比分配anchors。参数分别为Anchors、宽高比阈值（宽高比小于该阈值为匹配成功）。 def __call__(self, gt_boxes: Union[ndarray, Tensor], gt_labels: Union[ndarray, Tensor]): # 标注框与标签。标注框格式'xyxy' ``` | | | | | | :-----------------: | :--------------------------------: | :--------------------------------: | :--------------------------------: | | 按IOU分配anchors |  |  |  | | AnchorsAssignerIOU | threshold=0.5 | threshold=0.5 | threshold=0.5 | | 按宽高比分配anchors |  |  |  | | AnchorsAssignerWH | threshold=3 | threshold=3 | threshold=3 | ## AnchorsGenerator (anchors生成) ```python class AnchorsGenerator(object): def __init__(self, image_size: Tuple[int, int], # 输入图片尺寸 feature_maps_size: Tuple[Tuple[int, int], ...], # 多层特征图尺寸 anchors_size: Tuple[Tuple[Tuple[float, float], ...], ...], # 每层特征图上anchor尺寸 form: str='xyxy', # xyxy or cxcywh clip: bool=True): # 超出图片anchor是否截断 ``` ```python eg: anchors = AnchorsGenerator(image_size=(300, 300), feature_maps_size=((76, 76), (38, 38), (19, 19)), anchors_size=(((10, 13), (16, 30), (33, 23)), ((30, 61), (62, 45), (59, 119)), ((116, 90), (156, 198), (373, 326))), form='xyxy', clip=True ) -> anchors: [[ 0. 0. 6.97368421 8.47368421] [ 0. 0. 9.97368421 16.97368421] [ 0. 0. 18.47368421 13.47368421] ... [234.10526316 247.10526316 300. 300. ] [214.10526316 193.10526316 300. 300. ] [105.60526316 129.10526316 300. 300. ]] ``` ## Transforms （数据增强） 所实现的数据增强方式，均是针对目标检测任务。 输入均为 image, boxes, labels - ToTensor() # 转换npndarry为tensor。（一般不需要使用，在dataloader中会自动转换） - ConvertBoxesToValue() # 将boxes从百分比形式转换为值形式。 - ConvertBoxesToPercentage() # 将boxes从值形式转换为百分比形式。 - ConvertBoxesForm(from_form: str, to_form: str) # 转换boxes格式。'xyxy' 或 'cxcywh' - Resize(size: Tuple[int, int]) # resize - AdaptiveResize(size: Tuple[int, int], value: Union[Tuple[int, int, int], Tuple[Tuple[int, int, int], ...]]=(114, 114, 114)) # 自适应resize。保持图像宽高比，短边补黑边。 - Scaled(scale: Union[float, Tuple[float, float]]) # 放大缩小图片。输入缩放比例值或缩放比例范围。 - CropIou(iou:Union[float, Tuple[float, ...]]) # 按指定最小iou截图图片。指定一个或多个iou值。 - CropSize(size:Union[Tuple[int, int], Tuple[Tuple[int, int], ...]]) # 按指定尺寸截取图片。图片尺寸小于截取尺寸时，先resize大图片，然后截取。 - SubtractMeans(mean: Tuple[float, float, float]) # 减均值 - DivideStds(std: Tuple[float, float, float]) # 除方差 - GaussNoise(scale: Union[float, Tuple[float, float]]=(0, 0.1), probability: float=0.5) # 高斯噪声。指定一个值或一个范围 - SalePepperNoise(sale_scale: Union[float, Tuple[float, float]]=(0, 0.03), pepper_scale: Union[float, Tuple[float, float]]=(0, 0.03), probability: float=0.5) # 椒盐噪声。参数分别为盐噪声比例与椒噪声比例，指定一个值或一个范围。 - GaussBlur(ksize: Union[int, Tuple[int, ...]]=(3, 5, 7, 9), probability: float=0.5) # 高斯平滑（高斯模糊）。高斯核大小，指定一个值或多个值，必须为正奇数。 - MotionBlue(ksize: Union[int, Tuple[int, ...]]=(3, 5, 7, 9, 11), angle: Union[int, Tuple[int, int]] = (0, 30), probability: float=0.5) # 运动平滑（运动模糊）。参数为核大小与运动角度。核大小，指定一个或多个值，必须为正奇数；运动角度，指定一个值或一个范围。 - Cutout(num: Union[int, Tuple[int, ...]]=(1, 2, 3, 4), size: Union[Tuple[float, float], Tuple[Tuple[float, float], ...]]=((0.2, 0.2), (0.3, 0.3)), value:Union[Tuple[int, int, int], Tuple[Tuple[int, int, int], ...]]=(114, 114, 114), cover: float=0.5, probability: float=0.5) # 遮挡。参数为遮挡块数量、遮挡块尺寸系数、遮挡块填充值与遮挡阈值。遮挡块数量，指定一个值或多个值；遮挡块尺寸，指定一组值或多组值（宽高）；遮挡块填充值，指定一组值或多种值。 - RandomFlipLR(probability: float=0.5) # 左右翻转。 - RandomFlipUD(probability: float=0.5) # 上下翻转。 - ShuffleChannels(mode: Union[Tuple[int, int, int], Tuple[Tuple[int, int, int], ]]=((0, 2, 1), (1, 0, 2), (1, 2, 0),(2, 0, 1), (2, 1, 0)), probability: float=0.5) # 交换通道。指定一组值或多组值。 - ChangeContrast(scale:Union[float, Tuple[float, float]]=(-0.5, 0.5), probability: float=0.5) # 对比度调整。指定一个值或一个范围。 - ChangeHue(scale: Union[int, Tuple[int, int]]=(0, 360), probability: float=0.5) # 色调调整。指定一个值或一个范围。 - ChangeSaturation(scale: Union[float, Tuple[float, float]]=(-1, 1), probability: float=0.5) # 饱和度调整。指定一个值或一个范围。 - ChangeBrightness(scale: Union[float, Tuple[float, float]]=(-0.3, 0.3), probability: float=0.5) # 亮度调整。指定一个值或一个范围。 - Compose(transforms) ### 色彩调整 | | | | | | | | | | :--------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | | 亮度 | <img src="./images/ChangeBrightness_scale_-0.3.png" style="zoom:50%;" /> | <img src="./images/ChangeBrightness_scale_-0.2.png" style="zoom:50%;" /> | <img src="./images/ChangeBrightness_scale_-0.1.png" style="zoom:50%;" /> | <img src="./images/ChangeBrightness_scale_0.png" style="zoom:50%;" /> | <img src="./images/ChangeBrightness_scale_0.1.png" style="zoom:50%;" /> | <img src="./images/ChangeBrightness_scale_0.2.png" style="zoom:50%;" /> | <img src="./images/ChangeBrightness_scale_0.3.png" style="zoom:50%;" /> | | ChangeBrightness | -0.3 | -0.2 | -0.1 | 0 | 0.1 | 0.2 | 0.3 | | 对比度 | <img src="./images/ChangeContrast_scale_-0.5.png" style="zoom:50%;" /> | <img src="./images/ChangeContrast_scale_-0.3.png" style="zoom:50%;" /> | <img src="./images/ChangeContrast_scale_-0.2.png" style="zoom:50%;" /> | <img src="./images/ChangeContrast_scale_0.png" style="zoom:50%;" /> | <img src="./images/ChangeContrast_scale_0.2.png" style="zoom:50%;" /> | <img src="./images/ChangeContrast_scale_0.3.png" style="zoom:50%;" /> | <img src="./images/ChangeContrast_scale_0.5.png" style="zoom:50%;" /> | | ChangeContrast | -0.5 | -0.3 | -0.2 | 0 | 0.2 | 0.3 | 0.5 | | 色调 | <img src="./images/ChangeHue_scale_0.png" style="zoom:50%;" /> | <img src="./images/ChangeHue_scale_60.png" style="zoom:50%;" /> | <img src="./images/ChangeHue_scale_120.png" style="zoom:50%;" /> | <img src="./images/ChangeHue_scale_180.png" style="zoom:50%;" /> | <img src="./images/ChangeHue_scale_240.png" style="zoom:50%;" /> | <img src="./images/ChangeHue_scale_300.png" style="zoom:50%;" /> | <img src="./images/ChangeHue_scale_360.png" style="zoom:50%;" /> | | ChangeHue | 0 | 60 | 120 | 180 | 240 | 300 | 360 | | 饱和度 | <img src="./images/ChangeSaturation_scale_-1.png" style="zoom:50%;" /> | <img src="./images/ChangeSaturation_scale_-0.5.png" style="zoom:50%;" /> | <img src="./images/ChangeSaturation_scale_-0.3.png" style="zoom:50%;" /> | <img src="./images/ChangeSaturation_scale_0.png" style="zoom:50%;" /> | <img src="./images/ChangeSaturation_scale_0.3.png" style="zoom:50%;" /> | <img src="./images/ChangeSaturation_scale_0.5.png" style="zoom:50%;" /> | <img src="./images/ChangeSaturation_scale_1.png" style="zoom:50%;" /> | | ChangeSaturation | -1 | -0.5 | -0.3 | 0 | 0.3 | 0.5 | 1 | | 通道交换 | <img src="./images/ShuffleChannels_mode_0_1_2.png" style="zoom:50%;" /> | <img src="./images/ShuffleChannels_mode_0_2_1.png" style="zoom:50%;" /> | <img src="./images/ShuffleChannels_mode_1_0_2.png" style="zoom:50%;" /> | <img src="./images/ShuffleChannels_mode_1_2_0.png" style="zoom:50%;" /> | <img src="./images/ShuffleChannels_mode_2_0_1.png" style="zoom:50%;" /> | <img src="./images/ShuffleChannels_mode_2_1_0.png" style="zoom:50%;" /> | | | ShuffleChannels | (0, 1, 2) | (0, 2, 1) | (1, 0, 2) | (1, 2, 0) | (2, 0, 1) | (2, 1, 0) | | ### 翻转 | | | | | | :----------: | :--------------------------------------: | :----------: | :--------------------------------------: | | 左右翻转 | <img src="./images/RandomFlipLR.png" /> | 上下翻转 | <img src="./images/RandomFlipUD.png" /> | | RandomFlipLR | | RandomFlipUD | | ### 噪声 | | | | | | | | | | :-------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | | 椒盐噪声 | <img src="./images/SalePepperNoise_sale_0.01_pepper_0.01.png" style="zoom:50%;" /> | <img src="./images/SalePepperNoise_sale_0.02_pepper_0.02.png" style="zoom:50%;" /> | <img src="./images/SalePepperNoise_sale_0.03_pepper_0.03.png" style="zoom:50%;" /> | <img src="./images/SalePepperNoise_sale_0_pepper_0.01.png" style="zoom:50%;" /> | <img src="./images/SalePepperNoise_sale_0.01_pepper_0.png" style="zoom:50%;" /> | <img src="./images/SalePepperNoise_sale_0_pepper_0.02.png" style="zoom:50%;" /> | <img src="./images/SalePepperNoise_sale_0.02_pepper_0.png" style="zoom:50%;" /> | | SalePepperNoise | sale=0.01, pepper=0.01 | sale=0.02, pepper=0.02 | sale=0.03, pepper=0.03 | sale=0, pepper=0.01 | sale=0.01, pepper=0 | sale=0, pepper=0.02 | sale=0.02, pepper=0 | | 高斯噪声 | <img src="./images/GaussNoise_scale_0.01.png" style="zoom:50%;" /> | <img src="./images/GaussNoise_scale_0.05.png" style="zoom:50%;" /> | <img src="./images/GaussNoise_scale_0.1.png" style="zoom:50%;" /> | <img src="./images/GaussNoise_scale_0.2.png" style="zoom:50%;" /> | <img src="./images/GaussNoise_scale_0.3.png" style="zoom:50%;" /> | | | | GaussNoise | 0.01 | 0.05 | 0.1 | 0.2 | 0.3 | | | ### 模糊 | | | | | | | | | | :--------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | | 运动模糊 | <img src="./images/MotionBlue_ksize_3_angle_0.png" style="zoom:50%;" /> | <img src="./images/MotionBlue_ksize_5_angle_0.png" style="zoom:50%;" /> | <img src="./images/MotionBlue_ksize_7_angle_0.png" style="zoom:50%;" /> | <img src="./images/MotionBlue_ksize_9_angle_0.png" style="zoom:50%;" /> | <img src="./images/MotionBlue_ksize_11_angle_0.png" style="zoom:50%;" /> | <img src="./images/MotionBlue_ksize_11_angle_30.png" style="zoom:50%;" /> | <img src="./images/MotionBlue_ksize_11_angle_60.png" style="zoom:50%;" /> | | MotionBlue | ksize=3, angle=0 | ksize=5, angle=0 | ksize=7, angle=0 | ksize=9, angle=0 | ksize=11, angle=0 | ksize=11, angle=30 | ksize=11, angle=60 | | 高斯模糊 | <img src="./images/GaussBlur_ksize_1.png" style="zoom:50%;" /> | <img src="./images/GaussBlur_ksize_3.png" style="zoom:50%;" /> | <img src="./images/GaussBlur_ksize_5.png" style="zoom:50%;" /> | <img src="./images/GaussBlur_ksize_7.png" style="zoom:50%;" /> | <img src="./images/GaussBlur_ksize_9.png" style="zoom:50%;" /> | | | | GaussBlur | ksize=1 | ksize=3 | ksize=5 | ksize=7 | ksize=9 | | | ### 遮挡 | | | | | | | :----: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | | 遮挡 | <img src="./images/Cutout_num_1_size_03_03_value_114_114_114.png" style="zoom:50%;" /> | <img src="./images/Cutout_num_2_size_03_03_value_114_114_114.png" style="zoom:50%;" /> | <img src="./images/Cutout_num_2_size_02_015_value_0_0_0.png" style="zoom:50%;" /> | <img src="./images/Cutout_num_3_size_015_015_value_114_114_255.png" style="zoom:50%;" /> | | Cutout | num=1,size=(0.3, 0.3), value=(114,114,114), cover=0.5 | num=2,size=(0.3, 0.3), value=(114,114,114), cover=0.5 | num=2,size=(0.2, 0.15), value=(0,0,0), cover=0.5 | num=3,size=(0.15, 0.15), value=(114,114,255) | ### 裁剪 | | | | | | | | | :---------------: | :-------------------------------------: | :-------------------------------------: | :-------------------------------------: | :-------------------------------------: | :-------------------------------------: | :-------------------------------------: | | 按指定尺寸裁剪 |  |  |  |  |  |  | | CropSize | size=(100,100) | size=(300,300) | size=(300,400) | size=(400,300) | size=(400,400) | size=(500,500) | | 按指定最小iou裁剪 |  |  |  | | | | | CropIou | iou=0.3 | iou=0.5 | iou=0.7 | | | | ### 尺寸调整 | | | | | | | :------------: | :----------------------------------------------: | :----------------------------------------------: | :----------------------------------------------: | :----------------------------------------------: | | |  |  |  |  | | Resize | size=(300, 300) | size=(300, 400) | size=(400, 300) | size=(400, 400) | | |  |  |  |  | | AdaptiveResize | size=(300, 300) | size=(300, 400) | size=(400, 300) | size=(400, 400) | | 缩放 |  |  |  |  | | Scaled | scale=0.5 | scale=0.7 | scale=1.0 | scale=1.2 | # example ```python from changeable.dataset import VOCDataset from changeable.dataloader import dataloader from changeable.transforms import * from changeable.anchor import AnchorsAssignerWH, AnchorsGenerator from changeable.utils.display import draw_boxes, plot_image with open('classes.txt', 'r')as f: # 类别名文件，每行一个类别名 lines = f.readlines() classes_name = tuple([line.rstrip('\n') for line in lines]) dataset = VOCDataset(root='voc_root', # voc数据集根目录 classes_name=classes_name, is_train=True, transforms=Compose([ # 这里添加了所有的数据增强方式，只做例子演示用。 Resize((300, 300)), AdaptiveResize((300, 300)), Scaled(1.1), CropIou(0.5), CropSize((300, 300)), DivideStds((1,1,1)), SubtractMeans((0,0,0)), GaussNoise(), SalePepperNoise(), GaussBlur(), MotionBlue(), Cutout(), RandomFlipLR(), RandomFlipUD(), ShuffleChannels(), ChangeContrast(), ChangeHue(), ChangeBrightness(), ChangeSaturation(), ConvertBoxesToPercentage(), ConvertBoxesToValue(), ConvertBoxesForm('xyxy', 'cxcywh'), ConvertBoxesForm('cxcywh', 'xyxy'), ]) ) anchors = AnchorsGenerator(image_size=(600, 600), feature_maps_size=((76, 76), (38, 38), (19, 19)), anchors_size=(((10, 13), (16, 30), (33, 23)), ((30, 61), (62, 45), (59, 119)), ((116, 90), (156, 198), (373, 326))), form='xyxy', clip=True ) anchors_assigner = AnchorsAssignerWH(anchors, 3) loader = dataloader(dataset, batch_size=4, resize=(600, 600), use_mosaic=True, anchors_assigner=anchors_assigner, shuffle=True, num_workers=8) for i, (img, box, lab, ids) in enumerate(loader): print(i) print(img.size()) print(box.size()) print(lab.size()) img, box, lab, id = img[0], box[0], lab[0], ids[0] img = img.permute((1, 2, 0)).numpy() box, lab = box.numpy(), lab.numpy() box, lab = box[lab>0], lab[lab>0] img = draw_boxes(img, box, lab, label_name=classes_name) plot_image(img) ```
