6.1.5.1. hat.data¶

Main data module for training in HAT, which contains datasets, transforms, samplers.

6.1.5.1.1. Data¶

6.1.5.1.1.1. collates¶

`collate_2d`	Merge a list of samples to form a mini-batch of Tensor(s).
`collate_3d`	Merge a list of samples to form a mini-batch of Tensor(s).
`collate_psd`	Merge a list of samples to form a mini-batch of Tensor(s).
`CocktailCollate`	CocktailCollate.
`collate_lidar`	Merge a list of samples to form a mini-batch of Tensor(s).
`collate_2d_with_diff_im_hw`	Merge a list of samples to form a mini-batch of Tensor(s).
`collate_seq_with_diff_im_hw`	Merge a list of samples to form a mini-batch of Tensor(s).
`collate_nlu_with_pad`	Collate nlu func for dataloader.
`collate_2d_replace_empty`	Merge a list of samples to form a mini-batch of Tensor(s).
`default_collate_v2`	Entend torch.utils.data.default_collate.
`collate_2d_pad`	Merge a list of samples to form a mini-batch of Tensor(s).
`collate_mot_seq`	Collate for mot seq data.
`collate_2d_cat`	Merge a list of samples to form a mini-batch of Tensor(s).
`collate_lidar3d`

6.1.5.1.1.2. dataloaders¶

PassThroughDataLoader

Directly pass through input example.

6.1.5.1.1.3. datasets¶

`Cityscapes`	Cityscapes provides the method of reading cityscapes data from target pack type.
`CityscapesPacker`	CityscapesPacker is used for converting Cityscapes dataset in torchvision to target DataType format.
`RepeatDataset`	A wrapper of repeated dataset.
`ComposeDataset`	Dataset wrapper for multiple datasets with precise batch size.
`DistributedComposeRandomDataset`	Dataset wrapper for multiple datasets fair sample weights accross multi workers in a distributed environment.
`ResampleDataset`	A wrapper of resample dataset.
`ConcatDataset`	A wrapper of concatenated dataset with group flag.
`ImageNet`	ImageNet provides the method of reading imagenet data from target pack type.
`ImageNetPacker`	ImageNetPacker is used for converting ImageNet dataset in torchvision to DataType format.
`ImageNetFromImage`	ImageNet from image by torchvison.
`Kitti3D`	Kitti3D provides the method of reading kitti3d data from target pack type.
`Kitti3DReader`	Kitti3D dataset processor.
`Kitti3DDetectionPacker`	Kitti3DDetectionPacker is used for converting kitti3D dataset to target DataType format.
`Kitti3DDetection`	Kitti 3D Detection Dataset.
`Coco`	Coco provides the method of reading coco data from target pack type.
`CocoDetection`	Coco Detection Dataset.
`CocoDetectionPacker`	CocoDetectionPacker is used for packing coco dataset to target format.
`CocoFromImage`	Coco from image by torchvision.
`RandDataset`
`SimpleDataset`
`PascalVOC`	PascalVOC provides the method of reading voc data from target pack type.
`VOCDetectionPacker`	VOCDetectionPacker is used for packing voc dataset to target format.
`VOCFromImage`	VOC from image by torchvision.
`BatchTransformDataset`	Dataset which uses different transforms in different epochs.

6.1.5.1.1.4. samplers¶

`DistributedCycleMultiDatasetSampler`	In one epoch period, do cyclic sampling on the dataset according to iter_time.
`DistSamplerHook`	The hook api for torch.utils.data.DistributedDampler.
`SelectedSampler`	Distributed sampler that supports user-defined indices.
`DistributedGroupSampler`	Sampler that restricts data loading to a subset of the dataset.

6.1.5.1.1.5. transforms¶

`ConvertLayout`	ConvertLayout is used for layout convert.
`BgrToYuv444`	BgrToYuv444 is used for color format convert.
`BgrToYuv444V2`	BgrToYuv444V2 is used for color format convert.
`OneHot`	OneHot is used for convert layer to one-hot format.
`LabelSmooth`	LabelSmooth is used for label smooth.
`TimmTransforms`	Transforms of timm.
`TimmMixup`	Mixup of timm.
`Resize`	Resize image & bbox & mask & seg.
`Resize3D`	Resize 3D labels.
`RandomFlip`	Flip image & bbox & mask & seg & flow.
`Pad`
`Normalize`	Normalize image.
`RandomCrop`
`ToTensor`	Convert objects of various python types to torch.Tensor and convert the img to yuv444 format if to_yuv is True.
`Batchify`
`FixedCrop`	Crop image with fixed position and size.
`PresetCrop`	Crop image with preset roi param.
`ColorJitter`	Randomly change the brightness, contrast, saturation and hue of an image.
`DetInputPadding`
`AugmentHSV`	Random add color disturbance.
`RandomExpand`	Random expand the image & bboxes.
`MinIoURandomCrop`	Random crop the image & bboxes, the cropped patches have minimum IoU requirement with original image & bboxes, the IoU threshold is randomly selected from min_ious.
`ToFasterRCNNData`	Prepare faster-rcnn input data.
`ToLdmkRCNNData`	Transform dataset to RCNN input need.
`ToMultiTaskFasterRCNNData`	Convert multi-classes detection data to multi-task data.
`PadTensorListToBatch`	List of image tensor to be stacked vertically.
`HueSaturationValue`	Randomly change hue, saturation and value of the input image.
`RGBShift`	Randomly shift values for each channel of the input image.
`ReformatLanePolygon`
`PolygonToMask`
`MeanBlur`	Apply mean blur to the input image using a fix-sized kernel.
`MedianBlur`	Apply median blur to the input image using a fix-sized kernel.
`Mosaic`	Mosaic augmentation for detection task.
`RandomBrightnessContrast`	Randomly change brightness and contrast of the input image.
`ShiftScaleRotate`	Randomly apply affine transforms: translate, scale and rotate the input.
`RandomResizedCrop`	Torchvision's variant of crop a random part of the input, and rescale it to some size.
`AlbuImageOnlyTransform`	AlbuImageOnlyTransform used on img only.
`BoxJitter`	Jitter box to simulate the box predicted by the model.
`RandomSizeCrop`
`PlainCopyPaste`	Copy and paste instances plainly.
`SegRandomCrop`	Random crop on data with gt_seg label, can only be used for segmentation
`SegReWeightByArea`	Calculate the weight of each category according to the area of each category.
`LabelRemap`	Remap labels.
`SegOneHot`	OneHot is used for convert layer to one-hot format.
`SegResize`	Apply resize for both image and label.
`SegResizeAffine`	Resize image & seg.
`SegRandomAffine`	Apply random for both image and label.
`Scale`	Scale input according to a scale list.
`FlowRandomAffineScale`
`SegRandomCutOut`	CutOut operation for segmentation task.
`ListToDict`	Convert list args to dict.
`DeleteKeys`	Delete keys in input dict.
`RenameKeys`	Rename keys in input dict.
`Undistortion`	Convert a `PIL Image` or `numpy.ndarray` to
`PILToTensor`	Convert PIL Image to Tensor.
`TensorToNumpy`	Convert tensor to numpy.
`RandomSelectOne`	Select one of transforms to apply.
`MultiTaskAnnoWrapper`	Wrapper for multi-task anno generating.
`ConvertDataType`	Convert data type.
`RandomGray`	Transform RGB or BGR format into Gray format.
`JPEGCompress`	Do JPEG compression to downgrade image quality.
`SpatialVariantBrightness`	Spatial variant brightness, Enhanced Edition.
`GaussianBlur`	Randomly add guass blur on an image.
`MotionBlur`	Randomly add motion blur on an image.
`RandomDownSample`	First downsample and upsample to original size.
`Contrast`	Randomly jitters image contrast with a factor.
`GazeYUVTransform`	YUVTransform for Gaze Task.
`GazeRandomCropWoResize`	Random crop without resize.
`Clip`	Clip Data to [minimum, maximum].
`RandomColorJitter`	Randomly change the brightness, contrast, saturation and hue of an image.
`GazeRotate3DWithCrop`	Random rotate image, calculate ROI and random crop if necessary.
`eye_ldmk_mirror`	Flip eye landmarks.
`IterableDetRoIListTransform`	Iterable transformer base on roi list for object detection.
`IterableDetRoITransform`	Iterable transformer base on rois for object detection.
`PadDetData`
`DetAffineAugTransformer`	Affine augmentation for object detection.
`SeqRandomFlip`	Flip image & bbox & mask & seg & flow for sequence.
`SeqAugmentHSV`	Random add color disturbance for sequence.
`SeqResize`
`SeqPad`
`SeqToFasterRCNNData`
`SeqAlbuImageOnlyTransform`
`SeqBgrToYuv444`	BgrToYuv444 for sequence.
`SeqToTensor`	ToTensor for sequence.
`SeqNormalize`	Normalize for sequence.
`SeqRandomSizeCrop`	RandomSizeCrop for sequence.

6.1.5.1.1.5.1. lidar_utils¶

`DBFilterByDifficulty`	Filter sampled data by diffculties.
`DBFilterByMinNumPoint`	Filter sampled data by NumPoint.
`DataBaseSampler`
`VoxelGenerator`
`ObjectSample`	Sample GT objects to the data.
`ObjectNoise`	Apply noise to each GT objects in the scene.
`PointRandomFlip`	Flip the points & bbox.
`PointGlobalRotation`	Apply global rotation to a 3D scene.
`PointGlobalScaling`	Apply global scaling to a 3D scene.
`ShufflePoints`	Shuffle Points.
`ObjectRangeFilter`	Filter objects by point cloud range.
`LidarReformat`

6.1.5.1.2. API Reference¶

class hat.data.collates.CocktailCollate(ignore_id: int = - 1, batch_first: bool = True)¶

CocktailCollate.

鸡尾酒（多模）算法批量数据collate的Callable类. 默认需要处理的是 dict 类型数据的列表。

首先，将List[Dict[str, …]]转换成Dict[str, List] 然后，对dict中的 ‘images’, ‘audio’, ‘label’ 跟训练相关的数据。进行 pad_sequence 操作。对 ‘tokens’ 直接跳过。其他的key使用default_collate

参数

ignore_id – 被忽略的标签ID, 默认使用wenet中的IGNORE_ID即-1. 处理标签数据时，使用IGNORE_ID的值作为padding值
batch_first – 处理批量数据时, batch 的维度是否在第1位(数组编号0). 如果batch_first是True, 数组为 BxTx* 如果batch_first是False, 数组为 TxBx*

hat.data.collates.collate_2d(batch: List[Any]) → Union[torch.Tensor, Dict]¶

Merge a list of samples to form a mini-batch of Tensor(s).

Used in 2d task, for collating data with inconsistent shapes.

参数: batch (list) – list of data.

hat.data.collates.collate_2d_cat(batch: List[Any]) → Union[torch.Tensor, Dict]¶

Merge a list of samples to form a mini-batch of Tensor(s).

Used in 2d task, for collating data with the first dimension inconsistent. If the data shape is (n,c,h,w), concat on aixs 0 directly.

参数: batch (list) – list of data.

hat.data.collates.collate_2d_pad(batch: List[Any]) → Union[torch.Tensor, Dict]¶

Merge a list of samples to form a mini-batch of Tensor(s).

Used in 2d task, for collating data with inconsistent shapes. Images with different shapes will pad to max shapes by axis.

参数: batch (list) – list of data.

hat.data.collates.collate_2d_replace_empty(batch: List[Any], prob: float = 0.0) → Union[torch.Tensor, Dict]¶

Merge a list of samples to form a mini-batch of Tensor(s).

This function also replaces those detection samples that have no positive training targets with eligible ones.This can improve training effectiveness and efficiency when there are many images having no training targets in the dataset.

参数

batch – list of data.
prob – the probability of conducting empty-gt image replacement.

hat.data.collates.collate_2d_with_diff_im_hw(batch: List[Any]) → Union[torch.Tensor, Dict]¶

Merge a list of samples to form a mini-batch of Tensor(s).

Used in 2d task, for collating data with different image heights or widths. These inconsisten images will be vstacked in batch transform.

参数: batch (list) – list of data.

hat.data.collates.collate_3d(batch_data: List[Any])¶

Merge a list of samples to form a mini-batch of Tensor(s).

Used in bev task. * If output tensor from dataset shape is (n,c,h,w),concat on aixs 0 directly. * If output tensor from dataset shape is (c,h,w),expand_dim on axis 0 and concat.

参数: batch (list) – list of data.

hat.data.collates.collate_lidar(batch_list: List[Any]) → Union[torch.Tensor, Dict]¶

Merge a list of samples to form a mini-batch of Tensor(s).

Used in rad task, for collating data with inconsistent shapes. Rad(Realtime and Accurate 3D Object Detection).

First converts List[Dict[str, …] or List[Dict]] to Dict[str, List], then process values whoses keys are related to training.

参数: batch (list) – list of data.

hat.data.collates.collate_mot_seq(batch: List[Dict]) → Union[torch.Tensor, Dict]¶

Collate for mot seq data.

参数: batch (list) – list of data.

hat.data.collates.collate_nlu_with_pad(batch_dic: List[Dict], total_sequence_length: int = 30) → Dict¶: Collate nlu func for dataloader.

hat.data.collates.collate_psd(batch: List[Any])¶

Merge a list of samples to form a mini-batch of Tensor(s).

Used in parking slot detection(psd) task. For collating data with inconsistent shapes.

参数: batch – list of data.

hat.data.collates.collate_seq_with_diff_im_hw(batch: List[Dict]) → Union[torch.Tensor, Dict]¶

Merge a list of samples to form a mini-batch of Tensor(s).

Used in sequence task, for collating data with different image heights or widths. These inconsisten images will be vstacked in batch transform.

参数: batch (list) – list of data.

hat.data.collates.default_collate_v2(batch)¶

Entend torch.utils.data.default_collate.

It can handle classes that cannot be converted to torch.tensor and convert them to lists instead of reporting errors directly. Examples: batch=[dict(input_x=A), dict(input_x=B)] where input_x can not be converted to torch.Tensor output=dict(input_x=[A, B]).

class hat.data.dataloaders.PassThroughDataLoader(data: Any, *, length: int, clone: bool = False)¶

Directly pass through input example.

参数

data (Any) – Input data
length (int) – Length of dataloader
clone (bool, optional) – Whether clone input data

class hat.data.datasets.BatchTransformDataset(dataset: torch.utils.data.dataset.Dataset, transforms_cfgs: List, epoch_steps: List)¶

Dataset which uses different transforms in different epochs.

参数

dataset – Target dataset.
transforms_cfgs – The list of different transform configs.
epoch_steps – Effective epoch of different transforms.

class hat.data.datasets.Cityscapes(data_path: str, transforms: Optional[list] = None, pack_type: Optional[str] = None, pack_kwargs: Optional[dict] = None, color_space: str = 'bgr')¶

Cityscapes provides the method of reading cityscapes data from target pack type.

参数

data_path – The path of packed file.
pack_type – The pack type.
transfroms – Transfroms of cityscapes before using.
pack_kwargs – Kwargs for pack type.
color_space – color space of data.

class hat.data.datasets.CityscapesPacker(src_data_dir: str, target_data_dir: str, split_name: str, num_workers: int, pack_type: str, num_samples: Optional[int] = None, **kwargs)¶

CityscapesPacker is used for converting Cityscapes dataset in torchvision to target DataType format.

参数

src_data_dir (str) – The dir of original cityscapes data.
target_data_dir (str) – Path for packed file.
split_name (str) – Split name of data, such as train, val and so on.
num_workers (int) – Num workers for reading data using multiprocessing.
pack_type (str) – The file type for packing.
num_samples (int) – the number of samples you want to pack. You will pack all the samples if num_samples is None.

pack_data(idx)¶

Read orginal data from Folder with some process.

参数: idx (int) – Idx for reading.
返回: Processed data for pack.

class hat.data.datasets.Coco(data_path: str, transforms: Optional[List] = None, pack_type: Optional[str] = None, pack_kwargs: Optional[dict] = None)¶

Coco provides the method of reading coco data from target pack type.

参数

data_path (str) – The path of packed file.
transforms (list) – Transfroms of data before using.
pack_type (str) – The pack type.
pack_kwargs (dict) – Kwargs for pack type.

class hat.data.datasets.CocoDetection(root, annFile, num_classes=80, transform=None, target_transform=None, transforms=None)¶

Coco Detection Dataset.

参数

root (string) – Root directory where images are downloaded to.
annFile (string) – Path to json annotation file.
num_classes (int) – The number of classes of coco. 80 or 91.
transform (callable, optional) – A function transform that takes in an PIL image and returns a transformed version. E.g, transforms.ToTensor
target_transform (callable, optional) – A function transform that takes in the target and transforms it.
transforms (callable, optional) – A function transform that takes input sample and its target as entry and returns a transformed version.

class hat.data.datasets.CocoDetectionPacker(src_data_dir: str, target_data_dir: str, split_name: str, num_workers: int, pack_type: str, num_classes: int = 80, num_samples: Optional[int] = None, **kwargs)¶

CocoDetectionPacker is used for packing coco dataset to target format.

参数

src_data_dir (str) – The dir of original coco data.
target_data_dir (str) – Path for packed file.
split_name (str) – Split name of data, such as train, val and so on.
num_workers (int) – The num workers for reading data using multiprocessing.
pack_type (str) – The file type for packing.
num_classes (int) – The num of classes produced.
num_samples (int) – the number of samples you want to pack. You will pack all the samples if num_samples is None.

pack_data(idx)¶

Read orginal data from Folder with some process.

参数: idx (int) – Idx for reading.
返回: Processed data for pack.

class hat.data.datasets.CocoFromImage(*args, **kwargs)¶

Coco from image by torchvision.

The params of COCOFromImage is same as params of torchvision.dataset.CocoDetection.

class hat.data.datasets.ComposeDataset(datasets: List[Dict], batchsize_list: List[int])¶

Dataset wrapper for multiple datasets with precise batch size.

参数

datasets – config for each dataset.
batchsize_list – batchsize for each task dataset.

class hat.data.datasets.ConcatDataset(datasets, with_flag: bool = False, record_index: bool = False)¶

A wrapper of concatenated dataset with group flag.

Same as torch.utils.data.dataset.ConcatDataset, addititionally concatenat the group flag of all dataset.

参数

datasets – A list of datasets.
with_flag – Whether to concatenate datasets flags. If True, concatenate all datasets flag ( all datasets must has flag attribute in this case). Default to False.
record_index – Whether to record the index. If True, record the index. Default to False.

class hat.data.datasets.DistributedComposeRandomDataset(datasets: List[torch.utils.data.dataset.Dataset], sample_weights: List[int], shuffle=True, seed=0, multi_sample_output=False)¶

Dataset wrapper for multiple datasets fair sample weights accross multi workers in a distributed environment.

Each datsaet is cutted by (num_workers x num_ranks).

参数

datasets – list of datasets.
sample_weights – sample weights for each dataset.
shuffle – shuffle each dataset when set to True
seed – random seed for shuffle
multi_sample_output – whether dataset outputs multiple samples at the same time.

reinforce_type(expected_type)¶: Reinforce the type for DataPipe instance. And the ‘expected_type’ is required to be a subtype of the original type hint to restrict the type requirement of DataPipe instance.

class hat.data.datasets.ImageNet(data_path: str, out_pil: bool = False, transforms: Optional[List] = None, pack_type: Optional[str] = None, pack_kwargs: Optional[dict] = None)¶

ImageNet provides the method of reading imagenet data from target pack type.

参数

data_path (str) – The path of packed file.
transforms (list) – Transforms of voc before using.
pack_type (str) – The pack type.
pack_kwargs (dict) – Kwargs for pack type.

class hat.data.datasets.ImageNetFromImage(transforms=None, *args, **kwargs)¶

ImageNet from image by torchvison.

The params of ImageNetFromImage are same as params of torchvision.datasets.ImageNet.

class hat.data.datasets.ImageNetPacker(src_data_dir: str, target_data_dir: str, split_name: str, num_workers: int, pack_type: str, num_samples: Optional[int] = None, **kwargs)¶

ImageNetPacker is used for converting ImageNet dataset in torchvision to DataType format.

参数

src_data_dir (str) – The dir of original imagenet data.
target_data_dir (str) – Path for LMDB file.
split_name (str) – Split name of data, such as train, val and so on.
num_workers (int) – Num workers for reading data using multiprocessing.
pack_type (str) – The file type for packing.
num_samples (int) – the number of samples you want to pack. You will pack all the samples if num_samples is None.

pack_data(idx)¶

Read orginal data from Folder with some process.

参数: idx (int) – Idx for reading.
返回: Processed data for pack.

class hat.data.datasets.Kitti3D(data_path: str, num_point_feature: int = 4, transforms: Optional[List] = None, pack_type: Optional[str] = None, pack_kwargs: Optional[dict] = None)¶

Kitti3D provides the method of reading kitti3d data from target pack type.

参数

data_path – The path of LMDB file.
transforms – Transforms of voc before using.
pack_type – The pack type.
pack_kwargs – Kwargs for pack type.

class hat.data.datasets.Kitti3DDetection(source_path: str, split_name: str, transforms: Optional[Callable] = None, num_point_feature: int = 4)¶

Kitti 3D Detection Dataset.

参数

source_path – Root directory where images are downloaded to.
split_name – Dataset split, ‘train’ or ‘val’.
transforms – A function transform that takes input sample and its target as entry and returns a transformed version.
num_point_feature – Number of feature in points, default 4 (x, y, z, r).

class hat.data.datasets.Kitti3DDetectionPacker(src_data_dir: str, target_data_dir: str, split_name: str, num_workers: int, pack_type: str, num_samples: Optional[int] = None, **kwargs)¶

Kitti3DDetectionPacker is used for converting kitti3D dataset to target DataType format.

参数

src_data_dir – The dir of original kitti2D data.
target_data_dir – Path for LMDB file.
split_name – Dataset split, ‘train’ or ‘val’.
num_workers – The num workers for reading data using multiprocessing.
pack_type – The file type for packing.
num_samples – the number of samples you want to pack. You will pack all the samples if num_samples is None.

pack_data(idx)¶

Read orginal data from Folder with some process.

参数: idx (int) – Idx for reading.
返回: Processed data for pack.

class hat.data.datasets.Kitti3DReader(data_dir: str, split_name: str = 'train', num_point_feature: int = 4)¶

Kitti3D dataset processor.

参数

data_dir – Root directory path of Kitti3D dataset. And the directory structure of data_dir should be like this: ` |--- data_dir | |--- ImageSets | | |--- train.txt | | |--- val.txt | | |--- ... | |--- training | | |--- calib | | |--- image_2 | | |--- label_2 | | |--- velodyne | |--- testing | | |--- ... `
split – Dataset split, in [“train”, “val”, “test”].
num_point_feature – Number of feature in points, default 4 (x, y, z, r).

generate_reduced_pointcloud(points: numpy.ndarray, rect: numpy.ndarray, Trv2c: numpy.ndarray, P2: numpy.ndarray, image_shape: numpy.ndarray) → numpy.ndarray¶

Generate reduced pointcloud.

参数

points – Point cloud, shape=[N, 3] or shape=[N, 4].
rect – matrix rect, shape=[4, 4].
Trv2c – Translate matrix vel2cam, shape=[4, 4].
P2 – Project matrix, shape=[4, 4].
image_shape – Image shape, (H, W, …) format.

get_calib(index: int, extend_matrix: bool = True) → Dict¶

Get the calibration information of one sample.

参数

index – Int value in sample name. For example, the index value of sample ‘000026.bin’ will be int(26).
extend_matrix – Whether to pad calibration matrix from shape (3, 4) to (4,4).

返回

Calibration info.

返回类型

Dict

get_img(index: int) → Dict¶

Get the image information of one sample.

参数: index – Int value in sample name.
返回: Image info.
返回类型: Dict

get_label_annotation(index: int, add_difficulty: bool = True, add_num_points_in_gt: bool = True) → Dict¶

Get the annotation of one sample.

参数: index – Int value in sample name.
返回: annotaions.
返回类型: Dict

get_ponitcloud_from_bin(index: int, remove_outside: bool = False) → numpy.ndarray¶

Get the points cloud data of one sample.

参数: index – Int value in sample name.
返回: Points cloud data.
返回类型: np.ndarray

get_split_img_ids() → List[int]¶

Get all index of split dataset.

返回: All index of split dataset.
返回类型: List[int]

class hat.data.datasets.PascalVOC(data_path: str, transforms: Optional[List] = None, pack_type: Optional[str] = None, pack_kwargs: Optional[dict] = None)¶

PascalVOC provides the method of reading voc data from target pack type.

参数

data_path (str) – The path of packed file.
transforms (list) – Transforms of voc before using.
pack_type (str) – The pack type.
pack_kwargs (dict) – Kwargs for pack type.

class hat.data.datasets.RandDataset(length: int, example: Any, clone: bool = True, flag: int = 1)¶

class hat.data.datasets.RepeatDataset(dataset, times)¶

A wrapper of repeated dataset.

Using RepeatDataset can reduce the data loading time between epochs.

参数

dataset (torch.utils.data.Dataset) – The datasets for repeating.
times (int) – Repeat times.

class hat.data.datasets.ResampleDataset(dataset, with_flag: bool = False, resample_interval: int = 1)¶

A wrapper of resample dataset.

Using ResampleDataset can resample on original dataset: with specific interval.

参数

dataset (dict) – The datasets for resampling.
with_flag (bool) – Whether to use dataset.flag. If True, resampling dataset.flag with resample_interval ( dataset must has flag attribute in this case.)
resample_interval (int) – resample interval.

class hat.data.datasets.SimpleDataset(start: int, length: int, flag: int = 1)¶

class hat.data.datasets.VOCDetectionPacker(src_data_dir: str, target_data_dir: str, split_name: str, num_workers: int, pack_type: str, num_samples: Optional[int] = None, **kwargs)¶

VOCDetectionPacker is used for packing voc dataset to target format.

参数

src_data_dir (str) – Dir of original voc data.
target_data_dir (str) – Path for packed file.
split_name (str) – Split name of data, such as trainval and test.
num_workers (int) – Num workers for reading data using multiprocessing.
pack_type (str) – The file type for packing.
num_samples (int) – the number of samples you want to pack. You will pack all the samples if num_samples is None.

pack_data(idx)¶

Read orginal data from Folder with some process.

参数: idx (int) – Idx for reading.
返回: Processed data for pack.

class hat.data.datasets.VOCFromImage(size=416, *args, **kwargs)¶

VOC from image by torchvision.

The params of VOCFromImage is same as params of torchvision.dataset.VOCDetection.

class hat.data.samplers.DistSamplerHook(dataset, num_replicas: Optional[int] = None, rank: Optional[int] = None, shuffle: bool = True, seed: int = 0, drop_last: bool = False)¶

The hook api for torch.utils.data.DistributedDampler. Used to get local rank and num_replicas before create DistributedSampler.

参数

dataset – compose dataset
num_replicas – same as DistributedSampler
rank – Same as DistributedSampler
shuffle – if shuffle data
seed – random seed

class hat.data.samplers.DistributedCycleMultiDatasetSampler(dataset: hat.data.datasets.dataset_wrappers.ComposeDataset, batchsize_list: List[int], num_replicas: Optional[int] = None, rank: Optional[int] = None, shuffle: bool = True, seed: int = 0)¶

In one epoch period, do cyclic sampling on the dataset according to iter_time.

参数

dataset – compose dataset
num_replicas (int) – same as DistributedSampler
rank (int) – Same as DistributedSampler
shuffle – if shuffle data
seed – random seed

class hat.data.samplers.DistributedGroupSampler(dataset, samples_per_gpu: int = 1, num_replicas: Optional[int] = None, rank: Optional[int] = None, seed: int = 0)¶

Sampler that restricts data loading to a subset of the dataset.

Each batch data indices are sampled from one group in all of the groups. Groups are organized according to the dataset flags.

注解

Dataset is assumed to be constant size and must has flag attribute. Different number in flag array represent different groups. for example, in aspect ratio group flag, there are two groups, in which 0 represent h/w >= 1 and 1 represent h/w < 1 group. Dataset flag must is numpy array instance, the dtype must is np.uint8 and length at axis 0 must equal to the dataset length.

参数

dataset – Dataset used for sampling.
samples_per_gpu – Number samplers for each gpu. Default is 1.
num_replicas – Number of processes participating in distributed training.
rank – Rank of the current process within num_replicas.
seed – random seed used in torch.Generator(). This number should be identical across all processes in the distributed group. Default: 0.

set_epoch(epoch)¶

Sets the epoch for this sampler. When shuffle=True, this ensures all replicas use a different random ordering for each epoch. Otherwise, the next iteration of this sampler will yield the same ordering.

参数: epoch (int) – Epoch number.

class hat.data.samplers.SelectedSampler(indices_function: Callable, dataset: torch.utils.data.dataset.Dataset, *, num_replicas: Optional[int] = None, rank: Optional[int] = None, shuffle: bool = True, seed: int = 0, drop_last: bool = False)¶

Distributed sampler that supports user-defined indices.

参数

indices_function (Callable) – Callback function given by user. Input are dataset and return a indices list.
dataset – Dataset used for sampling.
num_replicas (int, optional) – Number of processes participating in distributed training. By default, world_size is retrieved from the current distributed group.
rank (int, optional) – Rank of the current process in num_replicas. By default, rank is retrieved from the current distributed group.
shuffle (bool, optional) – If True (default), sampler will shuffle the indices.
seed (int, optional) – random seed used to shuffle the sampler if shuffle=True. This number should be identical across all processes in the distributed group. Default: 0.
drop_last (bool, optional) – if True, then the sampler will drop the tail of the data to make it evenly divisible across the number of replicas. If False, the sampler will add extra indices to make the data evenly divisible across the replicas. Default: False.

警告

In distributed mode, calling the set_epoch() method at the beginning of each epoch before creating the DataLoader iterator is necessary to make shuffling work properly across multiple epochs. Otherwise, the same ordering will be always used.

set_epoch(epoch: int) → None¶

Sets the epoch for this sampler. When shuffle=True, this ensures all replicas use a different random ordering for each epoch. Otherwise, the next iteration of this sampler will yield the same ordering.

参数: epoch (int) – Epoch number.

class hat.data.transforms.AlbuImageOnlyTransform(albu_params: List[Dict])¶

AlbuImageOnlyTransform used on img only.

Composed by list of albu ImageOnlyTransform.

参数: albu_params – List of albu iamge only transform.

Examples:

dict(
    type="AlbuImageOnlyTransform",
    albu_params=[
        dict(
            name="RandomBrightnessContrast",
            p=0.3,
        ),
        dict(
            name="GaussNoise",
            var_limit=50.0,
            p=0.5,
        ),
        dict(
            name="Blur",
            p=0.2,
            blur_limit=(3, 15),
        ),
        dict(
            name="ToGray",
            p=0.2,
        ),
    ],
)

check_transform(transform)¶

Check transform is ImageOnlyTransform.

only support ImageOnlyTransform till now.

class hat.data.transforms.AugmentHSV(hgain=0.5, sgain=0.5, vgain=0.5, p=1.0)¶

Random add color disturbance.

Convert RGB img to HSV, and then randomly change the hue, saturation and value.

注解

Affected keys: ‘img’.

参数

hgain (float) – Gain of hue.
sgain (float) – Gain of saturation.
vgain (float) – Gain of value.
p (float) – Prob.

class hat.data.transforms.BgrToYuv444(affect_key='img', rgb_input=False)¶

BgrToYuv444 is used for color format convert.

注解

Affected keys: ‘img’.

参数: rgb_input (bool) – The input is rgb input or not.

class hat.data.transforms.BgrToYuv444V2(rgb_input: bool = False, swing: str = 'full')¶

BgrToYuv444V2 is used for color format convert.

BgrToYuv444V2 implements by calling rgb2centered_yuv functions which has been verified to get the basically same YUV output on J5.

注解

Affected keys: ‘img’.

参数

rgb_input – The input is rgb input or not.
swing – “studio” for YUV studio swing (Y: -112~107, U, V: -112~112). “full” for YUV full swing (Y, U, V: -128~127). default is “full”

class hat.data.transforms.BoxJitter(exp_ratio: float = 1.0, exp_jitter: float = 0.0, center_shift: float = 0.0)¶

Jitter box to simulate the box predicted by the model.

Usually used in tasks that use ground truth boxes for training.

参数

exp_ratio – Ratio of the expansion of box. Defaults to 1.0.
exp_jitter – Jitter of expansion ratio . Defaults to 0.0.
center_shift – Box center shift range. Defaults to 0.0.

class hat.data.transforms.Clip(minimum=0.0, maximum=255.0)¶

Clip Data to [minimum, maximum].

参数

minimum – The minimum number of data. Defaults 0.
maximum – The maximum number of data. Defaults 255.

class hat.data.transforms.ColorJitter(brightness=0.5, contrast=(0.5, 1.5), saturation=(0.5, 1.5), hue=0.1)¶

Randomly change the brightness, contrast, saturation and hue of an image.

For det and dict input are the main differences with ColorJitter in torchvision and the default settings have been changed to the most common settings.

注解

Affected keys: ‘img’.

参数

brightness (float or tuple of float (min, max)) – How much to jitter brightness.
contrast (float or tuple of float (min, max)) – How much to jitter contrast.
saturation (float or tuple of float (min, max)) – How much to jitter saturation.
hue (float or tuple of float (min, max)) – How much to jitter hue.

class hat.data.transforms.Contrast(p: float = 0.08, contrast: float = 0.5)¶

Randomly jitters image contrast with a factor.

注解

Affected keys: ‘img’.

参数

p – prob
contrast – How much to jitter contrast.
range (The contrast jitter ratio) –
[0 –
1] –

class hat.data.transforms.ConvertDataType(convert_map: Optional[Dict] = None)¶

Convert data type.

参数: convert_map – The mapping dict for to be converted data name and type. Only for np.ndarray and torch.Tensor.

class hat.data.transforms.ConvertLayout(hwc2chw=True, keys=None)¶

ConvertLayout is used for layout convert.

注解

Affected keys: ‘img’.

参数

hwc2chw (bool) – Whether to convert hwc to chw.
keys (list) –

class hat.data.transforms.DeleteKeys(keys: List[str])¶

Delete keys in input dict.

参数: keys – key list to detele

class hat.data.transforms.DetAffineAugTransformer(target_wh, flip_prob, scale_type='W', inter_method=10, use_pyramid=True, pyramid_min_step=0.7, pyramid_max_step=0.8, pixel_center_aligned=True, center_aligned=False, rand_scale_range=(1.0, 1.0), rand_translation_ratio=0.0, rand_aspect_ratio=0.0, rand_rotation_angle=0.0, norm_wh=None, norm_scale=None, resize_wh=None, min_valid_area=8, min_valid_clip_area_ratio=0.5, min_edge_size=2, clip_bbox=True, keep_aspect_ratio=False)¶

Affine augmentation for object detection.

参数

resize_wh – list/tuple of 2 int Resize input image to target size, by default None
**kwargs – Please see get_affine_image_resize() and ImageAffineTransform

class hat.data.transforms.FixedCrop(size=None, min_area=- 1, min_iou=- 1, dynamic_roi_params=None, discriminate_ignore_classes=False)¶

Crop image with fixed position and size.

注解

Affected keys: ‘img’, ‘img_shape’, ‘pad_shape’, ‘layout’, ‘before_crop_shape’, ‘crop_offset’, ‘gt_bboxes’, ‘gt_classes’.

inverse_transform(inputs, task_type, inverse_info)¶

Inverse option of transform to map the prediction to the original image.

参数

inputs (array) – Prediction
task_type (str) – detection or segmentation.
inverse_info (dict) – The transform keyword is the key, and the corresponding value is the value.

class hat.data.transforms.GaussianBlur(p: float = 0.08, kernel_size_min: int = 2, kernel_size_max: int = 9, sigma_min: float = 0.0, sigma_max: float = 0.0)¶

Randomly add guass blur on an image.

注解

Affected keys: ‘img’.

参数

p – prob
kernel_size_min – min size of guass kernel
kernel_size_max – max size of guass kernel
sigma_min – min sigma of guass kernel
sigma_max – max sigma of guass kernel

class hat.data.transforms.GazeRandomCropWoResize(size=(192, 320), area=(0.08, 1.0), ratio=(0.75, 1.3333333333333333), prob: float = 1.0, is_train: bool = True)¶

Random crop without resize.

More notes ref to https://horizonrobotics.feishu.cn/docx/LKhddopAeoXJmXxa6KocbwJdnSg. # noqa

class hat.data.transforms.GazeRotate3DWithCrop(is_train=True, head_pose_type='euler z-xy degree', rand_crop_scale=(0.85, 1.0), rand_crop_ratio=(1.25, 2), rand_crop_cropper_border=5, rotate_type='pos_map_uniform', rotate_augm_prob: float = 1, pos_map_range_pitch=(- 17, 17), pos_map_range_yaw=(- 20, 20), pos_map_range_roll=(- 20, 20), delta_rpy_range=([0, 0], [0, 0], [0, 0]), seperate_ldmk=False, seperate_ldmk_roll_range=(0, 0), crop_size=(256, 128), to_yuv420sp=True, standard_focal=600, cropping_ratio=0.25, rand_inter_type=False)¶

Random rotate image, calculate ROI and random crop if necessary.

Meanwhile, pos map is generated.

参数

is_train – To apply 3d rotate augm in train mod or test mod. Defaults to True.
head_pose_type – Type of head pose. Defaults to “euler z-xy degree”.
rand_crop_scale – Scale of rand crop. Defaults to (0.85, 1.0).
rand_crop_ratio – Ratio of rand crop. Defaults to (1.25, 2).
rand_crop_cropper_border – Expanded pixel size. Defaults to 5.
rotate_type – 3D rotate augm type. Defaults to “pos_map_uniform”.
rotate_augm_prob – Prob to do 3d rotate augm. Defaults to 1.
pos_map_range_pitch – Rotate range in pitch dimension.
pos_map_range_yaw – Rotate range in yaw dimension.
pos_map_range_roll – Rotate range in roll dimension.
delta_rpy_range – _description_.
seperate_ldmk – _description_. Defaults to False.
seperate_ldmk_roll_range – _description_. Defaults to (0, 0).
crop_size – Crop size. Defaults to (256, 128).
to_yuv420sp – Whether transform to yuv420sp. Defaults to True.
standard_focal – Standard focal of camera. Defaults to 600.
cropping_ratio – Ratio of crop when calc crop roi with rotated face ldmks.
rand_inter_type – Whether use rand inter type. Defaults to False.

class hat.data.transforms.GazeYUVTransform(rgb_data=False, nc=3)¶

YUVTransform for Gaze Task.

This pipeline: bgr_to_yuv444 -> equalizehist -> yuv444_to_yuv444_int8 :param rgb_data: whether input data is rgb format :param nc: output channels of data

Inputs:

data: input tensor with (H x W x C) shape.

Outputs:

out: output tensor with same shape as data.

class hat.data.transforms.HueSaturationValue(hue_range: Tuple[float, float] = (- 20, 20), sat_range: Tuple[float, float] = (- 30, 30), val_range: Tuple[float, float] = (- 20, 20), p: float = 0.5)¶

Randomly change hue, saturation and value of the input image.

Used for unit8 np.ndarray, RGB image input. Unlike AugmentHSV, this transform uses addition to shift value. This transform is same as albumentations.augmentations.transforms.HueSaturationValue

参数

hue_range – range for changing hue. Default: (-20, 20).
sat_range – range for changing saturation. Default: (-30, 30).
val_range – range for changing value. Default: (-20, 20).
p – probability of applying the transform. Default: 0.5.

class hat.data.transforms.IterableDetRoIListTransform(target_wh, flip_prob, img_scale_range=(0.5, 2.0), roi_scale_range=(0.8, 1.25), min_sample_num=1, max_sample_num=1, center_aligned=True, inter_method=10, use_pyramid=True, pyramid_min_step=0.7, pyramid_max_step=0.8, pixel_center_aligned=True, min_valid_area=8, min_valid_clip_area_ratio=0.5, min_edge_size=2, rand_translation_ratio=0, rand_aspect_ratio=0, rand_rotation_angle=0, reselect_ratio=0, clip_bbox=True, rand_sampling_bbox=True, resize_wh=None, keep_aspect_ratio=False, roi_list=None, append_gt=False)¶

Iterable transformer base on roi list for object detection.

参数

resize_wh (list/tuple of 2 int, optional) – Resize input image to target size, by default None
roi_list (ndarray, optional) – Transform the specified image region
append_gt (bool, optional) – Append the groundtruth to roi_list
**kwargs – Please see AffineMatFromROIBoxGenerator and ImageAffineTransform

class hat.data.transforms.IterableDetRoITransform(target_wh, flip_prob, img_scale_range=(0.5, 2.0), roi_scale_range=(0.8, 1.25), min_sample_num=1, max_sample_num=1, center_aligned=True, inter_method=10, use_pyramid=True, pyramid_min_step=0.7, pyramid_max_step=0.8, pixel_center_aligned=True, min_valid_area=8, min_valid_clip_area_ratio=0.5, min_edge_size=2, rand_translation_ratio=0, rand_aspect_ratio=0, rand_rotation_angle=0, reselect_ratio=0, clip_bbox=True, rand_sampling_bbox=True, resize_wh=None, keep_aspect_ratio=False)¶

Iterable transformer base on rois for object detection.

参数

resize_wh (list/tuple of 2 int, optional) – Resize input image to target size, by default None
**kwargs – Please see AffineMatFromROIBoxGenerator and ImageAffineTransform

class hat.data.transforms.JPEGCompress(p: float = 0.08, max_quality: int = 95, min_quality: int = 30)¶

Do JPEG compression to downgrade image quality.

注解

Affected keys: ‘img’.

参数

p – prob
max_quality – (0, 100] JPEG compression highest quality
min_quality – (0, 100] JPEG compression lowest quality

class hat.data.transforms.LabelRemap(mapping: Sequence)¶

Remap labels.

注解

Affected keys: ‘gt_seg’.

参数: mapping (Sequence) – Mapping from input to output.

class hat.data.transforms.LabelSmooth(num_classes, eta=0.1)¶

LabelSmooth is used for label smooth.

注解

Affected keys: ‘labels’.

参数

num_classes (int) – Num classes.
eta (float) – Eta of label smooth.

class hat.data.transforms.ListToDict(keys: List[str])¶

Convert list args to dict.

参数: keys – keys for each object in args.

class hat.data.transforms.MeanBlur(ksize: int = 3, p: float = 0.5)¶

Apply mean blur to the input image using a fix-sized kernel.

Used for np.ndarray.

参数

ksize – maximum kernel size for blurring the input image. Default: 3.
p – probability of applying the transform. Default: 0.5.

class hat.data.transforms.MedianBlur(ksize: int = 3, p: float = 0.5)¶

Apply median blur to the input image using a fix-sized kernel.

Used for np.ndarray.

参数

ksize – maximum kernel size for blurring the input image. Default: 3.
p – probability of applying the transform. Default: 0.5.

class hat.data.transforms.MinIoURandomCrop(min_ious=(0.1, 0.3, 0.5, 0.7, 0.9), min_crop_size=0.3, bbox_clip_border=True, repeat_num=50)¶

Random crop the image & bboxes, the cropped patches have minimum IoU requirement with original image & bboxes, the IoU threshold is randomly selected from min_ious.

注解

Affected keys: ‘img’, ‘gt_bboxes’, ‘gt_classes’, ‘gt_difficult’.

参数

min_ious (tuple) – minimum IoU threshold for all intersections with
boxes (bounding) –
min_crop_size (float) – minimum crop’s size (i.e. h,w := a*h, a*w,
min_crop_size) (where a >=) –
bbox_clip_border (bool) – Whether clip the objects outside the border of the image. Defaults to True.
repeat_num (float) – Max repeat num for finding avaiable bbox.

class hat.data.transforms.Mosaic(image_size: int = 512, degrees: int = 10, translate: float = 0.1, scale: float = 0.1, shear: int = 10, perspective: float = 0.0, mixup: bool = True)¶

Mosaic augmentation for detection task.

参数

image_size – Image size after mosaic pipeline. Default: (512, 512).
degrees – Rotation degree. Defaults to 10.
translate – translate value for warpPerspective. Defaults to 0.1.
scale – Random scale value. Defaults to 0.1.
shear – Shear value for warpPerspective. Defaults to 10.
perspective – perspective value for warpPerspective. Defaults to 0.0.
mixup – Whether use mixup. Defaults to True.

class hat.data.transforms.MotionBlur(p: float = 0.08, length_min: int = 9, length_max: int = 18, angle_min: float = 1, angle_max: float = 359)¶

Randomly add motion blur on an image.

注解

Affected keys: ‘img’.

参数

p – prob
length_min – min size of motion blur
length_max – max size of motion blur
angle_min – min angle of motion blur
angle_max – max angle of motion blur

class hat.data.transforms.MultiTaskAnnoWrapper(sub_transforms: Dict[str, Any], unikeys: Tuple[str] = (), repkeys: Tuple[str] = ())¶

Wrapper for multi-task anno generating.

参数

sub_transforms – The mapping dict for task-wise transforms.
unikeys – Keys of unique annotations in each task.
repkeys – Keys of repeated annotations for all tasks.

class hat.data.transforms.Normalize(mean: Union[float, Sequence[float]], std: Union[float, Sequence[float]], raw_norm=False)¶

Normalize image.

注解

Affected keys: ‘img’, ‘layout’.

参数

mean – mean of normalize.
std – std of normalize.
raw_norm (bool) – Whether to open raw_norm.

class hat.data.transforms.OneHot(num_classes)¶

OneHot is used for convert layer to one-hot format.

注解

Affected keys: ‘labels’.

参数: num_classes (int) – Num classes.

class hat.data.transforms.PILToTensor¶: Convert PIL Image to Tensor.

class hat.data.transforms.PadTensorListToBatch(pad_val: int = 0, seg_pad_val: Optional[int] = 255)¶

List of image tensor to be stacked vertically.

Used for diff shape tensors list.

参数

pad_val – Values to be filled in padding areas for img. Default to 0.
seg_pad_val – Value to be filled in padding areas for gt_seg. Default to 255.

class hat.data.transforms.PlainCopyPaste(min_ins_num: int = 1, cp_prob: float = 0.0)¶

Copy and paste instances plainly.

参数

min_ins_num – Min instances num of the image after paste.
cp_prob – Probability of applying this transformation.

class hat.data.transforms.PresetCrop(crop_top: int = 220, crop_bottom: int = 128, crop_left: int = 0, crop_right: int = 0, min_area: float = - 1, min_iou: float = - 1)¶

Crop image with preset roi param.

inverse_transform(inputs, task_type, inverse_info)¶

Inverse option of transform to map the prediction to the original image.

参数

inputs (array) – Prediction
task_type (str) – detection or segmentation.
inverse_info (dict) – not used yet.

class hat.data.transforms.RGBShift(r_shift_limit: Tuple[float, float] = (- 20, 20), g_shift_limit: Tuple[float, float] = (- 20, 20), b_shift_limit: Tuple[float, float] = (- 20, 20), p: float = 0.5)¶

Randomly shift values for each channel of the input image.

Used for np.ndarray. This transform is same as albumentations.augmentations.transforms.RGBShift.

参数

r_shift_limit – range for changing values for the red channel. Default: (-20, 20).
g_shift_limit – range for changing values for the green channel. Default: (-20, 20).
b_shift_limit – range for changing values for the blue channel. Default: (-20, 20).
p – probability of applying the transform. Default: 0.5.

class hat.data.transforms.RandomBrightnessContrast(brightness_limit: Tuple[float, float] = (- 0.2, 0.2), contrast_limit: Tuple[float, float] = (- 0.2, 0.2), brightness_by_max: bool = True, p=0.5)¶

Randomly change brightness and contrast of the input image.

Used for unit8 np.ndarray. This transform is same as albumentations.augmentations.transforms.RandomBrightnessContrast.

参数

brightness_limit – factor range for changing brightness. Default: (-0.2, 0.2).
contrast_limit – factor range for changing contrast. Default: (-0.2, 0.2).
brightness_by_max – If True adjust contrast by image dtype maximum, else adjust contrast by image mean.
p – probability of applying the transform. Default: 0.5.

class hat.data.transforms.RandomColorJitter(brightness=0.5, contrast=(0.5, 1.5), saturation=(0.5, 1.5), hue=0.1, prob=0.5)¶

Randomly change the brightness, contrast, saturation and hue of an image. # noqa

More notes ref to https://horizonrobotics.feishu.cn/docx/LKhddopAeoXJmXxa6KocbwJdnSg. # noqa

class hat.data.transforms.RandomDownSample(p: float = 0.2, data_shape: Optional[Tuple] = (3, 112, 112), min_downsample_width: int = 60, inter_method: int = 1)¶

First downsample and upsample to original size.

注解

Affected keys: ‘img’.

参数

p – prob
data_shape – C, H, W
min_downsample_width – minimum downsample width
inter_method – interpolation method index

class hat.data.transforms.RandomExpand(mean=(0, 0, 0), ratio_range=(1, 4), prob=0.5)¶

Random expand the image & bboxes.

Randomly place the original image on a canvas of ‘ratio’ x original image size filled with mean values. The ratio is in the range of ratio_range.

注解

Affected keys: ‘img’, ‘gt_bboxes’.

参数

ratio_range (tuple) – range of expand ratio.
prob (float) – probability of applying this transformation

class hat.data.transforms.RandomFlip(px: Optional[float] = 0.5, py: Optional[float] = 0)¶

Flip image & bbox & mask & seg & flow.

注解

Affected keys: ‘img’, ‘ori_img’, ‘img_shape’, ‘pad_shape’, ‘gt_bboxes’, ‘gt_seg’, ‘gt_flow’, ‘gt_mask’, ‘gt_ldmk’, ‘ldmk_pairs’.

参数

px – Horizontal flip probability, range between [0, 1].
py – Vertical flip probability, range between [0, 1].

class hat.data.transforms.RandomGray(p: float = 0.08, rgb_data: bool = True)¶

Transform RGB or BGR format into Gray format.

注解

Affected keys: ‘img’.

参数

p – prob
rgb_data – Default=True Whether the input data is in RGB format. If not, it should be in BGR format.

class hat.data.transforms.RandomResizedCrop(height: int, width: int, scale: Tuple[float, float] = (0.08, 1.0), ratio: Tuple[float, float] = (0.75, 1.3333333333333333), interpolation: int = 1, p: float = 1.0)¶

Torchvision’s variant of crop a random part of the input, and rescale it to some size.

Used for np.ndarray. This transform is same as albumentations.augmentations.transforms.RandomResizedCrop.

参数

height – height after crop and resize.
width – width after crop and resize.
scale – range of size of the origin size cropped
ratio – range of aspect ratio of the origin aspect ratio cropped.
interpolation – flag that is used to specify the interpolation algorithm. Should be one of: cv2.INTER_NEAREST, cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_LANCZOS4. Default: cv2.INTER_LINEAR.
p – probability of applying the transform. Default: 1.

class hat.data.transforms.RandomSelectOne(transforms: List, p: float = 0.5, p_trans: Optional[List] = None)¶

Select one of transforms to apply.

参数

transforms – list of transformations to compose.
p – probability of applying selected transform. Default: 0.5.
p_trans – list of possibility of transformations.

class hat.data.transforms.RenameKeys(keys: List[str], split='|')¶

Rename keys in input dict.

参数: keys – key list to rename, in “old_name | new_name” format.

class hat.data.transforms.Resize(img_scale: Optional[Union[Sequence[int], Sequence[Sequence[int]]]] = None, max_scale: Optional[Union[Sequence[int], Sequence[Sequence[int]]]] = None, multiscale_mode: str = 'range', ratio_range: Optional[Tuple[float, float]] = None, keep_ratio: bool = True, pad_to_keep_ratio: bool = False, raw_scaler_enable: bool = False, sample1c_enable: bool = False, divisor: int = 1, rm_neg_coords: bool = True)¶

Resize image & bbox & mask & seg.

注解

Affected keys: ‘img’, ‘ori_img’, ‘img_shape’, ‘pad_shape’, ‘resized_shape’, ‘pad_shape’, ‘scale_factor’, ‘gt_bboxes’, ‘gt_seg’, ‘gt_ldmk’.

参数

img_scale – See above.
max_scale – The max size of image. If the image’s shape > max_scale, The image is resized to max_scale
multiscale_mode – Value must be one of “range” or “value”. This transform resizes the input image and bbox to same scale factor. There are 3 multiscale modes: ‘ratio_range’ is not None: randomly sample a ratio from the ratio range and multiply with the image scale. e.g. Resize(img_scale=(400, 500)), multiscale_mode=’range’, ratio_range=(0.5, 2.0) ‘ratio_range’ is None and ‘multiscale_mode’ == “range”: randomly sample a scale from a range, the length of img_scale[tuple] must be 2, which represent small img_scale and large img_scale. e.g. Resize(img_scale=((100, 200), (400,500)), multiscale_mode=’range’) ‘ratio_range’ is None and ‘multiscale_mode’ == “value”: randomly sample a scale from multiple scales. e.g. Resize(img_scale=((100, 200), (300, 400), (400, 500)), multiscale_mode=’value’)))
ratio_range – Scale factor range like (min_ratio, max_ratio).
keep_ratio – Whether to keep the aspect ratio when resizing the image.
pad_to_keep_ratio – Whether to pad image to keep the same shape and aspect ratio when resizing the image to target shape.
raw_scaler_enable – Whether to enable raw scaler when resize the image.
sample1c_enable – Whether to sample one channel after resize the image.
divisor – Width and height are rounded to multiples of divisor.
rm_neg_coords – Whether to rm negative coordinates.

inverse_transform(inputs, task_type, inverse_info)¶

Inverse option of transform to map the prediction to the original image.

参数

inputs (array|Tensor) – Prediction.
task_type (str) – detection or segmentation.
inverse_info (dict) – The transform keyword is the key, and the corresponding value is the value.

class hat.data.transforms.Resize3D(img_scale=None, multiscale_mode='range', ratio_range=None, keep_ratio=True, bbox_clip_border=True, backend='cv2', interpolation='nearest', override=False, cam2img_keep_ratio=False)¶

Resize 3D labels.

Different from 2D Resize, we accept img_scale=None and ratio_range is not None. In that case we will take the input img scale as the ori_scale for rescaling with ratio_range.

参数

img_scale – Images scales for resizing.
multiscale_mode – Either “range” or “value”.
ratio_range – (min_ratio, max_ratio).
keep_ratio – Whether to keep the aspect ratio when resizing the image.
bbox_clip_border – Whether to clip the objects outside the border of the image.
backend (str) – Image resize backend, choices are ‘cv2’ and ‘pillow’.
interpolation (str) – Interpolation method, accepted values are “nearest”, “bilinear”, “bicubic”, “area”, “lanczos” for ‘cv2’ backend, “nearest”, “bilinear” for ‘pillow’ backend.
override (bool, optional) – Whether to override scale and scale_factor so as to call resize twice.

class hat.data.transforms.Scale(scales: Union[numbers.Real, Sequence], mode: str = 'nearest', mul_scale: bool = False)¶

Scale input according to a scale list.

注解

Affected keys: ‘img’, ‘gt_flow’, ‘gt_ori_flow’, ‘gt_seg’.

参数

scales (Union[Real, Sequence]) – The scales to apply on input.
mode (str) – algorithm used for upsampling: 'nearest' | 'bilinear' | 'area'. Default: 'nearest'
mul_scale (bool) – Whether to multiply the scale coefficient.

class hat.data.transforms.SegOneHot(num_classes: int)¶

OneHot is used for convert layer to one-hot format.

注解

Affected keys: ‘gt_seg’.

参数: num_classes (int) – Num classes.

class hat.data.transforms.SegRandomAffine(degrees: Union[Sequence, float] = 0, translate: Tuple = None, scale: Tuple = None, shear: Union[Sequence, float] = None, interpolation: torchvision.transforms.functional.InterpolationMode = InterpolationMode.NEAREST, fill: Union[tuple, int] = 0, label_fill_value: Union[tuple, int] = - 1, rotate_p: float = 1.0, translate_p: float = 1.0, scale_p: float = 1.0)¶

Apply random for both image and label.

Please refer to RandomAffine for details.

注解

Affected keys: ‘img’, ‘gt_flow’, ‘gt_seg’.

参数

label_fill_value (tuple or int, optional) – Fill value for label. Defaults to -1.
translate_p – Translate flip probability, range between [0, 1].
scale_p – Scale flip probability, range between [0, 1].

class hat.data.transforms.SegRandomCrop(size, cat_max_ratio=1.0, ignore_index=255)¶

Random crop on data with gt_seg label, can only be used for segmentation: task.

注解

Affected keys: ‘img’, ‘img_shape’, ‘pad_shape’, ‘layout’, ‘gt_seg’.

参数

size (tuple) – Expected size after cropping, (h, w).
cat_max_ratio (float, optional) – The maximum ratio that single category could occupy.
ignore_index (int, optional) – When considering the cat_max_ratio condition, the area corresponding to ignore_index will be ignored.

get_crop_bbox(data)¶: Randomly get a crop bounding box.

class hat.data.transforms.SegRandomCutOut(prob: float, n_holes: Union[int, Tuple[int, int]], cutout_shape: Optional[Union[Tuple[int, int], Tuple[Tuple[int, int], ...]]] = None, cutout_ratio: Optional[Union[Tuple[int, int], Tuple[Tuple[int, int], ...]]] = None, fill_in: Tuple[float, float, float] = (0, 0, 0), seg_fill_in: Optional[int] = None)¶

CutOut operation for segmentation task.

Randomly drop some regions of image used in Cutout.

参数

prob – Cutout probability.
n_holes – Number of regions to be dropped. If it is given as a list,
interval (number of holes will be randomly selected from the closed) – [n_holes[0], n_holes[1]].
cutout_shape – The candidate shape of dropped regions. It can be tuple[int, int] to use a fixed cutout shape, or list[tuple[int, int]] to randomly choose shape from the list.
cutout_ratio – The candidate ratio of dropped regions. It can be tuple[float, float] to use a fixed ratio or list[tuple[float, float]] to randomly choose ratio from the list. Please note that cutout_shape and cutout_ratio cannot be both given at the same time.
fill_in – The value of pixel to fill in the dropped regions. Default is (0, 0, 0).
seg_fill_in – The labels of pixel to fill in the dropped regions. If seg_fill_in is None, skip. Default is None.

class hat.data.transforms.SegReWeightByArea(seg_num_classes, lower_bound: int = 0.5, ignore_index: int = 255)¶

Calculate the weight of each category according to the area of each category.

For each category, the calculation formula of weight is as follows: weight = max(1.0 - seg_area / total_area, lower_bound)

注解

Affected keys: ‘gt_seg’, ‘gt_seg_weight’.

参数

seg_num_classes (int) – Number of segmentation categories.
lower_bound (float) – Lower bound of weight.
ignore_index (int) – Index of ignore class.

class hat.data.transforms.SegResize(size, interpolation=InterpolationMode.BILINEAR)¶

Apply resize for both image and label.

注解

Affected keys: ‘img’, ‘gt_seg’.

参数

size – target size of resize.
interpolation – interpolation method of resize.

forward(data)¶

参数: img (PIL Image or Tensor) – Image to be scaled.
返回: Rescaled image.
返回类型: PIL Image or Tensor

class hat.data.transforms.SegResizeAffine(img_scale: Optional[Union[Sequence[int], Sequence[Sequence[int]]]] = None, max_scale: Optional[Union[Sequence[int], Sequence[Sequence[int]]]] = None, multiscale_mode: str = 'range', ratio_range: Optional[Tuple[float, float]] = None, keep_ratio: bool = True)¶

Resize image & seg.

注解

Affected keys: ‘img’, ‘img_shape’, ‘pad_shape’, ‘resized_shape’, ‘scale_factor’, ‘gt_seg’, ‘gt_polygons’.

参数

img_scale – (height, width) or a list of [(height1, width1), (height2, width2), …] for image resize.
max_scale – The max size of image. If the image’s shape > max_scale, The image is resized to max_scale
multiscale_mode – Value must be one of “range” or “value”. This transform resizes the input image and bbox to same scale factor. There are 3 multiscale modes: ‘ratio_range’ is not None: randomly sample a ratio from the ratio range and multiply with the image scale. e.g. Resize(img_scale=(400, 500)), multiscale_mode=’range’, ratio_range=(0.5, 2.0) ‘ratio_range’ is None and ‘multiscale_mode’ == “range”: randomly sample a scale from a range, the length of img_scale[tuple] must be 2, which represent small img_scale and large img_scale. e.g. Resize(img_scale=((100, 200), (400,500)), multiscale_mode=’range’) ‘ratio_range’ is None and ‘multiscale_mode’ == “value”: randomly sample a scale from multiple scales. e.g. Resize(img_scale=((100, 200), (300, 400), (400, 500)), multiscale_mode=’value’)))
ratio_range – Scale factor range like (min_ratio, max_ratio).
keep_ratio – Whether to keep the aspect ratio when resizing the image.

inverse_transform(inputs: numpy.ndarray, task_type: str, inverse_info: Dict[str, Any])¶

Inverse option of transform to map the prediction to the original image.

参数

inputs – Prediction.
task_type – support segmentation only.
inverse_info – The transform keyword is the key, and the corresponding value is the value.

class hat.data.transforms.SeqAlbuImageOnlyTransform(albu_params: List[Dict])¶

class hat.data.transforms.SeqAugmentHSV(hgain=0.5, sgain=0.5, vgain=0.5, p=1.0)¶: Random add color disturbance for sequence.

class hat.data.transforms.SeqBgrToYuv444(affect_key='img', rgb_input=False)¶: BgrToYuv444 for sequence.

class hat.data.transforms.SeqNormalize(mean: Union[float, Sequence[float]], std: Union[float, Sequence[float]], raw_norm=False)¶: Normalize for sequence.

class hat.data.transforms.SeqRandomFlip(px: Optional[float] = 0.5, py: Optional[float] = 0)¶: Flip image & bbox & mask & seg & flow for sequence.

class hat.data.transforms.SeqRandomSizeCrop(min_size: int, max_size: int, **kwargs)¶: RandomSizeCrop for sequence.

class hat.data.transforms.SeqResize(img_scale: Optional[Union[Sequence[int], Sequence[Sequence[int]]]] = None, max_scale: Optional[Union[Sequence[int], Sequence[Sequence[int]]]] = None, multiscale_mode: str = 'range', ratio_range: Optional[Tuple[float, float]] = None, keep_ratio: bool = True, pad_to_keep_ratio: bool = False, raw_scaler_enable: bool = False, sample1c_enable: bool = False, divisor: int = 1, rm_neg_coords: bool = True)¶

class hat.data.transforms.SeqToFasterRCNNData(max_gt_boxes_num=500, max_ig_regions_num=500)¶

class hat.data.transforms.SeqToTensor(to_yuv: bool = False, use_yuv_v2: bool = True)¶: ToTensor for sequence.

class hat.data.transforms.ShiftScaleRotate(shift_limit: Tuple[float, float] = (- 0.0625, 0.0625), scale_limit: Tuple[float, float] = (- 0.1, 0.1), rotate_limit: Tuple[float, float] = (- 45.0, 45.0), interpolation: int = 1, border_mode: int = 4, value: Optional[int] = None, p: float = 0.5)¶

Randomly apply affine transforms: translate, scale and rotate the input.

Used for np.ndarray hwc img. This transform is same as albumentations.augmentations.transforms.ShiftScaleRotate.

参数

shift_limit – shift factor range for both height and width. Absolute values for lower and upper bounds should lie in range [0, 1]. Default: (-0.0625, 0.0625).
scale_limit – scaling factor range. Default: (-0.1, 0.1).
rotate_limit – rotation range. Default: (-45, 45).
interpolation – flag that is used to specify the interpolation algorithm. Should be one of: cv2.INTER_NEAREST, cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_LANCZOS4. Default: cv2.INTER_LINEAR.
border_mode – flag that is used to specify the pixel extrapolation method. Should be one of: cv2.BORDER_CONSTANT, cv2.BORDER_REPLICATE, cv2.BORDER_REFLECT, cv2.BORDER_WRAP, cv2.BORDER_REFLECT_101. Default: cv2.BORDER_REFLECT_101
value – padding value if border_mode is cv2.BORDER_CONSTANT.
p – probability of applying the transform. Default: 0.5.

class hat.data.transforms.SpatialVariantBrightness(p: float = 0.08, brightness: float = 0.6, max_template_type: int = 3, online_template: bool = False)¶

Spatial variant brightness, Enhanced Edition. Powered by xin.wang@horizon.ai.

注解

Affected keys: ‘img’.

参数

p – prob
brightness – default is 0.6 Brightness ratio for this augmentation, the value choice in Uniform ~ [-brightness, brigheness].
max_template_type – default is 3 Max number of template type in once process. Note, the selection process is repeated.
online_template – default is False Template generated online or offline. “False” is recommended to get fast speed.

class hat.data.transforms.TensorToNumpy¶: Convert tensor to numpy.

class hat.data.transforms.TimmMixup(*args, **kwargs)¶

Mixup of timm.

注解

Affected keys: ‘img’, ‘labels’.

参数: timm.data.Mixup (args are the same as) –

class hat.data.transforms.TimmTransforms(*args, **kwargs)¶

Transforms of timm.

注解

Affected keys: ‘img’.

参数: timm.data.create_transform (args are the same as) –

class hat.data.transforms.ToFasterRCNNData(max_gt_boxes_num=500, max_ig_regions_num=500)¶

Prepare faster-rcnn input data.

Convert gt_bboxes (n, 4) & gt_classes (n, ) to gt_boxes (n, 5), gt_boxes_num (1, ), ig_regions (m, 5), ig_regions_num (m, ); If gt_ids exists, it will be concated into gt_boxes, resulting in gt_boxes array shape expanding from nx5 to nx6.

Convert key img_shape to im_hw; Convert image Layout to chw;

参数

max_gt_boxes_num (int) – Max gt bboxes number in one image, Default 500.
max_ig_regions_num (int) – Max ignore regions number in one image, Default 500.

返回

Result dict with: gt_boxes (max_gt_boxes_num, 5 or 6), gt_boxes_num (1, ), ig_regions (max_ig_regions_num, 5 or 6), ig_regions_num (1, ), im_hw (2,) layout convert to “chw”.

返回类型

dict

class hat.data.transforms.ToLdmkRCNNData(num_ldmk=15, max_gt_boxes_num=1000, max_ig_regions_num=1000)¶

Transform dataset to RCNN input need.

This class is used to stack landmark with boxes, and typically used to facilitate landmark and boxes matching in anchor-based model.

参数

num_ldmk – Number of landmark. Defaults to 15.
max_gt_boxes_num – Max gt bboxes number in one image. Defaults to 1000.
max_ig_regions_num – Max ignore regions number in one image. Defaults to 1000.

class hat.data.transforms.ToMultiTaskFasterRCNNData(taskname_clsidx_map: Dict[str, int], max_gt_boxes_num: int = 500, max_ig_regions_num: int = 500, num_ldmk: int = 15)¶

Convert multi-classes detection data to multi-task data.

Each class will be convert to a detection task.

参数

taskname_clsidx_map – {cls1: cls_idx1, cls2: cls_idx2}.
max_gt_boxes_num – Same as ToFasterRCNNData. Defaults to 500.
max_ig_regions_num – Same as ToFasterRCNNData. Defaults to 500.
num_ldmk – Number of human ldmk. Defaults to 15.

返回

Result dict with: ”task1”: FasterRCNNDataDict1, “task2”: FasterRCNNDataDict2,

返回类型

dict

class hat.data.transforms.ToTensor(to_yuv: bool = False, use_yuv_v2: bool = True)¶

Convert objects of various python types to torch.Tensor and convert the img to yuv444 format if to_yuv is True.

Supported types are: numpy.ndarray, torch.Tensor, Sequence, int, float.

注解

Affected keys: ‘img’, ‘img_shape’, ‘pad_shape’, ‘layout’, ‘gt_bboxes’, ‘gt_seg’, ‘gt_seg_weights’, ‘gt_flow’, ‘color_space’.

参数

to_yuv – If true, convert the img to yuv444 format.
use_yuv_v2 – If true, use BgrToYuv444V2 when convert img to yuv format.

class hat.data.transforms.Undistortion¶

Convert a PIL Image or numpy.ndarray to: undistor PIL Image or numpy.ndarray.

hat.data.transforms.eye_ldmk_mirror(eye_ldmk, normd=True)¶

Flip eye landmarks.

Eye landmarks(21 points) here are already computed as ratio within final input image.

class hat.data.transforms.lidar_utils.DBFilterByDifficulty(filter_by_difficulty)¶

Filter sampled data by diffculties.

参数: removed_difficulties (list) – class diffculties

class hat.data.transforms.lidar_utils.DBFilterByMinNumPoint(filter_by_min_num_points)¶

Filter sampled data by NumPoint.

参数: min_gt_point_dict (dict) – class numpoint thershold

class hat.data.transforms.lidar_utils.ObjectNoise(gt_rotation_noise: List[float], gt_loc_noise_std: List[float], global_random_rot_range: List[float], num_try: int = 100)¶

Apply noise to each GT objects in the scene.

参数

gt_rotation_noise – Object rotation range.
gt_loc_noise_std – Object noise std.
global_random_rot_range – Global rotation to the scene.
num_try – Number of times to try if the noise applied is invalid.

class hat.data.transforms.lidar_utils.ObjectRangeFilter(point_cloud_range: List[float])¶

Filter objects by point cloud range.

参数: point_cloud_range – Point cloud range.

class hat.data.transforms.lidar_utils.ObjectSample(db_sampler: Callable, class_names: List[str], random_crop: bool = False, remove_points_after_sample: bool = False, remove_outside_points: bool = False)¶

Sample GT objects to the data.

参数

db_sampler – Database sampler.
class_names – Class names.
random_crop – Whether to random crop.
remove_points_after_sample – Whether to remove points after sample.
remove_outside_points – Whether to remove outsize points.

class hat.data.transforms.lidar_utils.PointGlobalRotation(rotation: float = 0.7853981633974483)¶

Apply global rotation to a 3D scene.

参数: rotation – Range of rotation angle.

class hat.data.transforms.lidar_utils.PointGlobalScaling(min_scale: float = 0.95, max_scale: float = 1.05)¶

Apply global scaling to a 3D scene.

参数

min_scale – Min scale ratio.
max_scale – Max scale ratio.

class hat.data.transforms.lidar_utils.PointRandomFlip(probability: float = 0.5)¶

Flip the points & bbox.

参数: probability – The flipping probability.

class hat.data.transforms.lidar_utils.ShufflePoints(shuffle: bool = True)¶

Shuffle Points.

参数: shuffle – Whether to shuffle