mmagic.structures

Package Contents

Classes

DataSample

A data structure interface of MMagic. They are used as interfaces

class mmagic.structures.DataSample(*, metainfo: Optional[dict] = None, **kwargs)

Bases: mmengine.structures.BaseDataElement

A data structure interface of MMagic. They are used as interfaces between different components, e.g., model, visualizer, evaluator, etc. Typically, DataSample contains all the information and data from ground- truth and predictions.

DataSample inherits from BaseDataElement. See more details in:

https://mmengine.readthedocs.io/en/latest/advanced_tutorials/data_element.html Specifically, an instance of BaseDataElement consists of two components, - metainfo, which contains some meta information,

e.g., img_shape, img_id, color_order, etc.

• data, which contains the data used in the loop.

The attributes in DataSample are divided into several parts:

• gt_img: Ground truth image(s).

• pred_img: Image(s) of model predictions.

• ref_img: Reference image(s).

• mask: Mask in Inpainting.

• trimap: Trimap in Matting.

• gt_alpha: Ground truth alpha image in Matting.

• pred_alpha: Predicted alpha image in Matting.

• gt_fg: Ground truth foreground image in Matting.

• pred_fg: Predicted foreground image in Matting.

• gt_bg: Ground truth background image in Matting.

• pred_bg: Predicted background image in Matting.

• gt_merged: Ground truth merged image in Matting.

Examples:

 >>> import torch
 >>> import numpy as np
 >>> from mmagic.structures import DataSample
 >>> img_meta = dict(img_shape=(800, 1196, 3))
 >>> img = torch.rand((3, 800, 1196))
 >>> data_sample = DataSample(gt_img=img, metainfo=img_meta)
 >>> assert 'img_shape' in data_sample.metainfo_keys()
 >>> data_sample
<DataSample(

    META INFORMATION
    img_shape: (800, 1196, 3)

    DATA FIELDS
    gt_img: tensor(...)
) at 0x1f6a5a99a00>

We also support stack and split operation to handle a batch of data samples:

>>> import torch
>>> import numpy as np
>>> from mmagic.structures import DataSample
>>> img_meta1 = img_meta2 = dict(img_shape=(800, 1196, 3))
>>> img1 = torch.rand((3, 800, 1196))
>>> img2 = torch.rand((3, 800, 1196))
>>> data_sample1 = DataSample(gt_img=img1, metainfo=img_meta1)
>>> data_sample2 = DataSample(gt_img=img2, metainfo=img_meta1)

>>> # stack them and then use as batched-tensor!
>>> data_sample = DataSample.stack([data_sample1, data_sample2])
>>> print(data_sample.gt_img.shape)
torch.Size([2, 3, 800, 1196])
>>> print(data_sample.metainfo)
{'img_shape': [(800, 1196, 3), (800, 1196, 3)]}

>>> # split them if you want
>>> data_sample1_, data_sample2_ = data_sample.split()
>>> assert (data_sample1_.gt_img == img1).all()
>>> assert (data_sample2_.gt_img == img2).all()

property gt_label

This the function to fetch gt label.

Returns

gt label.

Return type

LabelData

META_KEYS

DATA_KEYS

set_predefined_data(data: dict) → None

set or change pre-defined key-value pairs in data_field by parameter data.

Parameters

data (dict) – A dict contains annotations of image or model predictions.

set_tensor_data(data: dict) → None

convert input data to tensor, and then set or change key-value pairs in data_field by parameter data.

Parameters

data (dict) – A dict contains annotations of image or model predictions.

set_gt_label(value: Union[numpy.ndarray, torch.Tensor, Sequence[numbers.Number], numbers.Number]) → DataSample

Set label of gt_label.

gt_label()

Delete gt label.

classmethod stack(data_samples: Sequence[DataSample]) → DataSample

Stack a list of data samples to one. All tensor fields will be stacked at first dimension. Otherwise the values will be saved in a list.

Parameters

data_samples (Sequence['DataSample']) – A sequence of DataSample to stack.

Returns

The stacked data sample.

Return type

DataSample

split(allow_nonseq_value: bool = False) → Sequence[DataSample]

Split a sequence of data sample in the first dimension.

Parameters

allow_nonseq_value (bool) – Whether allow non-sequential data in split operation. If True, non-sequential data will be copied for all split data samples. Otherwise, an error will be raised. Defaults to False.

Returns

The list of data samples after splitting.

Return type

Sequence[DataSample]

__len__()

Get the length of the data sample.