PackInputs¶

class mmpretrain.datasets.transforms.PackInputs(input_key='img', algorithm_keys=(), meta_keys=('sample_idx', 'img_path', 'ori_shape', 'img_shape', 'scale_factor', 'flip', 'flip_direction'))[source]¶

Pack the inputs data.

Required Keys:

input_key
*algorithm_keys
*meta_keys

Deleted Keys:

All other keys in the dict.

Added Keys:

inputs (torch.Tensor): The forward data of models.
data_samples (DataSample): The annotation info of the sample.

Parameters:

input_key (str) – The key of element to feed into the model forwarding. Defaults to ‘img’.
algorithm_keys (Sequence[str]) – The keys of custom elements to be used in the algorithm. Defaults to an empty tuple.
meta_keys (Sequence[str]) – The keys of meta information to be saved in the data sample. Defaults to PackInputs.DEFAULT_META_KEYS.

Default algorithm keys

Besides the specified algorithm_keys, we will set some default keys into the output data sample and do some formatting. Therefore, you don’t need to set these keys in the algorithm_keys.

gt_label: The ground-truth label. The value will be converted into a 1-D tensor.
gt_score: The ground-truth score. The value will be converted into a 1-D tensor.
mask: The mask for some self-supervise tasks. The value will be converted into a tensor.

Default meta keys

sample_idx: The id of the image sample.
img_path: The path to the image file.
ori_shape: The original shape of the image as a tuple (H, W).
img_shape: The shape of the image after the pipeline as a tuple (H, W).
scale_factor: The scale factor between the resized image and the original image.
flip: A boolean indicating if image flip transform was used.
flip_direction: The flipping direction.

transform(results)[source]¶: Method to pack the input data.