Data Transformation

In the OpenMMLab algorithm library, dataset construction and data preparation are decoupled. Usually, the construction of the dataset only parses the dataset and records the basic information of each sample, while the data preparation is a series of data transformations including data loading, preprocessing, formatting, and other operations performed according to the basic information of the sample.

Design of data transformation

In MMCV, we use various callable data transformation classes to manipulate data. These data transformation classes can accept several configuration parameters for the instantiation and then process the input data dictionary by __call__ method. All data transformation methods accept a dictionary as the input and produce the output as a dictionary as well. A simple example is as follows:

>>> import numpy as np
>>> from mmcv.transforms import Resize
>>> transform = Resize(scale=(224, 224))
>>> data_dict = {'img': np.random.rand(256, 256, 3)}
>>> data_dict = transform(data_dict)
>>> print(data_dict['img'].shape)
(224, 224, 3)

The data transformation class reads some fields of the input dictionary and may add or update some fields. The keys of these fields are mostly fixed. For example, Resize will always read fields such as "img" in the input dictionary. More information about the conventions for input and output fields could be found in the documentation of the corresponding class.


By convention, the order of image shape which is used as initialization parameters in data transformation (such as Resize, Pad) is (width, height). In the dictionary returned by the data transformation, the image related shape, such as img_shape, ori_shape, pad_shape, etc., is (height, width).

MMCV provides a unified base class called BaseTransform for all data transformation classes:

class BaseTransform(metaclass=ABCMeta):

    def __call__(self, results: dict) -> dict:

        return self.transform(results)

    def transform(self, results: dict) -> dict:

All data transformation classes must inherit BaseTransform and implement the transform method. Both the input and output of the transform method are a dictionary. In the Custom data transformation class section, we will describe how to implement a data transformation class in more detail.

Data pipeline

As mentioned above, the inputs and outputs of all data transformations are dictionaries. Moreover, according to the [Convention on Datasets] (TODO) in OpenMMLab, the basic information of each sample in the dataset is also a dictionary. This way, we can connect all data transformation operations end to end and combine them into a data pipeline. This pipeline inputs the information dictionary of the samples in the dataset and outputs the information dictionary after a series of processing.

Taking the classification task as an example, we show a typical data pipeline in the figure below. For each sample, the information stored in the dataset is a dictionary, as shown on the far left in the figure. After each data transformation operation represented by the blue block, a new field (marked in green) will be added to the data dictionary or an existing field (marked in orange) will be updated.

The data pipeline is a list of several data transformation configuration dictionaries in the configuration file. Each dataset needs to set the parameter pipeline to define the data preparation operations the dataset needs to perform. The configuration of the above data pipeline in the configuration file is as follows:

pipeline = [
    dict(type='Resize', size=256, keep_ratio=True),
    dict(type='CenterCrop', crop_size=224),
    dict(type='Normalize', mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375]),

dataset = dict(

Common data transformation classes

The commonly used data transformation classes can be roughly divided into data loading, data preprocessing and augmentation, and data formatting. In MMCV, we provide some commonly used classes as follows:

Data loading

To support the loading of large-scale datasets, data is usually not loaded when Dataset is initialized. Only the corresponding path is loaded. Therefore, it is necessary to load specific data in the data pipeline.

Class Feature
LoadImageFromFile Load from file path
LoadAnnotations Load and organize the annotations (bbox, etc.)

Data preprocessing and enhancement

Data preprocessing and augmentation usually involve transforming the image itself, such as cropping, padding, scaling, etc.

Class Feature
Pad Padding
CenterCrop Center crop
Normalize Image normalization
Resize Resize to the specified size or ratio
RandomResize Scale the image randomly within the specified range
RandomMultiscaleResize Scale the image to a random size from multiple options
RandomGrayscale Random grayscale
RandomFlip Random flip
MultiScaleFlipAug Support scaling and flipping during the testing

Data formatting

Data formatting operations are type conversions performed on the data.

Class Feature
ToTensor Convert the specified data to torch.Tensor
ImageToTensor Convert the image to torch.Tensor

Customize data transformation classes

To implement a new data transformation class, you must inherit BaseTransform and implement the transform method. Here, we use a simple flip transform (MyFlip) as an example:

import random
import mmcv
from mmcv.transforms import BaseTransform, TRANSFORMS

class MyFlip(BaseTransform):
    def __init__(self, direction: str):
        self.direction = direction

    def transform(self, results: dict) -> dict:
        img = results['img']
        results['img'] = mmcv.imflip(img, direction=self.direction)
        return results

Now, we can instantiate MyFlip as a callable object to handle our data dictionary.

import numpy as np

transform = MyFlip(direction='horizontal')
data_dict = {'img': np.random.rand(224, 224, 3)}
data_dict = transform(data_dict)
processed_img = data_dict['img']

Alternatively, use MyFlip transform in the pipeline of the config file.

pipeline = [
    dict(type='MyFlip', direction='horizontal'),

It should be noted that if you want to use it in the configuration file, you must ensure that the file where the MyFlip class is located can be imported at the runtime.

Transform wrapper

Transform wrappers are a special class of data transformations. They do not operate on images, labels or other information in the data dictionary by themselves. Instead, they enhance the behavior of data transformations defined in them.


KeyMapper is used to map fields in the data dictionary. For example, image processing transforms usually get their values from the "img" field in the data dictionary. But sometimes we want these transforms to handle images in other fields in the data dictionary, such as the "gt_img" field.

When used with registry and configuration file, the field map wrapper should be used as follows:

pipeline = [
            'img': 'gt_img',  # map "gt_img" to "img"
            'mask': ...,  # The "mask" field in the raw data is not used. That is, for wrapped data transformations, the "mask" field is not included in the data
        auto_remap=True,  # remap "img" back to "gt_img" after the transformation
            # only need to specify "img" in `RandomFlip`

With KeyMapper, we don’t need to consider various possible input field names in the transform method when we implement the data transformation class. We only need to deal with the default fields.

RandomChoice and RandomApply

RandomChoice is used to randomly select a data transformation pipeline from the given choices. With this wrapper, we can easily implement some data augmentation functions, such as AutoAugment.

In configuration file, you can use RandomChoice as follows:

pipeline = [
                dict(type='Posterize', bits=4),
                dict(type='Rotate', angle=30.)
            ],  # the first combo option
                dict(type='Rotate', angle=30)
            ],  # the second combo option
        prob=[0.4, 0.6]  # the prob of each combo

RandomApply is used to randomly perform a combination of data transformations with a specified probability. For example:

pipeline = [
        transforms=[dict(type='Rotate', angle=30.)],
        prob=0.3)  # perform the transformation with prob as 0.3


Usually, a data transformation class only reads the target of an operation from one field. While we can also use KeyMapper to change the fields read, there is no way to apply transformations to the data of multiple fields at once. To achieve this, we need to use the multi-target extension wrapper TransformBroadcaster.

TransformBroadcaster has two uses, one is to apply data transformation to multiple specified fields, and the other is to apply data transformation to a group of targets under a field.

  1. Apply to multiple fields

    Suppose we need to apply a data transformation to images in two fields "lq" (low-quality) and "gt" (ground-truth).

    pipeline = [
            # apply to the "lq" and "gt" fields respectively, and set the "img" field to both
            mapping={'img': ['lq', 'gt']},
            # remap the "img" field back to the original field after the transformation
            # whether to share random variables in the transformation of each target
            # more introduction will be referred in the following chapters (random variable sharing)
                # only need to manipulate the "img" field in the `RandomFlip` class

    In the mapping setting of the multi-target extension, we can also use ... to ignore the specified original field. As shown in the following example, the wrapped RandomCrop will crop the image in the field "img" and update the size of the cropped image if the field "img_shape" exists. If we want to do the same random cropping for both image fields "lq" and "gt" at the same time but update the "img_shape" field only once, we can do it as in the example:

    pipeline = [
                'img': ['lq', 'gt'],
                'img_shape': ['img_shape', ...],
            # remap the "img" and "img_shape" fields back to their original fields after the transformation
            # whether to share random variables in the transformation of each target
            # more introduction will be referred in the following chapters (random variable sharing)
                # "img" and "img_shape" fields are manipulated in the `RandomCrop` class
                # if "img_shape" is missing, only operate on "img"
  2. A set of targets applied to a field

    Suppose we need to apply a data transformation to the "images" field, which is a list of images.

    pipeline = [
            # map each image under the "images" field to the "img" field
            mapping={'img': 'images'},
            # remap the images under the "img" field back to the list in the "images" field after the transformation
            # whether to share random variables in the transformation of each target
                # in the `RandomFlip` transformation class, we only need to manipulate the "img" field

Decorator cache_randomness

In TransformBroadcaster, we provide the share_random_params option to support sharing random states across multiple data transformations. For example, in a super-resolution task, we want to apply the same random transformations simultaneously to the low-resolution image and the original image. If we use this function in a custom data transformation class, we need to mark which random variables support sharing in the class. This can be achieved with the decorator cache_randomness.

Taking MyFlip from the above example, we want to perform flipping randomly with a certain probability:

from mmcv.transforms.utils import cache_randomness

class MyRandomFlip(BaseTransform):
    def __init__(self, prob: float, direction: str):
        self.prob = prob
        self.direction = direction

    @cache_randomness  # label the output of the method as a shareable random variable
    def do_flip(self):
        flip = True if random.random() > self.prob else False
        return flip

    def transform(self, results: dict) -> dict:
        img = results['img']
        if self.do_flip():
            results['img'] = mmcv.imflip(img, direction=self.direction)
        return results

In the above example, we decorate the do_flip method with cache_randomness, marking the method return value flip as a random variable that supports sharing. Therefore, in the transformation of TransformBroadcaster to multiple targets, the value of this variable will remain the same.

Decorator avoid_cache_randomness

In some cases, we cannot separate the process of generating random variables in data transformation into a class method. For example, modules from third-party libraries used in data transformation encapsulate the relevant parts of random variables inside, making them impossible to be extracted as class methods for data transformation. Such data transformations cannot support shared random variables through the decorator cache_randomness annotation, and thus cannot share random variables during multi-objective expansion.

To avoid misuse of such data transformations in multi-object extensions, we provide another decorator, avoid_cache_randomness, to mark such data transformations:

from mmcv.transforms.utils import avoid_cache_randomness

class MyRandomTransform(BaseTransform):

    def transform(self, results: dict) -> dict:

Data transformation classes marked with avoid_cache_randomness will throw an exception when their instance is wrapped by TransformBroadcaster and the parameter share_random_params is set to True. This reminds the user not to use it in this way.

There are a few things to keep in mind when using avoid_cache_randomness:

  1. avoid_cache_randomness is only used to decorate data transformation classes (subclasses of BaseTransfrom) and cannot be used to decorate other general classes, class methods, or functions

  2. When a data transformation decorated with avoid_cache_randomness is used as a base class, its subclasses will not inherit its feature. If the subclass is still unable to share random variables, avoid_cache_randomness should be used again.

  3. A data transformation needs to be modified with avoid_cache_randomness only when a data transformation is random and cannot share its random parameters. Data transformations without randomness require no decoration

Read the Docs v: 2.x
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.