Add stochastic mirror augmentation
Training image classifiers
Keras · Class
layers.RandomFlip
RandomFlip randomly mirrors images horizontally, vertically, or in both directions during training. It preserves image dimensions and returns the original input during inference.
How it works
RandomFlip is an image-preprocessing layer that applies randomized left-right, top-bottom, or combined flips when training is enabled. It accepts either one image or a batch, preserves the image dimensions, and leaves inputs unchanged during inference.
Syntax
keras.layers.RandomFlip(mode="horizontal_and_vertical", seed=None, data_format=None, **kwargs)Parameters
| Name | Type | Description |
|---|---|---|
mode | string | Flip direction: horizontal, vertical, or horizontal_and_vertical. Defaults to horizontal_and_vertical. |
seed | integer | Integer used to create a random seed. |
**kwargs | keyword arguments | Base-layer keyword arguments, such as name and dtype. |
Returns
A 3D unbatched or 4D batched tensor with the same (..., height, width, channels) shape convention as the input.
Key points
- Use RandomFlip as an in-model augmentation layer rather than implementing flip logic in the input loop.
- The selected mode determines whether the transformation mirrors width, height, or both axes.
- The mode argument accepts exactly three documented string choices.
- Provide an integer seed when you want to control the layer's random-number initialization.
- The layer handles either a single image tensor or a batch of images.
Common mistakes
- Augmentation is disabled during inference; pass training=True when explicitly calling the layer if you need a flip.
- Expect floating-point output unless you configure an appropriate dtype.
- The layer does not require normalized pixels, so scaling to [0, 1] is optional.
Example: Apply random horizontal flips
import keras
augment = keras.layers.RandomFlip(
mode="horizontal",
seed=42,
)
images = keras.random.uniform((8, 128, 128, 3))
flipped = augment(images, training=True)
unchanged = augment(images, training=False)
print(flipped.shape)
print(unchanged.shape)Creates a horizontal-flip augmenter, applies it during training, and shows that inference preserves the input shape and values.
When to use it
Augment images within a data pipeline
tf.data or grain input pipelines
Increase variation without changing annotations
Tasks with mirror-invariant labels
Interview questions
QWhat does RandomFlip do during training and inference?
It is a preprocessing layer for randomly flipping images during training. The mode selects horizontal, vertical, or both directions, while inference returns the input unchanged.
Follow-up: What happens if the layer is called directly outside a training context? Pass training=True to request augmentation; otherwise it follows inference behavior.
QWhich flip modes are supported?
The mode can be horizontal, vertical, or horizontal_and_vertical. Horizontal means left-to-right mirroring, and vertical means top-to-bottom mirroring.
Follow-up: How would you restrict augmentation to one direction? Construct the layer with mode="horizontal" or mode="vertical".
QWhat input and output shapes does the layer use?
It accepts a 3D unbatched or 4D batched image tensor in channels_last layout, with shape (..., height, width, channels). The output has the same rank and shape convention.
Follow-up: Can integer images be supplied? Yes; pixel values may be integer or floating point and may use ranges such as [0, 255] or [0, 1).
Related
keras.layers.RandomRotationkeras.layers.RandomCropkeras.layers.RandomTranslationkeras.layers.Rescaling