Albumentations Cheat Sheet

1Install & Compose2.x

pip install albumentations # or: pip install albumentationsx★
Same import either way. AlbumentationsX is the actively-developed, dual-licensed successor.
import albumentations as A from albumentations.pytorch import ToTensorV2★
Convention is the A alias. ToTensorV2 lives in the pytorch submodule.
tf = A.Compose([ A.RandomResizedCrop(size=(224,224)), A.HorizontalFlip(p=0.5), A.Normalize(), ToTensorV2(), ])★
A pipeline is an ordered list. A typical train transform: geometry → color → Normalize → ToTensorV2.

2Apply & read outputkeyword targets

out = tf(image=img) # img: numpy HWC, RGB, uint8★
Input is a NumPy array (e.g. from cv2 or np.array(PIL)) — not a PIL Image. Convert BGR→RGB if you used cv2.imread.
aug_img = out["image"]★
The result is a dict keyed by the targets you passed. After ToTensorV2 the image is a CHW tensor.
out = tf(image=img, mask=mask, bboxes=bb, keypoints=kp)
Pass every target that should transform together — all stay geometrically consistent.
# each call re-samples randomness -> a different augmentation
Calling tf again gives a new random augmentation; that's the point for training.

3Spatial & cropsmove geometry

A.HorizontalFlip(p=0.5) · A.VerticalFlip(p=0.5) · A.RandomRotate90()★
Flips/rotations — cheap, high-value. These also update masks, bboxes & keypoints.
A.Affine(scale=(0.8,1.2), translate_percent=0.1, rotate=(-15,15), p=0.7)★
General affine (the modern replacement for ShiftScaleRotate). A.Perspective() for perspective warps.
A.Resize(256,256) · A.RandomResizedCrop(size=(224,224)) · A.CenterCrop(224,224)★
Sizing. RandomResizedCrop is the standard train-time crop; Resize+CenterCrop for eval.
A.ElasticTransform() · A.GridDistortion() · A.CoarseDropout()
Elastic/distortion warps (great for medical/OCR); CoarseDropout (cutout) erases patches for regularization.

4Pixel-levelappearance only

A.RandomBrightnessContrast(p=0.5) · A.HueSaturationValue() · A.ColorJitter()★
Color/lighting jitter — robustness to camera & exposure variation. Applied to the image only (masks/boxes untouched).
A.GaussNoise() · A.GaussianBlur() · A.MotionBlur() · A.ImageCompression()
Noise/blur/JPEG-artifact simulation — mimic real-world degradation.
A.CLAHE() · A.Sharpen() · A.ToGray() · A.RandomGamma()
Contrast/detail/tone transforms.
A.Normalize(mean=(0.485,0.456,0.406), std=(0.229,0.224,0.225))★
Scales & normalizes (ImageNet stats by default) — put it just before ToTensorV2.

6Masks (segmentation)easiest target

out = tf(image=img, mask=mask); out["mask"]★
Just pass mask= — no extra config. Spatial transforms move it in lockstep with the image.
tf(image=img, masks=[m1, m2, ...]) # multiple masks
masks= (plural) for instance masks / multi-class stacks.
use nearest interpolation for masksnote
Resizes/warps use nearest-neighbor for masks automatically so label values aren't blended — but don't Normalize/blur a mask.

7Bounding boxesdetection

tf = A.Compose([...], bbox_params=A.BboxParams(format="pascal_voc", label_fields=["class_labels"]))★
You must declare bbox_params. format: pascal_voc [x_min,y_min,x_max,y_max], coco [x,y,w,h], yolo (normalized), albumentations.
out = tf(image=img, bboxes=bboxes, class_labels=labels) out["bboxes"], out["class_labels"]★
Pass labels via the declared label_fields so they stay aligned when boxes are dropped.
A.BboxParams(..., min_area=64, min_visibility=0.3)★
Drop boxes that a crop shrinks below a pixel area or visibility fraction — avoids degenerate labels.

8Keypointspose / landmarks

tf = A.Compose([...], keypoint_params=A.KeypointParams(format="xy", label_fields=["kp_labels"]))★
format: xy, yx, xya (angle), xys (scale), xyas. remove_invisible=True drops keypoints leaving the frame.
out = tf(image=img, keypoints=kps, kp_labels=labels)
Coordinates are transformed with the image; combine with bbox_params for full detection+pose pipelines.
tf(image=img, mask=m, bboxes=b, keypoints=k, class_labels=c)
All target types can be augmented in a single consistent call.

9PyTorch & serializationdatasets & replay

class DS(Dataset): def __getitem__(self, i): out = self.tf(image=np.array(img), mask=mask) return out["image"], out["mask"]★
Call the transform inside __getitem__; with ToTensorV2 the image is already a CHW tensor.
A.save(tf, "aug.json") · tf = A.load("aug.json")★
Serialize a pipeline to JSON/YAML — version-control your augmentation exactly.
tf = A.ReplayCompose([...]); out = tf(image=img) A.ReplayCompose.replay(out["replay"], image=img2)
Capture the exact random parameters and replay them on another image (e.g. paired/stereo data).
A.Compose([...], additional_targets={"image2": "image"})
Apply the same augmentation to extra targets (a second image, extra masks).

!Common gotchasread before shipping

NumPy in, not PIL
Pass np.array(pil_img). And if you loaded with cv2.imread, convert BGR→RGB first or colors are wrong.
ToTensorV2 must be last
Once it's a tensor, later NumPy transforms break. Order: augment → Normalize → ToTensorV2.
ToTensorV2 does NOT scale to [0,1]
Unlike torchvision's ToTensor. Use A.Normalize (which scales) before it, or your inputs stay 0–255.
declare bbox_params / keypoint_params
Passing bboxes= without BboxParams errors. And match the format to your annotation style.
don't blur/normalize masks
Keep pixel-level color transforms away from masks; only spatial transforms should touch them.

albumentations cheat sheet