batchgenerators

0.25.1 verified Mon Apr 27 auth: no python

Data augmentation toolkit for medical imaging and deep learning. Current version 0.25.1. Released on PyPI with irregular cadence; maintained by MIC-DKFZ.

pip install batchgenerators

Common errors

error AttributeError: module 'batchgenerators' has no attribute 'dataloading' ↓

cause Installed version is too old (pre-0.20) where dataloading was not a top-level module.

fix

Upgrade batchgenerators: pip install --upgrade batchgenerators

error ImportError: cannot import name 'DefaultTransformations' from 'batchgenerators.transforms' ↓

cause DefaultTransformations was removed in 0.25. Use explicit transforms instead.

fix

Use from batchgenerators.transforms import Compose, SpatialTransform, ... and compose your own pipeline.

error RuntimeError: DataLoader expects keyword argument 'data' but got 'dataset' ↓

cause Code written for older batchgenerators (<0.25) uses 'dataset', newer versions require 'data'.

fix

Change DataLoader(dataset=...) to DataLoader(data=...)

Warnings

breaking In version 0.25, DataLoader no longer accepts the 'dataset' argument; use 'data' instead. ↓

fix Replace DataLoader(dataset=...) with DataLoader(data=...)

deprecated batchgenerators.transforms.abstract_transforms is deprecated; use batchgenerators.transforms directly. ↓

fix Import from batchgenerators.transforms instead.

gotcha Memory usage can spike if num_threads is set too high in DataLoader, especially with large images. ↓

fix Keep num_threads <= number of CPU cores; use multiprocessing context if needed.

Imports

DataLoader

wrong

from batchgenerators.dataloading.data_loader import DataLoader

correct

from batchgenerators.dataloading import DataLoader

DataLoader is directly in batchgenerators.dataloading; submodule path changed.

DefaultTransformations

from batchgenerators.transforms import DefaultTransformations

Common transformation set.

Quickstart

Basic usage: create a DataLoader from a list of data, apply augmentations via Compose.

from batchgenerators.dataloading import DataLoader
from batchgenerators.transforms import Compose, SpatialTransform

# Simple data loader for a list of images
train_loader = DataLoader(my_data_list, batch_size=4, num_threads=2)
transforms = Compose([SpatialTransform(patch_size=(128,128,128))])
for batch in train_loader:
    data = batch['data']
    seg = batch['seg']
    data, seg = transforms(data, seg)
    break