batchgeneratorsv2

0.3.2 verified Sat May 09 auth: no python

Batchgenerators is a deep learning data augmentation and batching library originally for medical image segmentation. Version 2 overhauls the API for composable transformations, deterministic ordering, and reduced overhead. Current version 0.3.2 (pre-1.0, breaking changes likely). Released irregularly.

pip install batchgeneratorsv2

Common errors

error AttributeError: module 'batchgeneratorsv2' has no attribute 'BatchGenerator' ↓

cause Trying to import BatchGenerator from the top module instead of submodule.

fix

Use: from batchgeneratorsv2.generators import BatchGenerator

error ImportError: cannot import name 'Compose' from 'batchgeneratorsv2' ↓

cause Compose is not in the top-level package; it's in transforms.

fix

Use: from batchgeneratorsv2.transforms import Compose

error RuntimeError: DataLoader worker (pid(s) ...) exited unexpectedly ↓

cause Using num_processes > 1 with multiprocessing has known issues.

fix

Set num_processes=1 or use single-process mode.

Warnings

breaking batchgeneratorsv2 is a complete rewrite from batchgenerators (v1). Import paths, API, and behavior differ: v2 uses generators module, Compose is in transforms, and data is passed as dict with 'data' key. ↓

fix Do not mix v1 and v2 imports. Rewrite code to use v2 API: from batchgeneratorsv2.generators import BatchGenerator; from batchgeneratorsv2.transforms import Compose.

breaking v0.3.0 removed the old 'batchgeneratorsv2.transforms.abstract' submodule. All transforms now inherit from base classes in transforms.base. ↓

fix Use transforms.base.AbstractTransform instead of transforms.abstract.AbstractTransform.

deprecated Setting num_processes > 1 is experimental. May cause deadlocks or data corruption. ↓

fix Use num_processes=1 for stability, or test carefully with multiprocessing.

gotcha SpatialTransform's 'patch_size' must be smaller than input dimensions; otherwise, it silently fails or returns empty arrays. ↓

fix Ensure patch_size is less than or equal to the spatial dimensions of your input data.

Imports

BatchGenerator

wrong

from batchgeneratorsv2 import BatchGenerator

correct

from batchgeneratorsv2.generators import BatchGenerator

BatchGenerator is in the generators submodule, not top-level.

Compose

wrong

from batchgeneratorsv2 import Compose

correct

from batchgeneratorsv2.transforms import Compose

Compose moved to transforms submodule.

Quickstart

Minimal example: create a BatchGenerator with transforms and iterate.

import numpy as np
from batchgeneratorsv2.generators import BatchGenerator
from batchgeneratorsv2.transforms import Compose, SpatialTransform, ColorTransform

# Dummy data
data = np.random.rand(100, 3, 128, 128)  # N, C, H, W

# Define transforms
transforms = Compose([
    SpatialTransform(patch_size=(64, 64), do_rotation=True),
    ColorTransform(brightness_range=(0.9, 1.1), contrast_range=(0.9, 1.1))
])

# Create batch generator
batch_gen = BatchGenerator(
    data=data,
    batch_size=4,
    transforms=transforms,
    num_processes=1
)

# Iterate
for batch in batch_gen:
    x = batch['data']
    # x shape: (4, 3, 64, 64)
    break
print('Quickstart works: shape', x.shape)