MMCV: OpenMMLab Computer Vision Foundation

2.2.0 · active · verified Thu Apr 16

MMCV (OpenMMLab Computer Vision Foundation) is a foundational library for computer vision research and development, providing a comprehensive set of building blocks including rich data augmentation and transformation operations, various deep learning utilities, and support for hardware-specific optimizations. It is a core component of the OpenMMLab ecosystem. The current version is 2.2.0, with frequent releases to add features, bug fixes, and support for new PyTorch versions and hardware platforms.

Common errors

```
ModuleNotFoundError: No module named 'mmcv.runner'
```
cause Attempting to import training-related modules (like Runner) from MMCV in versions >= 2.0.0, where they have been moved to the MMEngine library.

fix Change the import statement to `from mmengine.runner import Runner`. Ensure `mmengine` is installed.
```
RuntimeError: CUDA error: no kernel image is available for execution on the device
```
cause This typically means `mmcv-full` was installed incorrectly, or the installed `mmcv-full` version's CUDA extensions do not match your current PyTorch or NVIDIA driver/CUDA toolkit environment.

fix Uninstall any existing `mmcv` and `mmcv-full`. Install `openmim` (`pip install -U openmim`) and then use `mim install mmcv-full` after carefully checking PyTorch and CUDA compatibility in the official documentation.
```
ImportError: cannot import name 'build_from_cfg' from 'mmcv.cnn'
```
cause The utility function `build_from_cfg` was moved to a different module path in MMCV v2.x.

fix Update the import statement to `from mmcv.registry import build_from_cfg`.

Warnings

breaking MMCV v2.0.0 introduced significant breaking changes by moving training-related modules (like `Runner`, `Hook`, `Optimizer`) to a new library, `MMEngine`. Direct imports from `mmcv.runner` or similar paths will fail.
Fix: Update imports to use `mmengine` for these modules (e.g., `from mmengine.runner import Runner`). Ensure `mmengine` is installed alongside `mmcv`.
gotcha Installing `mmcv-full` (the version with CUDA/hardware extensions) via `pip install mmcv-full` can be problematic and may result in a CPU-only build or version mismatches. The `mim` tool is highly recommended for managing `mmcv-full` installations.
Fix: First, install `openmim` (`pip install -U openmim`), then use `mim install mmcv-full`. Refer to the official installation guide for specific PyTorch and CUDA version compatibility.
gotcha MMCV-full has strict compatibility requirements with PyTorch and CUDA versions. Mismatched versions can lead to runtime errors (e.g., CUDA kernel errors) or unexpected behavior.
Fix: Always check the official MMCV installation guide for the exact PyTorch and CUDA versions supported by your chosen `mmcv-full` version. Use `mim install mmcv-full` to leverage OpenMMLab's pre-built packages which handle compatibility.
breaking The paths for data transformation pipelines and utilities (e.g., `Compose`) changed in v2.x. Old imports like `from mmcv.datasets.pipelines import Compose` are no longer correct.
Fix: Update import paths to `from mmcv.transforms import Compose` and similarly for other transformation components.

Install

pip install -U openmim mim install mmcv-full Recommended (with CUDA/hardware extensions)
pip install mmcv Basic version (without CUDA/hardware extensions)

Imports

Config
```
from mmcv import Config
```
While mmengine.Config is also commonly used in MMEngine-based projects, mmcv.Config remains for general configuration parsing.
Runner
wrong:
```
from mmcv.runner import Runner
```
correct:
```
from mmengine.runner import Runner
```
As of v2.0.0, training-related modules like Runner, Hook, and Parallel moved from MMCV to MMEngine.
build_from_cfg
wrong:
```
from mmcv.cnn import build_from_cfg
```
correct:
```
from mmcv.registry import build_from_cfg
```
The import path for build_from_cfg changed in v2.x to mmcv.registry.
Compose
wrong:
```
from mmcv.datasets.pipelines import Compose
```
correct:
```
from mmcv.transforms import Compose
```
Data augmentation and transformation pipelines moved to mmcv.transforms in v2.x.

Quickstart

This quickstart demonstrates loading a configuration using `mmcv.Config` and applying basic image transformations with `mmcv.transforms.Compose` and `Resize`/`RandomFlip`. It showcases the core utilities of MMCV for managing configurations and data pipelines, essential for computer vision tasks.

from mmcv import Config
from mmcv.transforms import Compose, Resize, RandomFlip
import numpy as np

# 1. Load a config (can be a dummy dict or a path to a .py config file)
cfg_dict = dict(
    model=dict(
        type='ResNet',
        depth=50,
        num_classes=1000
    ),
    data=dict(
        train_pipeline=[
            dict(type='Resize', scale=(224, 224)),
            dict(type='RandomFlip', prob=0.5)
        ]
    )
)
cfg = Config(cfg_dict)
print(f"Loaded config model type: {cfg.model.type}")

# 2. Use data transforms
pipeline = Compose([
    Resize(scale=(256, 256)),
    RandomFlip(prob=0.5)
])

# Dummy image data
dummy_img = np.random.rand(512, 512, 3).astype(np.uint8)
data_sample = dict(img=dummy_img, img_shape=dummy_img.shape[:2], ori_shape=dummy_img.shape[:2])

transformed_data = pipeline(data_sample)
print(f"Original image shape: {data_sample['img'].shape}")
print(f"Transformed image shape: {transformed_data['img'].shape}")

view raw JSON →