MMCV Full (OpenMMLab Computer Vision Foundation)

1.7.2 verified Mon Apr 27 auth: no python maintenance

OpenMMLab computer vision foundation library providing CUDA/C++ ops, data transforms, and utilities for deep learning. Current PyPI version is 1.7.2 (legacy mmcv-full). MMCV has been unified into a single `mmcv` package starting from v2.0.0; mmcv-full is no longer updated for v2.x. Release cadence: occasional.

pip install -U openmim && mim install mmcv-full

Common errors

error ModuleNotFoundError: No module named 'mmcv._ext' ↓

cause Mismatch between compiled ops and installed PyTorch/CUDA version.

fix

Use mim install mmcv-full to install a pre-built wheel matching your environment, or rebuild from source.

error ImportError: cannot import name 'fuse_conv_bn' from 'mmcv' ↓

cause Function was removed in mmcv v2.0.0. mmcv-full v1.7.2 still has it.

fix

If using v1.x, ensure correct version. For v2.x, use mmengine for fused BN.

error RuntimeError: Attempting to deserialize object on CUDA device 1 but torch.cuda.device_count() is 1 ↓

cause Model saved with multi-GPU but loaded on single GPU.

fix

Load with map_location='cuda:0' or set CUDA_VISIBLE_DEVICES appropriately.

Warnings

breaking mmcv-full v1.x and mmcv v2.x are not API compatible. mmcv-full v1.x uses `from mmcv.runner` etc. mmcv v2.x moved training modules to mmengine. ↓

fix For new projects, use `mmcv` (v2.x) with `mmengine`. For existing code, stay on mmcv-full or upgrade with adaptation.

gotcha pip install mmcv-full may install a CPU-only build or fail due to missing CUDA support. Always verify the built wheel matches your torch version and CUDA. ↓

fix Use `pip install -U openmim && mim install mmcv-full` to auto-select correct wheel.

deprecated mmcv-full is deprecated for v2.x. New features and fixes are only on the unified `mmcv` package (v2.x). ↓

fix Migrate to `mmcv` (v2.x) + `mmengine` for new projects.

Install

pip install mmcv-full

Imports

mmcv
wrong
```
from mmcv import *
```
correct
```
import mmcv
```
Wildcard import pollutes namespace; not recommended

Quickstart

Basic usage: image I/O and transforms.

import mmcv
import numpy as np
from PIL import Image

# Load an image from file
img = mmcv.imread('path/to/image.jpg')
print('Image shape:', img.shape)

# Apply a filter (example: Gaussian blur)
blurred = mmcv.gaussian_blur(img, kernel_size=5)

# Resize image
resized = mmcv.imresize(img, (224, 224))

# Visualize or save
mmcv.imwrite(resized, 'resized.jpg')