!!! danger "Important"
We currently only support Linux/MacOS installations
Depending on what version of cuda driver you'd like to use, you can install different versions of torch builds. If you're not sure which version to choose, we advise to use the current torch default cuda-10.2
=== "cuda-10.2"
pip install torch==1.10.0+cu102 torchvision==0.11.1+cu102 -f https://download.pytorch.org/whl/torch_stable.html
=== "cuda-11.1"
pip install torch==1.10.0+cu111 torchvision==0.11.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
=== "cpu"
pip install torch==1.10.0+cpu torchvision==0.11.1+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
??? note "checking your torch-cuda version"
To see what version of torch and cuda is installed in your current environment, run:
python -c "import torch;print(torch.__version__, torch.version.cuda)"
output:
1.10.1+cu102 10.2
Your installed torch version will determine which version of mmcv-full you can install.
Installing mmcv-full is optional, yet it will let you unleash the full potential of icevision and allow you to use the large library of models available in mmdet, therefore we strongly recommend doing it.
=== "cuda-10.2"
pip install mmcv-full==1.3.17 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.10.0/index.html pip install mmdet==2.17.0
=== "cuda-11.1"
pip install mmcv-full==1.3.17 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.10.0/index.html pip install mmdet==2.17.0
=== "cpu"
pip install mmcv-full==1.3.17 -f https://download.openmmlab.com/mmcv/dist/cpu/torch1.10.0/index.html pip install mmdet==2.17.0
??? "testing mmcv installation"
Installing `mmcv-full` can be tricky as it depends on both the exact `torch` and `cuda` version.
We highly recommend that you test your installation. You can verify it by executing the following command inside your virtual environment:
```bash
curl -sSL https://raw.githubusercontent.com/open-mmlab/mmcv/master/.dev_scripts/check_installation.py | python -
```
If everything went fine, you should see something like the following:
```
Start checking the installation of mmcv-full ...
CPU ops were compiled successfully.
CUDA ops were compiled successfully.
mmcv-full has been installed successfully.
Environment information:
-----------------------------------------------------------
sys.platform: linux
Python: 3.8.12 (default, Oct 12 2021, 13:49:34) [GCC 7.5.0]
CUDA available: True
GPU 0: GeForce RTX 2060
CUDA_HOME: /usr/local/cuda
NVCC: Build cuda_11.1.TC455_06.29069683_0
GCC: gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0
PyTorch: 1.10.0+cu111
PyTorch compiling details: PyTorch built with:
- GCC 7.3
- C++ Version: 201402
- Intel(R) Math Kernel Library Version 2020.0.0 Product Build 20191122 for Intel(R) 64 architecture applications
- Intel(R) MKL-DNN v2.2.3 (Git Hash 7336ca9f055cf1bfa13efb658fe15dc9b41f0740)
- OpenMP 201511 (a.k.a. OpenMP 4.5)
- LAPACK is enabled (usually provided by MKL)
- NNPACK is enabled
- CPU capability usage: AVX2
- CUDA Runtime 11.1
- NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86
- CuDNN 8.0.5
- Magma 2.5.2
- Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=11.1, CUDNN_VERSION=8.0.5, CXX_COMPILER=/opt/rh/devtoolset-7/root/usr/bin/c++, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_KINETO -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -DEDGE_PROFILER_USE_KINETO -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=1.10.0, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON,
TorchVision: 0.11.1+cu111
OpenCV: 4.5.4
MMCV: 1.3.17
MMCV Compiler: GCC 7.3
MMCV CUDA Compiler: 11.1
-----------------------------------------------------------
```
Icevision is distributed in 2 different eggs:
icevision[all]- recommended - complete icevision package with all dependenciesicevision[inference]- minimal dependencies, useful for deployment or simply parsing and viewing your dataset
we recommend to install the stable release but if you want to use the most recent, bleeding edge version of the library or would like to contribute, here is how to do it:
- stable
pip install icevision[all]- bleeding edge
pip install git+https://github.com/airctic/icevision.git@master#egg=icevision[all] --upgrade- editable mode (for developers)
git clone --depth=1 https://github.com/airctic/icevision.git
cd icevision
pip install -e .[dev]
pre-commit install??? "installing using different cuda version"
Installing icevision with different cuda version is possible, however it is only
recommended for more experienced users.
The main constraint here is `mmcv-full` and `torch` versions compatibility. In short,
torch is build for a specific cuda driver version, mmcv-full on the other hand is
distributed for a specific torch build.
To see which mmcv-full wheels are available for which versions of torch, check the
table at [mmcv installation guide](https://mmcv.readthedocs.io/en/latest/get_started/installation.html).
!!! note
running pip install icevision will install icevision[inference] by default
The easiest way to install icevision with all its dependencies is to use our conda
environment.yml file. Creating a conda environment is considered as a best practice
because it avoids polluting the default (base) environment, and reduces dependencies
conflicts.
!!! warning
using the environment.yml works only on cuda-10.2 enabled devices. If your GPU
architecture is Ampere or newer, you have to use the pip installation method.
If you are installing MMCV from the wheel like described above and still are having problems with CUDA you will probably have to compile it locally. Do that by running:
pip install mmcv-full
If you encounter the following error it means you will have to install CUDA manually (the one that comes with conda installation will not do).
OSError: CUDA_HOME environment variable is not set. Please set it to your CUDA install root.
Try installing it with:
sudo apt install nvidia-cuda-toolkit
Check the installation by running:
nvcc --version
If you encounter the following error, when installation process is building wheel for pycocotools:
unable to execute 'gcc': No such file or directory
error: command 'gcc' failed with exit status 1
Try installing gcc with:
sudo apt install gcc
Check the installation by running:
gcc --version
It should return something similar:
gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0
Copyright (C) 2019 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
After that try installing icevision again.