Wonder how to combine VISSL (Self-Supervised Learning library) with image segmentation from Detectron 2? Here is how.
The story behind this code, can be found in this blog post.
Download balloons dataset into the balloon folder.
Direct link https://github.com/matterport/Mask_RCNN/releases/download/v2.1/balloon_dataset.zip from https://github.com/matterport/Mask_RCNN/releases
The content of the folder should be:
balloon/
val/
...
train/
...
readme.txt
conda create --copy -n ssl_d2 -c pytorch python=3.8 pytorch torchvision torchaudio cudatoolkit=10.2
conda activate ssl_d2
Install APEX from NVIDIA. The installation instructions from vissl webpage works for older pytorch versions.
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir ./
Sanity check if CUDA is available:
python -c 'import torch;print(torch.cuda.is_available())'
Should print True
Clone vissl repo somewhere, code right here is on 0.1.6 release.
git clone --recursive https://github.com/facebookresearch/vissl.git && cd $HOME/vissl/
cd vissl
git checkout v0.1.6
git checkout -b v0.1.6
Install additional tools, as listed in 0.1.6 release instructions [copied here for convenience]
# install vissl dependencies
pip install --progress-bar off -r requirements.txt
pip install opencv-python
# update classy vision install to commit stable for vissl.
# Note: If building from vissl main, use classyvision main.
pip uninstall -y classy_vision
pip install classy-vision@https://github.com/facebookresearch/ClassyVision/tarball/4785d5ee19d3bcedd5b28c1eb51ea1f59188b54d
# update fairscale install to commit stable for vissl.
pip uninstall -y fairscale
pip install fairscale==0.4.6
# install vissl dev mode (e stands for editable)
pip install -e ".[dev]"
# verify installation
python -c 'import vissl, apex'
Last command should work without errors.
Follow the instructions here: https://detectron2.readthedocs.io/en/latest/tutorials/install.html#build-detectron2-from-source
I installed it from commit 717ab9f0aeca216a2f800e43d705766251ba3a55
There should be some errors regarding fvcore and hydra-core. Well . . .
Now, install some additional tools like jupyterlab
conda install jupyterlab ipympl
Launch a juputer lab and go to detectron_vissl_2.ipynb