detectron2-bdd100k

BDD100K Lane Segmentation using Detectron2 API

Installation

Learn your cuda and torch version

nvcc --version
pip list | grep torch

Install appropriate torch version

# https://pytorch.org/get-started/previous-versions/
pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

Install pyyaml

# install dependencies: 
!pip install pyyaml==5.1
!gcc --version
# opencv is pre-installed on colab

Check torch again

import torch, torchvision
print(torch.__version__, torch.cuda.is_available())

Install appropriate detectron2 version

# install detectron2: (Colab has CUDA 10.1 + torch 1.7)
# See https://detectron2.readthedocs.io/tutorials/install.html for instructions
import torch
assert torch.__version__.startswith("1.7")   # please manually install torch 1.7 if Colab changes its default version
!python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu110/torch1.7/index.html
exit(0)  # After installation, you need to "restart runtime" in Colab. This line can also restart runtime

Running

source run.sh

Docker

docker build -t detectron2-bdd100k .

add --no-cache flag if you do not want to use pre-downloaded images

docker run --gpus all -d detectron2-bdd100k bash run.sh

-d detaches you from the container add -it to have interactive pseudo-tty

alternatively

docker run --gpus all -d detectron2-bdd100k python train.py --trainer "custom" --bitmap_path "./subdataset/bitmaps" --images_path "./subdataset/imgs" --epochs 200 --batch_size 32 --eval_period 5 --checkpoint_period 5 --name "bdd100k"

check logs (stdout) of the progrom with the command

docker logs -f <containerId>

sample

docker logs -f d362659da5fc

Get the output file (checkpoints, tensorboard etc.)

docker cp <containerId>:/home/appuser/output .

sample:

docker cp d362659da5fc:/home/appuser/output .

You may use container name instead of container id