This repository assumes that the given data set is stored in the following locations:
$ ls data/input/xview3/*.csv
data/input/xview3/train.csv data/input/xview3/validation.csv
$ ls data/input/xview3/downloaded
00a035722196ee86t.tar.gz 4455faa0cb4824f4t.tar.gz 85fe34d1aee53a7ft.tar.gz c07f6ec980c2c149t.tar.gz
014261f774287442t.tar.gz 4518c556b38a5fa4t.tar.gz 864390795b0439b1t.tar.gz c0831dbd6d7f3c56t.tar.gz
(snip)
In the case of training, create a virtual env as follows:
# Setup: Create the virtual env
$ poetry config virtualenvs.in-project true
$ poetry install
$ poetry run pip install albumentations timm segmentation_models_pytorch torch==1.10.0+cu113 torchvision==0.11.1+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
To save storage space, my preprocessing code assumes that the input image file is the original file in .tar.gz format. It does NOT assume pre-extracted files. The batch size and other settings are based on the assumption that two RTX3080 GPU cards are used.
$ bash train.sh
My containerized inference code follows the xView3 evaluation protocol. The detailed usage of the xview3 evaluation protocol is described in https://iuu.xview.us/verify.
To build the Docker Image, you need the model weights generated by train.sh. The pre-computed files can be downloaded from my Google Drive.
You will also need to download and extract gshhg-shp-2.3.7.zip as external data.
GSHHG data can be found at https://www.ngdc.noaa.gov/mgg/shorelines/data/gshhg/latest/
The Docker image containing these model weights and external data has already been uploaded to Docker Hub (smly/kohei-xview3:latest).
## Setup: Build the docker image -----
$ docker build --no-cache -t kohei-xview3 .
## Inference -----
$ docker run \
--shm-size 16G \
--gpus=1 \
--mount type=bind,source=/home/xv3data,target=/on-docker/xv3data \
kohei-xview3 \
/on-docker/xv3data/ \
0157baf3866b2cf9v \
/on-docker/xv3data/prediction/prediction.csv