This is solution for the SpaceNet three challenge as submitted to be checked by the Topcoder team. This will give you some information about TopCoder platform.
More stuff from us
Training
- 6+ core modern CPU (Xeon, i7) for fast image pre-processing;
- The models were trained on 2 * GeForce 1080 Ti;
- Training time on my setup ~ 3 hours for models with 8-bit images as inputs;
- Disk space - ~30GB should be more than enough for the Docker image + files;
- The dataset weighs ~100-150 GB and is copied, so make some room;
Inference
- 6+ core modern CPU (Xeon, i7) for fast image pre-processing;
- On 2 * GeForce 1080 Ti inference takes 3-5 minutes;
- Graph creation takes 5-10 minutes;
Data download guide from the authors.
Final testing guide from the authors.
Steps to reproduce the result as per the guide:
-
You can clone the repository to see the code for yourself
git clone REPO_URL .; -
Download the
.zipfile to some folder; -
Download the Dockerfile to the same folder;
-
Building and running the container (assuming nvidia-docker2):
docker build -t aveysov .
docker run --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all -it -v /path/to/data:/home/keras/notebook/data -p 8888:8888 -p 6006:6006 --shm-size 8G aveysov
Jupyter notebook is launched under port 8888. Port 6006 is for tensorboard to monitor the training process, which is optional.
docker exec -it --user root 09654f4db9f9 /bin/bash
- Inside of the container you can invoke
sh train.sh - to train the model. Training for 1 epoch replaces the weights;
sh test.sh - to test the model and generate the linestrings;