Here are detailed instructions about Efficiency Prize stage for Severstal: Steel Defect Detection challenge.
- Participants who finish in the top 50 on the Kaggle Private Leaderboard will be allowed for Efficiency Prize stage.
- Participants should submit to Severstal Docker image of their solution and inference code with Dockerfile from which Docker image was built (see instruction).
- Participants are allowed to submit a different code than the one that places them in the Top 50, as long as that solution also scores within the Top 50.
- Submission deadline for the Efficiency Prize is 11:59 PM UTC November 10, 2019.
- Participant who provided solution that gets the fastest inference on test images and still scores within the Top 50 is declared the winner of Efficiency Prize.
- The winner should deliver to Severstal the final model's software code as used to generate the winning Submission and associated documentation.
- Main competition Rules also apply.
- Severstal will post all of the results of the evaluations for the Efficiency Prize in the forum for the Competition.
Docker containers provided by participants run on following hardware configuration.
Intel(R) Core(TM) i7-8700K CPU @ 3.70GHz
GeForce GTX 1080 Ti, Driver Version: 390.67
CUDA 10.0 cudnn 7
64GiB RAM
CentOS 7
This instruction is intended to help kagglers to build Docker image for their solutions. We assume you use Python to build your solution.
Repository contains two folders:
- template : Contains Dockerfile template and necessary code to wrap your solution for Efficiency scoring. Use it to build your own Docker image.
- example : Contains an example of using template to package up a dummy-model into Docker image which successfully runs on our hardware configuration. Example also was tested on AWS Deep Learning AMI K80 and on laptop with Core(TM) i7 CPU/GeForce GTX 860M/Ubuntu 18.04.3.
Steps to package up your solution into Docker image:
- Install Docker on your PC. Version 19.03 is required.
- Copy template folder to your PC.
- Put your solution code, model weights and all necessary files into template folder.
- Now you should wrap you inference code for Efficiency scoring. Open app.py template and place your inference code into
app()function. - Then you should modify Dockerfile template. Open Dockerfile template and place commands to install necessary libs and dependencies there.
- Open terminal and change dir to template folder with following command. Replace path_to_template_folder with path to template folder.
$ cd path_to_template_folder - Replace your_docker_image_name with some name and build Docker image with command:
$ sudo docker build -f Dockerfile -t your_docker_image_name .
How to test your Docker image:
- To test your Docker image with GPU usage you should install nvidia-docker first.
- Put test images to some path your_test_images_path.
- To run Docker container from your Docker image use one of the following commands. Replace your_test_images_path with path to some test images, your_submission_path with path where submission.csv to be saved and your_docker_image_name with the name you chose for your Docker image.
- To test container on GPU:
$ sudo docker run -e MODEL_RUN_DEVICE='cuda:0' -it --gpus all -v your_test_images_path:/usr/src/app/test_images -v your_submission_path:/usr/src/app/temp your_docker_image_name- To test container on CPU:
$ sudo docker run -e MODEL_RUN_DEVICE='cpu' --rm -v your_test_images_path:/usr/src/app/test_images -v your_submission_path:/usr/src/app/temp your_docker_image_name - If app.py runs successfully, you will see submission.csv in your_submission_path.