Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/lromul/ranzcr-clip

Kaggle | 22nd place solution for RANZCR CLiP - Catheter and Line Position Challenge.
https://github.com/lromul/ranzcr-clip

Last synced: 21 days ago
JSON representation

Kaggle | 22nd place solution for RANZCR CLiP - Catheter and Line Position Challenge.

Host: GitHub
URL: https://github.com/lromul/ranzcr-clip
Owner: lRomul
License: mit
Created: 2020-12-30T23:05:22.000Z (almost 4 years ago)
Default Branch: master
Last Pushed: 2021-04-19T20:39:23.000Z (over 3 years ago)
Last Synced: 2024-10-03T17:45:08.258Z (about 1 month ago)
Language: Python
Homepage:
Size: 231 KB
Stars: 11
Watchers: 1
Forks: 0
Open Issues: 0
Metadata Files:
- Readme: README.md
- License: LICENSE

Awesome Lists containing this project

README

# Solution for RANZCR CLiP - Catheter and Line Position Challenge

![header](https://user-images.githubusercontent.com/11138870/111442192-23d49a80-8719-11eb-8d4b-7828bdf5632f.png)

Source code of 22nd place solution for [RANZCR CLiP - Catheter and Line Position Challenge](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification).

## Solution

Key points:
* EfficientNet
* 1024x1024 image resolution
* Soft pseudo labels
* Some MLOps for training and making a submission

## Experiments

The progress of the solution during the competition can be seen in the laboratory journal.
It describes all the single models and ensembles and shows CV, Public/Private LB scores.

Link: https://docs.google.com/spreadsheets/d/112wrfuQjNXEFyqQLVhu79Vf0uOabnZ1MaayEts2Gvto/edit?usp=sharing

Experiments:
![experiments](https://user-images.githubusercontent.com/11138870/111454092-51bfdc00-8725-11eb-9fa4-2657868a33ce.png)
Ensembles:
![ensembles](https://user-images.githubusercontent.com/11138870/111455050-651f7700-8726-11eb-8fef-b4d800ef51a5.png)

## Quick setup and start

### Requirements

* Nvidia drivers >= 460, CUDA >= 11.2
* [Docker](https://www.docker.com/), [nvidia-docker](https://github.com/NVIDIA/nvidia-docker)

The provided Dockerfile is supplied to build an image with CUDA support and cuDNN.

### Preparations

* Clone the repo.
```bash
git clone [email protected]:lRomul/ranzcr-clip.git
cd ranzcr-clip
```

* Download and extract [dataset](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/data) to the `data` folder.

### Run

Batch size tuned for RTX 3090.

* Train first stage models
```bash
./train.sh b7v3_001 2 0,1 all # ./train.sh EXPERIMENT N_DEVICES DEVICES FOLDS
./train.sh b6v3_001 # default settings: ./train.sh EXPERIMENT all all all
```

* Make soft pseudo labels
```bash
make COMMAND="python predict.py --experiment b7v3_001"
make COMMAND="python predict_val.py --experiment b7v3_001"
make COMMAND="python predict.py --experiment b6v3_001"
make COMMAND="python predict_val.py --experiment b6v3_001"
```

* Train second stage models
```bash
./train.sh kdb3v3_b71_001
./train.sh kdb4v3_b61_002
./train.sh kdb4v3_b71_001
```

* Make submission
```bash
cd data/ranzcr-deps/
./download.sh kdb3v3_b71_001,kdb4v3_b61_002,kdb4v3_b71_001
```

* Upload the contents of the folder `data/ranzcr-deps/` to Kaggle dataset with the name `RANZCR CLiP Dataset`.

* Connect competition data and `RANZCR CLiP Dataset` to Kaggle Code. Run script code from `data/ranzcr-deps/kernel.py`.