This repository is WIP. (28 Mar 2021)
git clone https://github.com/analokmaus/kaggle-ranzcr-clip-public.git
cd kaggle-ranzcr-clip-public
git clone https://github.com/analokmaus/kuma_utils.git kuma_utils is a toolbox I use for competitions and work. Check it out!
conda env create -n {NEW NAME} -f environment.yamlWIP
IMPORTANT: timm version
Since segmentation_models_pytorch requires timm=0.3.2 which does not include ResNet200D.
I added latest timm=0.3.4 as timm_latest in the root directory.
In case you need ResNet200D, you must use import timm_latest.
┣ data
┃ ┣ ranzcr-clip
┃ ┣ (competition files)
┃ ┣ nih_chestxray
┃ ┃ ┣ (nih dataset)
┃ ┣ padchest
┃ ┃ ┣ (padchest dataset)
┃ ┣ mimic
┃ ┣ (mimic dataset)
┃
┣ kaggle-ranzcr-clip-public
┣ scripts
kaggle competitions download ranzcr-clip-catheter-line-classificationkaggle datasets download nih-chest-xrays/datakaggle datasets download raddar/padchest-tubesDue to the license, we cannot host MIMIC CXR dataset.
Please go to MIMIC CXR official website and download by yourself.
CV: 0.9661
Public LB: 0.970
Private LB: 0.973
python train.py --config Segmentation13
python train.py --config SegAndCls12
python inference.py --config SegAndCls12 # generate pseudo labels
python train_external.py --config PretrainStudent08l
python train.py --config SegAndCls12external6CV: 0.9660
Public LB: 0.972
Private LB: 0.973
python train.py --config Segmentation15
python train.py --config SegAndCls14
python inference.py --config SegAndCls14 # generate pseudo labels
python train_external.py --config PretrainStudent09
python train.py --config SegAndCls14external2CV: 0.9671
Public LB: 0.970
Private LB: 0.972
python nfnet_train/train_nfnet_f1_stage1.py
python nfnet_train/train_nfnet_f1_stage23.py
(move and rename weights by yourself)
python inference.py --config SingleModel02
python train_external.py --config Distillation03
python train.py --config SingleModel02external0Adjust batch_size and relevant parameters (learning rate etc.) when you run script.
A machine with four V100 16GB (64GB total) was used to train the following configs:
- Segmentation13
- Segmentation15
- SegAndCls12*
- SegAndCls14*
- PretrainStudent08*
- PretrainStudent09*
A machine with two GF RTX 3090 24GB (48GB total) was used to train the following configs:
- SingleModel02*
- Distillation03