TBrain histapathology segmentation contest.
The training relative resource can be download in 2022 STAS
Download custom training datasets at 2022 STAS/SEG_Train_Datasets
Download package by execute:
$sh preprocess.sh
Or according to the requirements : ./requirements/Pipfile
, download the
package by pip
.
./trainlist
is the list of cross validation training/validation set.
Config file for training and inferencing model is in ./cfg
.
setting.yaml
is the config for experiment name, image root, inference path,
use crossvalidation , use soup setting.
wandbcfg.yaml
is the config for training model hyperparameter.
After setting config file, start training:
$python training.py
The training ckpt will save in ./result
We provide Data Distributed Parallel and will automatically detect gpu number.
Modified inference.py
for to assign inference experiment ckpt.
Use voting.py
and embedding.py
to get majority agreement mask.
$python inference.py
$python voting.py
$python embedding.py
Pretrain weight can be download in 2022 STAS/model weight
Or best model ckpt in private score is base_DL_plus_10fd0
Name f1-score precision recall
U+_nc_ef4ap_FTL_10fd4_soup5 0.904259 0.899413 0.909157 V
4, 5
base_DL_plus_10fd0 0.89715 0.888939 0.905514
base_DL_plus_10fd3 0.900433 0.906386 0.894558 V
base_DL_plus_10fd7 0.898166 0.889426 0.907079
base_DL_plus_10fd8 0.901063 0.886515 0.916096 V
0, 3, 5, 7, X
U+_nc_ef4ap_sDL_10fd4 0.894808 0.868046 0.923273
U+_nc_ef4ap_sDL_10fd6 0.902042 0.917095 0.887474 V
U+_nc_ef4ap_sDL_10fd8 0.902314 0.902158 0.902471 V
U+_nc_ef4ap_FTL_10fd0 0.895147 0.883711 0.906883
U+_nc_ef4ap_FTL_10fd1 0.886003 0.872314 0.900128
U+_nc_ef4ap_FTL_10fd2 0.892547 0.885231 0.899985
U+_nc_ef4ap_FTL_10fd3_last 0.894663 0.915445 0.874804
U+_nc_ef4ap_FTL_10fd4 0.899646 0.897439 0.901864 V
U+_nc_ef4ap_FTL_10fd5 0.894417 0.888879 0.900025
U+_nc_ef4ap_FTL_10fd6 0.898047 0.882558 0.914091
U+_nc_ef4ap_FTL_10fd7 0.897821 0.886498 0.909436
U+_nc_ef4ap_FTL_10fd8 0.893720 0.891105 0.896352
U+_nc_ef4ap_FTL_10fd9 0.892483 0.884145 0.900980