Kaggle-PANDA-1st-place-solution
This is the 1st place solution of the PANDA Competition, where the specific writeup is here.
The codes and models are created by Team PND, @yukkyo and @kentaroy47.
Our model and codes are open sourced under CC-BY-NC 4.0. Please see LICENSE for specifics.
You can skip some steps (because some outputs are already in input dir).
Slide describing our solution!
https://docs.google.com/presentation/d/1Ies4vnyVtW5U3XNDr_fom43ZJDIodu1SV6DSK8di6fs/
1. Environment
You can choose using docker or not.
1.1 Don't use docker (haven't tested..)
- Ubuntu 18.04
- Python 3.7.2
- CUDA 10.2
- NVIDIA/apex == 1.0 installed
# main dependency
$ pip install -r docker/requirements.txt
# arutema code dependency
$ pip install git+https://github.com/ildoonet/pytorch-gradual-warmup-lr.git
$ pip install efficientnet_pytorch1.2 Use docker (Recommended)
# build
$ sh docker/build.sh
# run
$ sh docker/run.sh
# exec
$ sh docker/exec.sh2. Preparing
2.1 Get Data
Download only train_images and train_masks.
$ cd input
$ kaggle download ...
$ unzip ...(skip) 2.2 Grouping imageids by image hash threshold
- If you want to do it on your own: https://www.kaggle.com/yukkyo/imagehash-to-detect-duplicate-images-and-grouping
- We will just place the output of the script as:
input/duplicate_imgids_imghash_thres_090.csv
(skip) 2.3 Split kfold
$ cd src
$ python data_process/s00_make_k_fold.py- Is constant with fixed seed
- output:
input/train-5kfold.csv
2.4 Make tile pngs for training
$ cd src
$ python data_process/s07_simple_tile.py --mode 0
$ python data_process/s07_simple_tile.py --mode 2
$ python data_process/a00_save_tiles.py
$ cd ../input
$ cd numtile-64-tilesize-192-res-1-mode-0
$ unzip train.zip -d train
$ cd ..
$ cd numtile-64-tilesize-192-res-1-mode-2
$ unzip train.zip -d train
$ cd ..3. Train base model for removing noise(expected TitanRTX x 1)
Each fold needs about 18 hours.
$ cd src
$ python train.py --config configs/final_1.yaml --kfold 1
$ python train.py --config configs/final_1.yaml --kfold 2
$ python train.py --config configs/final_1.yaml --kfold 3
$ python train.py --config configs/final_1.yaml --kfold 4
$ python train.py --config configs/final_1.yaml --kfold 5- output:
output/model/final_1- Each weights and train logs
4. Predict to local validation for removing noise
Each fold needs about 1 hour.
$ cd src
$ python kernel.py --kfold 1
$ python kernel.py --kfold 2
$ python kernel.py --kfold 3
$ python kernel.py --kfold 4
$ python kernel.py --kfold 5- outputs are prediction results of the hold-out train data:
output/model/final_1/local_preds~~~.csv
5. Remove noise
$ cd src
$ python data_process/s12_remove_noise_by_local_preds.py- output:
output/model/final_1local_preds_final_1_efficientnet-b1.csv- Concatenated prediction results of the hold-out data
- This is used to clean labels
local_preds_final_1_efficientnet-b1_removed_noise_thresh_16.csv- Used to train Model 1
- Base label cleaning results
local_preds_final_1_efficientnet-b1_removed_noise_thresh_rad_13_08_ka_15_10.csv- Used to train Model 2
- Label cleaned to remove 20% Radboud labels
- FYI: we used this csv at final sub on competition: (did not fix seed at time)
input/train-5kfold_remove_noisy_by_0622_rad_13_08_ka_15_10.csv
6. Re-train 5-fold models with noise removed
-
You can replace
output/train-5kfold_remove_noisy.csvtoinput/train-5kfold_remove_noisy_by_0622_rad_13_08_ka_15_10.csvin config -
Only 1,4,5 folds are used for final inference
-
Each fold needs about 15 hours.
Training model 2(fam_taro model):
$ cd src
# only best LB folds are trained
$ python train.py --config configs/final_2.yaml --kfold 1
$ python train.py --config configs/final_2.yaml --kfold 4
$ python train.py --config configs/final_2.yaml --kfold 5Training model 1(arutema model):
Please run train_famdata-kfolds.ipynb on jupyter notebook or
# go to home
$ python train_famdata-kfolds.pyI haven't tested .py, so please try .ipynb for operation.
The final models are saved to models.
Each fold will take 4 hours.
Trained models
Models reproducing 1st place score is saved in ./final_models
7. Submit on Kaggle Notebook
-
kernels
- final sub on competition:
- score: public 0.904, private 0.940 (1st)
- url: https://www.kaggle.com/yukkyo/latesub-pote-fam-aru-ensemble-0722-ew-1-0-0?scriptVersionId=39271011
- reproduced results (seed fixed as in this scripts, you can reproduce)
- score: public 0.894, private 0.939 (1st)
- url: https://www.kaggle.com/kyoshioka47/late-famrepro-fam-reproaru-ensemble-0725?scriptVersionId=39879219
submitted_notebook.ipynb
- Simple 5-fold model to get private 0.935(3rd)
- final sub on competition:
-
You can change paths by changing bellow.
- You must change Kaggle Dataset path for using your reproduced weights
### Model 2
# Line [7]
class Config:
def __init__(self, on_kernel=True, kfold=1, debug=False):
...
...
...
# You can change weight name. But not need on this README
self.weight_name = "final_2_efficientnet-b1_kfold_{}_latest.pt"
self.weight_name = self.weight_name.format(kfold)
...
...
...
def get_weight_path(self):
if self.on_kernel:
# You should change this path to your Kaggle Dataset path
return os.path.join("../input/030-weight", self.weight_name)
else:
dir_name = self.weight_name.split("_")[0]
return os.path.join("../output/model", dir_name, self.weight_name)
### Model 1
# Line [13]
def load_models(model_files):
models = []
for model_f in model_files:
## You should change this path to your Kaggle Dataset path
model_f = os.path.join("../input/latesubspanda", model_f)
...
model_files = [
'efficientnet-b0famlabelsmodelsub_avgpool_tile36_imsize256_mixup_final_epoch20_fold0.pth',
]
model_files2 = [
'efficientnet-b0famlabelsmodelsub_avgpool_tile36_imsize256_mixup_final_epoch20_fold0.pth',
"efficientnet-b0famlabelsmodelsub_avgpool_tile36_imsize256_mixup_final_epoch20_fold1.pth",
"efficientnet-b0famlabelsmodelsub_avgpool_tile36_imsize256_mixup_final_epoch20_fold2.pth",
"efficientnet-b0famlabelsmodelsub_avgpool_tile36_imsize256_mixup_final_epoch20_fold3.pth",
"efficientnet-b0famlabelsmodelsub_avgpool_tile36_imsize256_mixup_final_epoch20_fold4.pth"
]