kaggle-cloud

This is an improved version of the code that was used during the competition which achieved

Public LB : 55/1,538

Private LB : 26/1,538

Inspired from 6th place solution, I added pretraining-step using to make use of train+test set.

Installation

All requirements should be installed from requirements.txt. Using Anaconda is strongly recommended.

conda create -n cloud python=3.6
source activate cloud
pip install -r requirements.txt

Prepare dataset

$ kaggle competitions download -c understanding_cloud_organization
$ unzip understanding_cloud_organization.zip -d data

After downloading and unzipping, the data directory should look like this:

data
  +- train.csv
  +- sample_submission.csv
  +- train_images
  +- test_images

Make Split

To make 5-fold stratified split, run

python make_splits.py

This should save train_splits.csv in data-folder

Pre-Training

To pretrain using train and test set, run

python train.py --log_dir runs --model_name FPN_effb4 --pretrain True --num_epochs 5 --scheduler_name linear_warmup --data_dir ./data

This should save checkpoints in ./runs/FPN_effb4/fold_0/checkpoints

Main Training

The 2nd-stage training was initialized from the last checkpoint of pre-training step

python train.py --log_dir runs --model_name FPN_effb4 --fold 0 --num_epochs 30 --scheduler_name multistep --initial_ckpt ./runs/FPN_effb4/fold_0/checkpoints/4_train_loss.pth

To train with different fold, just change the fold number

Take EMA

To take Exponential-moving-average of top-5 checkpoints, run

python swa.py --log_dir runs --model_name FPN_effb4 --fold 0

This should save a 0_model_ema_5.pth checkpoint file in ./runs/FPN_effb4/fold_0/checkpoints I have uploaded this file as Kaggle dataset. You can download it via

kaggle datasets download -d bibek777/cloudfold0

Make Submission

To create a submssion-csv file using the best checkpoint file, run

python submit.py --initial_ckpt ./runs/FPN_effb4/fold_0/checkpoints/0_model_ema_5.pth --sub_name submission_fold0.csv

This should create a submission_fold0.csv file. To submit this to kaggle, run

kaggle competitions submit -c understanding_cloud_organization -f submission_fold0.csv -m "fold0_submission"

Train 5-folds and enemble them to get better results on LB

imbibekk / Kaggle_Cloud