Ecosyste.ms: Awesome

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

Awesome Lists | Featured Topics | Projects

https://github.com/lopuhin/kaggle-lions-2017

2nd place solution for Kaggle NOAA Fisheries Steller Sea Lion Population Count
https://github.com/lopuhin/kaggle-lions-2017

Last synced: 17 days ago
JSON representation

2nd place solution for Kaggle NOAA Fisheries Steller Sea Lion Population Count

Awesome Lists containing this project

README 

        
NOAA Fisheries Steller Sea Lion Population Count

================================================



This is a 2nd place solution for

`Kaggle NOAA Fisheries Steller Sea Lion Population Count `_.



A short overview of the solution on kaggle forums:

https://www.kaggle.com/c/noaa-fisheries-steller-sea-lion-population-count/discussion/35422



Installation

------------



Install required packages (Ubuntu 16.04 or above, CUDA is assumed

to be already installed)::



    sudo apt install \

        gcc make python3-venv python3-pip python3-dev python3-tk libgeos-dev



Make a virtual environment and install python packages::



    python3.5 -m venv venv

    . venv/bin/activate

    pip install -U pip wheel



Install pytorch ``0.1.12.post2``: go to http://pytorch.org, select the version

according to your CUDA and python version, and install it. For example,

for CUDA 8.0 on Python 3.5 and Ubuntu, the command would be::



    pip install http://download.pytorch.org/whl/cu80/torch-0.1.12.post2-cp35-cp35m-linux_x86_64.whl



Install the rest of python packages::



    pip install -r requirements.txt



Download training and testing data (``Test``, ``Train``, ``TrainDotted``)

and place it in ``./data`` folder so it looks like this::



    data

    ├── coords-threeplusone-v0.4.csv

    ├── MismatchedTrainImages.txt

    ├── sample_submission.csv

    ├── Test

    ├── Train

    └── TrainDotted



Training and making a submission

--------------------------------



Make a directory ``_runs``::



    mkdir _runs



Train UNet (this takes about 20 hours and needs 8GB of GPU memory)::



    ./unet.py _runs/unet-stratified-scale-0.8-1.6-oversample0.2 \

        --stratified \

        --batch-size 32 \

        --min-scale 0.8 --max-scale 1.6 \

        --n-epochs 13 \

        --oversample 0.2



**Note:** it may be that 8 GB is slightly not enough and the code may crash

due to GPU memory error after running for one epoch. In this case, run training

in a bash loop (the model is saved and restored successfully).



Result of this step (``best-model.pt``) is provided at

``models/unet-stratified-scale-0.8-1.6-oversample0.2/`` in the archive.



Make predictions for validation set (and all other images not used due to ``--limit``,

if ``--predict_all_valid`` option is used instead of ``--predict_valid``, like below),

also don't forget to pass all the other params from training

(this should take less than an hour)::



    ./unet.py _runs/unet-stratified-scale-0.8-1.6-oversample0.2 \

        --stratified \

        --batch-size 32 \

        --min-scale 0.8 --max-scale 1.6 \

        --n-epochs 13 --oversample 0.2 \

        --mode predict_all_valid



Train a regression model on this predictions (takes less than 10 minutes)::



    ./make_submission.py _runs/unet-stratified-scale-0.8-1.6-oversample0.2 train



Result of this step (``regressor.joblib``) is provided at

``models/unet-stratified-scale-0.8-1.6-oversample0.2/`` in the archive.



Now you need to predict all test (this takes about 12 hours)::



    ./unet.py _runs/unet-stratified-scale-0.8-1.6-oversample0.2 \

        --stratified \

        --batch-size 32 \

        --min-scale 0.8 --max-scale 1.6 \

        --n-epochs 13 --oversample 0.2 \

        --mode predict_test



Now make submission with (takes a few hours)::



    ./make_submission.py _runs/unet-stratified-scale-0.8-1.6-oversample0.2 predict



Submission file will be in

``_runs/unet-stratified-scale-0.8-1.6-oversample0.2/unet-stratified-scale-0.8-1.6-oversample0.2.csv``