The code in this repo is all you need to make a first submission to the Humpback Whale Identification Competition. It uses the FastAi library release 1.0.36.post1 for anything up to point 7 in the Navigating through the repository list below (this is important - you are likely to encounter an error if you use any other version of the library). Subsequently I switch to 1.0.39.
For additional information please refer to discussion threads on Kaggle forums: classification, feature learning, detection.
Some people reported issues with running the first_submission notebook. If you encounter the issue, you should be okay to skip to the subsequent notebooks. The one that scores 0.760 on the LB is only_known_train.ipynb
.
- Install the fastai library, specifically version 1.0.36.post1. The easiest way to do it is to follow the developer install as outlined in the README of the fastai repository. Once you perform the installation, navigate to the fastai directory and execute
git checkout 1.0.36.post1
. You can verify that this worked by executing the following inside jupyter notebook or a Python REPL:
import fastai
fastai.__version__
- Clone this repository. cd into data. Download competition data by running
kaggle competitions download -c humpback-whale-identification
. You might need to agree to competition rules on competition website if you get a 403. - Create the train directory and extract files via running
mkdir train && unzip train.zip -d train
- Do the same for test:
mkdir test && unzip test.zip -d test
- Open
first_submission.ipynb
in jupyter notebook and run all cells.
Here is the order in which I worked on the notebooks:
- first_submission - getting all the basics in place
- new_whale_detector - binary classifer known_whale / new_whale
- oversample - addressing class imbalance
- only_known_research - how to modify the architecture and what hyperparams to use
- only_known_train - training on full dataset
- resize - resize the images before training to free up CPU
- siamese network - a fully working prototype of a siamese network
- !!! Important !!! - to make use of some of the new functionality available in fast.ai at this point I switch to 1.0.39.
- fluke detection - train a model to draw bounding boxes surrounding flukes
- !!! Important !!! - here I switch to fastai master to incorporate a bug fix, will annotate with version once a new release comes out
- fluke detection redux - better results, less code, works with current fastai master
##KL: Getting started instructions 1/16/2019
- source deactivate (always do this if your command line starts with (fastai))?
- In fastai: git clone https://github.com/kevinlinke/whale.git
- In fastai/whale/env: pip install -r requirements.txt
- Download Kaggle json on your computer, then on your computer: scp Downloads/kaggle.json paperspace@<your.ip>:~/.kaggle/kaggle.json
- In fastai/whale/data: kaggle competitions download -c humpback-whale-identification
- In fastai/whale/data: mkdir train && unzip train.zip -d train
- In fastai/whale/data: mkdir test && unzip test.zip -d test
- Optional, in fastai/whale/data: rm train.zip && rm test.zip
- In fastai/whale: jupyter notebook --generate-config
- In fastai/whale: jupyter notebook --no-browser --port=8889 --NotebookApp.allow_remote_access=True
- On your computer: ssh -N -L localhost:8888:localhost:8889 paperspace@<your.ip>
- Run the first 5 cells of only_known_research
- Run oversample
- Run resize with SZ = 224
- Run resize with SZ = 448
- ...