Right Whale Recognition

• Competition page at Kaggle
• Video recording of presentation

These steps take about 6 hours on a system with 8 processors and an NVIDIA Titan X GPU. Tested only on Ubuntu.

1. Download and install neon 1.1.4

   git clone https://github.com/NervanaSystems/neon.git
   cd neon
   git checkout e09fc11
   make
   source .venv/bin/activate
   

2. Install prerequisites

   pip install scipy scikit-image
   

3. Download the following files from Kaggle:

   imgs.zip
   train.csv
   w_7489.jpg
   sample_submission.csv
   

   Save these to a directory that we will refer to as /path/to/data.

4. Clone this repository

   git clone https://github.com/anlthms/whale-2015.git
   cd whale-2015
   

5. Train models and generate predictions

   ./run.sh /path/to/data
   

6. Evaluate predictions

   Submit subm.csv.gz to Kaggle

• To run on a system that does not have a GPU:

    ./run.sh /path/to/data -bcpu

• For quicker results, decrease imwidth in run.sh.
• The script run.sh first prepares the data for training. If you want to repeat the preparation step, delete the file /path/to/data/prepdone before invoking run.sh again.
• If using a GPU, the results are non-deterministic regardless of how the random number generator is seeded.
• The localizer uses a heuristic to determine when to stop training. If a good optimum is not detected, a message that says "WARNING: model may not be optimal" is displayed.