Machine Learning in Practice

Repository for the Dog Breed Identification Kaggle competition. Provided a strictly canine subset of ImageNet, we create a classification system to correctly classify breeds of dogs.

Data

Download here. Make sure you download and extract the data in a folder called data (this folder is in .gitignore so the data does not have to reside on the repository) inside where you cloned this repository.

this/repository/data
├── labels.csv
├── sample_submission.csv
├── test
└── train

System requirements

• Python >3.6
• keras (will install numpy and scipy as well)
• sklearn (machine learning package)
• matplotlib (visualization)
• tensorflow-gpu (keras backend; you can also use regular tensorflow)
• pandas (easy data inspection)
• tqdm (progress bar)
• PIL (image library)
• h5py (HDF5 binary data format)
• jupyter (optional, to work with notebooks)
• seaborn (optional, to plot confusion matrices)

Virtual environment

Ideally you work in a Python virtual environment. If you don't know how to set this up, here are some instructions.

To create a new virtual environment, path can be anything you choose (for example /home/koen/venvs/mlip):

$ python3 -m venv <PATH>

Activate virtual environment:

$ source <PATH>/bin/activate

Once activated, anything you install using pip is installed in the virtual environment separately from your system Python:

$ pip install keras sklearn matplotlib tensorflow-gpu pandas tqdm pillow h5py jupyter

Code formatting

Code should be formatted according to PEP8 guidelines. 4 spaces indentation etc. :)

TensorBoard

train.py logs training progress to the ./training_log directory; these logs can be visualized using tensorboard (example command from within project directory).

$ tensorboard --logdir=./training_log