Skip to content

Latest commit

 

History

History
71 lines (55 loc) · 2.26 KB

README.md

File metadata and controls

71 lines (55 loc) · 2.26 KB
Raw

Transfer Learning

This project is the implementation of the learning protocol in the paper "Using transfer learning to detect galaxy mergers". This code can be used to exactly replicate the results of the paper with the original merger dataset, or for training with a completely new dataset.

Prerequisites

Linux or OSX

NVIDIA GPU + CUDA CuDNN for training on GPU's

Dependencies

The code is tested on python 3.5

Training requires the following python packages: tensorflow, numpy, keras, tqdm

Get our code

Clone the repo:

git clone https://github.com/SpaceML/merger_transfer_learning.git
cd  merger_transfer_learning/

Run our code

Directory structure for dataset

The original dataset can be found in this Google Drive folder: https://drive.google.com/open?id=1klSL9wZs3cZqAqmvFJVHxkPyXCbXN0-j

The code assumes a directory structure like the following one:

dataset_root/
└── merger
    └── training
        └── images
            ├── 00001.jpeg
             ⋮
    └── validation
        └── images
            ├── 01001.jpeg
             ⋮
    └── test
        └── images
            ├── 02001.jpeg
             ⋮
└── noninteracting
    └── training
        └── images
            ├── 00001.jpeg
             ⋮
    └── validation
        └── images
            ├── 01001.jpeg
             ⋮
    └── test
        └── images
            ├── 02001.jpeg
             ⋮

If you wish to use different class label names, please modify the respective default arguments in make_generators(path_images, positive_class, negative_class) in the file training.py.

If you use your own data set, please also update the corresponding default arguments in training_loop(...) in the file training.py.

Training

python training.py --path_images /path/to/dataset --path_checkpoints /where/checkpoints/go --path_statuses /where/logging/goes --mode [transferlearning|randinit]

The training will not stop automatically, you can monitor the progress in terms of the performance metrics on the validation set, and terminate the training manually as soon as overfitting occurs.