This repo is about finetuning some famous convolutional neural nets for MARVEL dataset (ship image classification) using TensorFlow.
ConvNets:
Requirements:
- Python 2.7 (Not tested with Python 3)
- Tensorflow >=1.0
- NumPy
- OpenCV2
MARVEL is a dataset contains over 2M ship images collected from shipspotting.com. For image classification in the paper they use 237K images labelled in 26 superclasses.
You can download the whole dataset with python repo they provided.
Or you can download just needed images directly from this dropbox link.
After downloading the dataset, you need to update the paths data/train.txt and data/val.txt.
Make sure dataset is downloaded and file paths are updated.
# Go to related folder that you want to finetune
cd vggnet
# Download the weights
./download_weights.sh
# See finetuning options (there is some difference between them, like dropout or resnet depth)
python finetune.py --help
# Start finetuning
python finetune.py [options]
# You can observe finetuning with the tensorboard (default train_root_dir is ../training)
tensorboard --logdir ../training