grseb9s / hyperspectral-fcn

HyperSpectral FCN

This is an implementation of FCN and Attention layers to recognize HyperSpectral images.

Prerequisites

• Python 2.7
  • Numpy
  • Caffe
  • Scipy
• An Nvidia GPU (Not required but recommended)
• A Training and Validation dataset
• VGG model trained weights

Setup

Anaconda is recommended for Python and its prerequisites. This implementation uses Python 2.7 so be sure to install the right one. You can install Anaconda from here.

The models are currently made to train from the VOC 2012 dataset and SBD dataset. Put these datasets in data/VOC and data/sbdd respectively.

You can get the trained weights for the VGG model on the Caffe Model Zoo or right here. Put the .caffemodel file in data/ilsvrc-nets.

Training

To train this model, run solve.py in fcn32s, then in fcn16s, and finally in fcn8s. Alternatively, run solve.py in fcn8s-atonce.

You can change what model the 16s and 8s models run off of in solve.py by replacing:

weights = '../fcn16s/snapshot/train_iter_8000.caffemodel'

with the model you want to train with:

weights = '../path/to/your/model.caffemodel'

Todo

• Make module to import HyperSpectral images
• Add the Attention Layer
• Optional CPU/GPU
• Visualization