This is the technical documentation of our project with RSIM of TU BERLIN.

Setup

First install the conda environment at cluster_instructions/conda_env.yml and activate it.

Data Set (Deep Globe Patches)

Then download the Deepglobe dataset from KAGGLE. Place it in the project with the path data/deepglobe. Then run the patch-sampling pipeline: data_pipeline/deepglobe/patch_sampling.py. This creates LMDB files for the train, test and valid set at data/deepglobe_patches/[train/test/valid]/.

Word Embeddings

The folder data/glove contains the embeddings for the deepglobe-labels for embeddings spaces of 50 and 300. If other embeddings are necessary, or a different dataset is used. These can be created by modifying the src/wordembedding/glove.py and running it. To use this downlaod the glove txt files from Stanford. Currently d = [50,100,200,300] are available there. For different embedding size, retrain the glove model.

Run Example

The most important parameters are contained in this exemplary run. For all parameters check out src/config_args.py, for our parameterized runs check out cluster_instructions/<model>_<loss>.sh

python main.py -model CbMLC -loss weighted_bce -optim sgd -d_model 50-lr 0.0001 -add_noise 0.1 -sub_noise 0.1