Summary

Concise steps to train the model. Make sure the environment is installed.

For more info on how to use the command line, see here.

Training Dataset Preprocessing

cd <path_to_SIT_FUSE>/src/sit_fuse/datasets/
# Can be run outside of the repo via command line or in a script as well

# generate vectorized dataset
python3 sf_dataset.py -y ../config/<folder>/<yaml_file>
# E.g. set <path_to_yaml> to ../config/model/emas_fire_dbn_multi_layer_pl.yaml 

The YAML file will have info on the training parameters and the files to be included in the training. Ensure the paths and parameters are up to date.

See Code Configuration for documentation on setting up different YAML files.

To learn more about YAML files, see this website and Creating the YAML file.

# generate tiled dataset
python3 sf_dataset_conv.py -y <path_to_yaml>
# Same YAML file

Training Command

cd <path_to_SIT_FUSE>/src/sit_fuse/train/
# Same process as above

# Pretrain encoder
python3 pretrain_encoder.py -y <path_to_yaml>
# Same YAML file as above

# Finetune encoder and train deep clustering MLP head
python3 finetune_dc_mlp_head.py -y <path_to_yaml>

# Train heirarchichal deep clustering MLP heads
python3 train_heirarchichal_deep_cluster.py -y <path_to_yaml>

Inference Command

cd <path_to_SIT_FUSE>/src/sit_fuse/inference/
# Same process as above

python3 generate_output.py -y <path_to_yaml>
# Same yaml file as above

Summary

An example training:

cd /app/SIT_FUSE/sit_fuse/datasets
python3 sf_dataset.py -y ../config/model/gk2a_test_dbn_multi_layer_pl.yaml
python3 sf_dataset_conv.py -y <path_to_yaml>
cd /train
python3 pretrain_encoder.py -y ../config/model/gk2a_test_dbn_multi_layer_pl.yaml
python3 finetune_dc_mlp_head.py -y ../config/model/gk2a_test_dbn_multi_layer_pl.yaml
python3 train_heirarchichal_deep_cluster.py -y ../config/model/gk2a_test_dbn_multi_layer_pl.yaml
cd /inference
python3 generate_output.py -y ../config/model/gk2a_test_dbn_multi_layer_pl.yaml