General:

To run the code, use Docker:

cd Docker
docker build -t "pytorch_extended:20.02"
docker run ...

Alternatively, use the environment.yml with anaconda.

Rerun experiments from FlowFrontNet paper:

1. Download the data and checkpoints from here: https://figshare.com/s/dde2f78958173c23aee4. There are two big Zip files: SensorToDryspot, SensorToFlowFront. To recreate the experiments from the paper, we need both.

   • The SensorToDryspot dataset can be used for:
     • Feedforward baseline
     • Finetuning FlowFrontNet with a pretrained Deconv / Conv
   • The SensorToFlowFront dataset can be used to train the Deconv / Conv Network to produce FlowFrontImages

2. Unzip those files in a certain data_path: SensorToFlowFront and SensorToDrySpot

3. Start Trainings:

   • Start the following script for 1140 sensors to flowfront:
     python3 -u ModelTrainerScripts.model_trainer_sensor_1140_to_flow.py --demo data_path/SensorToFlowFront

   • To use the fine-tuned model for binary classification:
     python3 -u ModelTrainerScripts.model_trainer_sensor_1140_to_dryspot.py --demo data_path/SensorToDrySpot --checkpoint_path checkpoint_path

   • For the baseline, run:
     python3 -u ModelTrainerScripts.model_trainer_sensor_1140_dryspot_end_to_end_dense.py --demo data_path/SensorToDrySpot

4. Evaluation:

   • Start the following script for 1140 sensors to flowfront:
     python3 -u ModelTrainerScripts.model_trainer_sensor_1140_to_flow.py --demo data_path/SensorToFlowFront--eval eval_output_path --checkpoint_path checkpoint_path

   • To use the fine-tuned model for binary classification:
     python3 -u ModelTrainerScripts.model_trainer_sensor_1140_to_dryspot.py --demo data_path/SensorToDrySpot --eval eval_output_path --checkpoint_path checkpoint_path

   • For the baseline, run:
     python3 -u ModelTrainerScripts.model_trainer_sensor_1140_dryspot_end_to_end_dense.py --demo data_path/SensorToDrySpot --eval eval_output_path --checkpoint_path checkpoint_path

Caution: New Folders with logs, tensorboard files etc. will be created in the directory of the Datasets, corresponding to the task: SensorToFlowFront or SensorToDryspot. For the trainings and evaluations with 80 and 20 sensors use the respective ModelTrainerScripts.model_trainer_sensor_*_... scripts.