flevoland

code to setup dedicated dockers and run the processing pipeline

1. make sure the dockerfile is set up properly and running with full access to the machine's GPU (read the dockerfile_readMe.txt for information on how to set this up)

2. if planning on running this on AWS - check out the aws_GPU_instance.txt for detailed instructions on how to properly set the versions of the prerequisite libraries.

3. Run the random_hyper.py file in order to get the random space values for learning rate and momentum. Make sure you get a good spread of values within the random space and that there are no obvious clusters of points. Put the values you are happy with in the hyper_tuning_BBCH_p3.sh in the brackets without commas under declare -a learning_rate() and declare -a momentum() respectively

4. run the hyper_tuning_BBCH_p3.sh script. NOTE: in order for this script to run smoothly you must run it while in the /scripts folder and must have the following directory tree:

├── inputs
│   ├── dataset_BBCH
│   │   ├── BSO0
│   │   ├── CAR1
│   │   ├── CAR4 
│   │   ├── GMA2
│   │   ├── GRA1
│   │   ├── MAI3
│   │   ├── MAI7
│   │   ├── ONI1
│   │   ├── ONI48
│   │   ├── POT1
│   │   ├── POT6
│   │   ├── POT8
│   │   ├── POT9
│   │   ├── SBT14
│   │   ├── SBT39
│   │   ├── SCR2
│   │   ├── TSH0
│   │   ├── VEG1
│   │   ├── WWH2
│   │   ├── WWH3
│   │   └── WWH7
│   ├── flevo5_test_set_scr3000.csv
│   ├── gt_BBCH.csv
├── outputs
├── scripts
│   ├── flevo5_analysisReadyCNNOUTFiles.py
│   ├── flevo5_excludeBSOnTSH.R
│   ├── hyper-tuning-BBCH_p3.sh
│   ├── __init__.pyc
│   ├── label_image_list.py
│   ├── overallstats.py
│   ├── parallel_inference.sh
│   ├── retrain.py
│   ├── Stats-afterparallel.sh
│   ├── tesorboard_vals.py
│   └── train_best_model.sh

5. After training completes, run the inference part in parallel via parallel_inference.sh. NOTE: You must first have installed GNU parallel on you machine. A comprehensive guide on how to install and use it can be found here : https://www.youtube.com/watch?v=OpaiGYxkSuQ&list=PL284C9FF2488BC6D1&index=3&t=2s . NOTE2: one shoule set the JOBS variable in the bash script to however many processes one wants to run in parallel. This decision should be guided by the amount of processing power available (number of CPUs). NOTE3: make sure you have set the environmental variables right in order to allow GNU parallel to be called from any directory, otherwise you would have to specify the entire path when you call it in the final (line 60) of the parallel_inference.sh script.

6. After inferencing run flevo5_analysisReadyCNNOUTFiles.py in order to format the cnn_output_data_check.csv for each model.

7. After formating is complete run the flevo5_excludeBSOnTSH.R in order to create the dfExcludeAll_noDup_best.csv which lists all the images from all the runs that have been labelled as either bare soil ot trash/other. This file should then be located at ../inputs/dfExcludeAll_noDup_best.csv.

8. After formating is complete run the Stats-afterparallel.sh script in order to output a single csv with all the results. It should be located in ../outputs/Random_search_BBCH/Results-all-images.csv.

9. Identify the top performing models and complete steps 4,5,6,7,8 with augmentations in order to push performance higher. NOTE: make sure to change the paths where processing is happening in order not to overwrite the results from the hyper-paramater tuning steps.

10. The folder exploreResultsAndFigures contains all the scripts used to create the figures in the paper and to explore the results in general.