Skip to content

Usage_guide.md

Latest commit

 

History

History
180 lines (116 loc) · 6.76 KB

Usage_guide.md

File metadata and controls

180 lines (116 loc) · 6.76 KB
Raw

INFERENCE GUIDE

1. Download the pretrained models

OD model

SS model

2. Run inference on your images

2.1 OD Inference

2.1.1 Run yolov5 model inference

cd Halimeda/object_detection/yolov5/

python detect.py --weights $path_to_model.pt \
    --project $path_to_your_project_folder --name $results_od_folder --data data/halimeda.yaml \
    --source $path_to_your_images --conf-thres 0.1 --iou-thres 0.5 \
    --save-txt --save-conf --imgsz 1024

ARGS EXPLAINED:

$path_to_model.pt: path to OD model downloaded from the link to zenodo
$path_to_your_project_folder: path to your project folder
$results_od_folder: inference results will be stored in $path_to_your_project_folder/$results_od_folder path
$path_to_your_images: the path containing the images for performing inference.

2.1.2 Convert the bounding boxes predictions to a coverage

cd Halimeda/scripts/

python coverage_od.py --shape 1024 --path_txt $path_to_your_project_folder/$results_od_folder/labels \
    --path_out $path_to_your_output_coverage_folder --grid 500

ARGS EXPLAINED:

$path_txt: path to the predicted bounding boxes in txt form resulting from executing detect.py
$path_to_your_output_coverage_folder: path to the folder where the bounding box predictions converted to grayscale images are stored

2.1.3 Optional Apply a threshold to the coverage grayscale images

cd Halimeda/scripts
python apply_theshold.py --input_dir $path_to_your_output_coverage_folder --output_dir $path_to_your_output_coverage_folder_thresholded --threshold $threshold_value

ARGS EXPLAINED:

$path_to_your_output_coverage_folder: path to the coverage grayscale images
$path_to_your_output_coverage_folder_thresholded: path to the output binarized images. (For sparse scenarios) -> pixels in white are halimeda and pixels in black will be considered background.
$threshold_value: threshold to binarize the images. For an OD scenario, the recommended experimental values are 43 or 75.

2.2 SS inference

2.2.1 Run the trained SS_model "model.h5" inference

cd semantic_segmentation/SS_halimeda/scripts
python3 inference.py --run_path SS_Halimeda/model/ --data_path $path_to_your_images --shape 1024 --path_out results_ss_folder

ARGS EXPLAINED:

$results_ss_folder: inference results will be stored in $results_ss_folder path
$path_to_your_images: the path containing the images for performing inference.

2.2.2 Optional Apply a threshold to the predictions

cd Halimeda/scripts
python apply_theshold.py --input_dir $results_ss_folder --output_dir $results_ss_folder_thresholded --threshold $threshold_value

ARGS EXPLAINED:

$results_ss_folder: path to the predictions of the SS net
$results_ss_folder_thresholded: path to the output binarized images. (For dense scenarios) -> pixels in white are halimeda and pixels in black will be considered background.
$threshold_value: threshold to binarize the images. For an SS scenario, the recommended experimental value is 123

3 Merge

3.1 Weighted merge

cd Halimeda/scripts
python w_merge.py --path_od $path_to_your_output_coverage_folder \
    --path_ss $results_ss_folder \
    --path_merge $results_w_merge_folder

ARGS EXPLAINED:

$path_to_your_output_coverage_folder: path to the coverage grayscale images corresponding to the OD predictions
$results_ss_folder: path to the predictions of the SS net
$results_w_merge_folder: path where the results of the weighted merge will be stored.

3.2 AUC merge

cd Halimeda/scripts

python merge_blobs.py --path_od $path_to_your_project_folder/$results_od_folder/labels \
    --path_ss $results_ss_folder \
    --path_merge $results_AUC_merge_folder --iter 1  --ns 12 3 1 2 1 1 1 1 1 1

TUNING GUIDE:

In this section, examples of calls for every tuning step are provided.

For more information on how to reparameterize every step of the algorithm, please refer to our paper.

In Case that you want to retrain the OD model:

cd Halimeda/object_detection/yolov5

python train.py --img 1024 --batch 8 --epochs 200 --data halimeda.yaml --weights yolov5x.pt --project projects/halimeda/final_trainings/yolo_XL/ --name base_1  --single-cls --hyp data/hyps/final/hyp.base.yaml

Keep in mind:

  • You must modify halimeda.yaml to match the paths to your dataset folder.
  • If you want to train from pretrained ImageNet weights use --weights yolov5x.pt, if you want to retrain our model use --weights $path_to_model.pt with the corresponding path where you have stored our downloaded model.
  • Change the hyperparameters in data/hyps/final/hyp.base.yaml according to your needs.

Eval command:

python3 Halimeda/scripts/eval.py --run_name 000033 --path_pred ../runs/1/000033/inference --path_out ../runs/1/000033/ --path_gt ../data/test/mask --name xxx.py --pred_path /mnt/c/Users/haddo/yolov5/projects/halimeda/final_trainings/yolo_XL/base_3/inference_test/coverage --gt_im_path  /mnt/c/Users/haddo/yolov5/datasets/halimeda/coverage/test/ --gt_label_path /mnt/c/Users/haddo/yolov5/datasets/halimeda/labels/test/ --run_name base_3 --save_path /mnt/c/Users/haddo/yolov5/projects/halimeda/final_trainings/yolo_XL/base_3/inference_test --shape 1024

In Case that you want to retrain the SS model:

Following, you have an example of calls, substitute in each instruction substitute for your paths and config

python3 train.py --run_path ../runs/1/0009/ --data_path ../data/1/ --shape 1024 --batch 3 --learning 0.009

python3 inference.py --run_path ../runs/5/0003_da/ --data_path ../data/5_da/val/img --shape 1024

Find best threshold with val partition

python3 eval.py --run_name 0009 --path_pred ../runs/1/0009/inference_val --path_out ../runs/1/0009/ --path_gt ../data/1/val/mask --name 'val'

Final evaluation over the independent test set with the obtained thr

python3 eval_thr.py --run_name 000033 --path_pred ../runs/5/000033/inference_test --path_out ../runs/5/000033/ --path_gt ../data/test_ss/mask --name 'test' --thr 123

Tuning of the w_merge and AUC_merge:

Find the best w_merge weights:

python merge_weights.py --path_od /mnt/c/Users/haddo/DL_stack/Halimeda/inference_val/OD/ \
                    --path_ss /mnt/c/Users/haddo/DL_stack/Halimeda/inference_val/SS/ \
                    --path_merge /mnt/c/Users/haddo/DL_stack/Halimeda/weighted_merge/w_sweep/

Find the n thresholds:

python get_n.py --path_od ../inference_val/OD_labels/ \
    --path_ss ../inference_val/SS/ --path_ss_gt ../dataset/all_val/gt/ --path_out ../AUC_merge/cthr0/ --iter 1