CLI API¶
fb_predict¶
Inference CLI script for flatbug.
A comprehensive CLI API for flatbug inference with support for hyperparameter configuration, flexible input parsing, output format specification, and hardware specification.
- Usage:
fb_predict -i INPUT_PATH_OR_DIRECTORY -o OUTPUT_DIRECTORY [OPTIONS]- Options:
- -h, --help
show this help message and exit
- -i INPUT, --input INPUT
A image file or a directory of image files
- -o OUTPUT_DIR, --output OUTPUT_DIR
The result directory
- -w MODEL_WEIGHTS, --model-weights MODEL_WEIGHTS
The .pt file
- -p INPUT_PATTERN, --input-pattern INPUT_PATTERN
The pattern to match the images. Default is ‘[^/]*.([jJ][pP][eE]{0,1}[gG]|[pP][nN][gG])$’ i.e. jpg/jpeg/png case-insensitive.
- -n MAX_IMAGES, --max-images MAX_IMAGES
Maximum number of images to process. Default is None. Truncates in alphabetical order.
- -R, --recursive
Process images nested within subdirectories of the input.
- -s SCALE_BEFORE, --scale-before SCALE_BEFORE
Downscale the image before detection, but crops from the original image.
- --single-scale
Use single scale.
- -g GPU, --gpu GPU
Which device to use for inference. Default is ‘cuda:0’, i.e. the first GPU.
- -d DTYPE, --dtype DTYPE
Which dtype to use for inference. Default is ‘float16’.
- -f, --fast
Use fast mode.
- --config CONFIG
The config file.
- --no-crops
Do not save the crops.
- --no-overviews
Do not save the overviews.
- --no-metadata
Do not save the metadata.
- --only-overviews
Only save the overviews.
- --long-format
Use long format for storing results.
- -S, --no-save
Do not save the results.
- -C, --no-compiled-coco
Skip the production of a compiled COCO file (for all images).
- -v, --verbose
Verbose mode.
fb_eval¶
Custom flatbug evaluation script. Requires prediction to have been run already, and ground truth labels should be supplied in COCO format.
For end-to-end use the script scripts/eval/end_to_end_eval.sh (requires R for summary statistics and figures).
- Usage:
fb_evaluate -p PREDICTIONS -g GROUND_TRUTH -I IMAGE_DIRECTORY -o OUTPUT_DIRECTORY [OPTIONS]- Options:
- -h, --help
show this help message and exit
- -p PREDICTIONS, --predictions PREDICTIONS
Path or pattern to the predictions files
- -g GROUND_TRUTH, --ground_truth GROUND_TRUTH
Path to the ground truth file
- -I IMAGE_DIRECTORY, --image_directory IMAGE_DIRECTORY
Path to the image directory
- -o OUTPUT_DIRECTORY, --output_directory OUTPUT_DIRECTORY
Path to the output directory
- --config CONFIG
Path to the configuration file
- -P, --plot
Plot the matches and the IoU matrix
- -b, --no_boxes
Do not plot the bounding boxes
- -c, --coco_predictions
Whether the predictions are already in a COCO format (legacy)
- -s SCALE, --scale SCALE
Scale of the output images. Defaults to 1. Lower is faster.
- -n N
Number of images to process. Defaults to -1 (all images)
- --workers WORKERS
Number of workers to use for the evaluation. Defaults to 8.
- --combine
Combine the results into a single CSV file
fb_train¶
flatbug training script.
The flatbug training script uses a lightly modified YOLO training interface (https://docs.ultralytics.com/modes/train/), with a few additional parameters.
See scripts/experiments/best_train/default.yaml for an example training config.
- Usage:
fb_train [-d DATA_DIR] [-c CONFIG_FILE] [-r]- Options:
- -h, --help
show this help message and exit
- -d DATA_DIR, --data-dir DATA_DIR
The directory containing the prepared data (i.e., the output of fb_prepare.py
- -c CONFIG_FILE, --config-file CONFIG_FILE
A YAML-formatted config file that overrides the default training meta-parameters
- -r, --resume
resume training