Scripts Reference

train_mrcnn_model.py

Train a new mrcnn model starting from pretrained COCO weights

optional arguments:

-h, --help	show this help message and exit
--steps STEPS	1, 2, or 3. How many of the steps to train: (heads, 4+, entire model)
--model_dir MODEL_DIR
	Directory in which to create new training subdirectory for checkpoints and tensorboard logs.
--data_dir DATA_DIR
	Directory in which to find train and test subdirectories containing labeled images. Extract OCR top and base depths from images with pytesseract

get_ocr_depths.py

Extract OCR top and base depths from images with pytesseract

optional arguments:

-h, --help	show this help message and exit
--root_dir ROOT_DIR
	A common parent directory of all target <subdir> directories.
--subdir SUBDIR
	A string contained in the name of all target subdirectories.
--save_name SAVE_NAME
	Name of depths csv file(s) to be saved in matching subdirs.
--force	Flag to force overwrite of any existing <save_name>.csv files.
--inspect	Flag to inspect images and print OCR output whenever there is an issue.

As an example, you can test the script by running:

This should save a new file at tests/data/two_image_dataset/auto_depths_test.csv, with contents like:

,top,bottom
S00101409.jpeg,2348.0,2350.0
S00111582.jpeg,7716.0,2220.0

Note that 7716.0 is a misread, and should have been 2218.0. At least with our BGS images, some manual corrections are usually required, but this provides a template for the --depth_csv file required to run process_directory.py.

process_directory.py

Process directory of raw images with Mask R-CNN and save results as a CoreColumn.

The path given should contain images as jpeg files, and a depth_csv file in the format:

         ,    top,    bottom
<filename1>, <top1>, <bottom1>
...
<filenameN>, <topN>, <bottomN>

NOTE: model Config, class_names, and segmentation layout_params can only be changed manually at the top of script, and default to those configured in defaults.py

positional arguments:

path Path to directory of images (and depth information csv) to process.

optional arguments:

-h, --help	show help message and exit
--model_dir MODEL_DIR
	Directory to load mrcnn model from. Default=``defaults.MODEL_DIR``
--weights_path WEIGHTS_PATH
	Path to model weights to load. Default=``defaults.CB_MODEL_PATH``
--add_tol ADD_TOL
	Gap tolerance when adding CoreColumn objects, default=5.0.
--add_mode ADD_MODE
	CoreColumn.add_mode. One of {‘fill’, ‘collapse’}.
--depth_csv DEPTH_CSV
	Name of filename + (top, bottom) csv to read from path, default=``’auto_depths.csv’``
--save_dir SAVE_DIR
	Path to save CoreColumn to, default=None will save to path
--save_name SAVE_NAME
	Name to use for CoreColumn.save, default=None results in CoreColumn_<top>_<base>
--save_mode SAVE_MODE
	One of {‘pickle’, ‘numpy’}. Whether to save as single pkl file or multiple npy files

Assuming you’ve downloaded and unzipped the assets folder in the default location, you can test the script with default parameters by running:

$ cd scripts
$ python process_directory.py ../tests/data/two_image_dataset --depth_csv dummy_depths.csv

This should save the aggregated CoreColumn to tests/data/two_image_dataset/CoreColumn_1.00_5.00.pkl.

prune_imageData.py

Remove the imageData field from all JSON files in tree below path:

positional arguments:

path Path to parent of all target JSON files.

optional arguments:

-h, --help

show this help message and exit