Segmentation dataset

Make sure to checkout with —recurse-submodules, there’s a submodule ‘tools’ for common bits and pieces)
It loads a folder full of images and masks (pixel value corresponds to class) with a config file.

An example dataset to train on can be found at:
https://drive.google.com/file/d/0B_mOCEqr7usZa2hra0xZUTJscE0/view

Or classes/images from the COCO/Pascal VOC dataset can be imported from scripts in the import/ folder.

View the training or testing set and mask annotations:

python -m dataset.view --input /path/to/dataset --train (or --test)

Useful to check the preprocessing of images.

View a mask file

python view_labels.py some/file.jpg.mask

Train a model:

python main.py --lr 0.1 --batch_size 4 --input /path/to/dataset --model "unet --depth 5" --epoch_size 1024

Common options:

--load, load from a previous checkpoint and cointunue training
--model, specify model and model parameters (use quotes)
--show, show results of evaluating the model in training (sanity check)

Evaluate a model on new image(s):

python test.py --batch /path/to/images --model log/model.pth --save results_path (and/or --show)
python test.py --image /my/image.jpg --model log/model.pth --show