In the following, the whole project is briefly presented in how it works and how the various scripts should be executed to reproduce our results.
Table of Contents
This project is about Image Segmentation, a machine learning task which consists in delineating specific shapes and patterns in input images. To produce trained models for further use, please referd to run_nn.sh or test_*.py; make sure to create the folders models/ and models/trained_models/ first. To see an example of segmentation, please refer to presentation_notebook.ipynb, in which all the necessary steps (if a suitably trained model is provided) are contained.
The preprocessing engine handles the input datasets selected. datasetscanner.py and datasetformatter.py take care of producing logs necessary for effisdataset.py and colombadataset.py in order for the dataloaders to work properly.
test_*.py files handle the training or fine-tuning of a U-Net over specific datasets/foundation models, if one wishes to produce results all at once, run_nn.sh will do the job. At the current state of the implementation no pre-trained models are present and will be added with suitable sources in later updates.
The Jupyter Notebook presentation_notebook.ipynb contains just a live demostration of the complete pipeline, from tiling, through predictions and eventually assembling of the final activation. (Smooth merging to be added on later releases.)
If one wishes to train, validate and test a specific model over a selected dataset (whose paths have to be appropriately configured in config/config.json), interface.py comes in handy. Please run
python3 interface.py --helpfor a set of instruction regarding the possible parameters to be used.
Francesco Pagano [email protected]
Sofia Fanigliulo [email protected]
Giuseppe Esposito [email protected]
Aurora Gensale [email protected]
Elena Di Felice [email protected]