conda env create -f venv.yml
conda activate histpath-project
jupyter notebook
Generating slides
datafolder - location where your .mrxs files live. image_size - default 1024 square image
Default output is /output folder for generated images and annotations.
python3 generate_slides.py <datafolder>
This creates a folder with a structure that can be viewed with fiftyone - fiftyone dataset.
view:
fiftyone app view --dataset-dir output/ --type fiftyone.types.COCODetectionDataset
This enables to look at ground truth and a check if annotations are correctly implemented.
Splitting training and testing datasets
The default implementation will find the output folder and generate a coco folder with train/test split with 80% of training images.
python3 split_coco.py
Training classifier
Training loop expects to receive training images from coco/train folder and test images from coco/test folder. Folder structure is default fiftyone dataset folder structure as was for output folder.
Default command will train 1 epoch with FastRCNNPredictor default weights and save a model in models/ directory with a name generated from the time the training finishes.
python3 train.py
Adding a dataset to predict on.
You could use another split or whole image to predict values on.
To predict on a second image generate dataset with generate_slides.py <dir_to_another_image_folder> --output_dir test_dataset/
This will generate images and annotations based of another image and save into test_dataset/ folder.
You can add dataset to fiftyone dataset lists with command
fiftyone datasets create -n mydataset -d test_dataset/ -t fiftyone.types.COCODetectionDataset
Predicting on a dataset.
To view not only ground truth, but also predictions that model makes you can add predictions into fiftyone dataset.
The weights are the weights that were saved after running train.py script and should live inside models/ directory.
Prediction label should be something clear that predictions are annotated with ex, <model_name_predictions> so you could compare multiple models as the annotations can be switched on and off based on the label.
python3 predict.py test_dataset/ mydataset <weights.pth> <prediction_label>
In the end the precision, recall, f1-score and support values are displayed for all classes.
Viewing predictions
To launch a fiftyone app in a browser:
fiftyone app launch
In upper menu you should see a select a dataset button and should be able to see a mydataset.
Labels have now two options to select from - ground_truth and your <prediction_label>
This project was done in a neural networks course in Tartu University. The time limit although didn't allow me to improve the model as most of the time went to generate a system that allowed to handle WSI images - split images into datasets, train and visualize results. The next step would be to improve the model as the model currently is just out of the box implementation. This repo however allows to modify model and visualize results in few steps and that makes playing with different solutions much faster.