britishgeologicalsurvey / corescore Goto Github PK

We currently have a single metric to assess rock quality that is implemented in fastaiv1 as foreground_acc and assesses the accuracy of the foreground (rock) segmentation. We should revisit this and implement more metrics. This can be done by having multiple metrics, e.g starting from the crude, like the total foreground accuracy and then measure the accuracy for each respective class. We should also add existing core quality assessments e.g the Rock Quality Designation (RQD) with the caveat that the segmentation has to be thoroughly tested for its performance before we have concrete results.

Libraries like talos add an automated hyperameter optimization layer. We should implement this (or any other similar libraries) and add it to the training workflow.

We've had good questions about the model's inner workings and this is an opportunity to try captum (most likely) or similar toolkit for model explainability.

https://github.com/pytorch/captum/blob/master/tutorials/Resnet_TorchVision_Ablation.ipynb - this is a promising notebook on "Inspecting influential image parts" for semantic segmentation with pytorch

The training script, train.py only accepts the number of epochs and learning rate. The default batch size is set t to 1 and as a result the training has been very slow. It will be good to be able handle more arguments and see how they affect performance.

Acceptance criteria:

• Batch size
• Weight decay
• Path

/labels API endpoint returns an image mask when POSTed a base64 encoded image wrapped in the JSON that LabelTool expects.

We should enable predicted annotations in LabelTool, and optionally extend the single-page JavaScript preview to show suggested segmentation overlaid on new images; or both, if we also want a simple interface to show the metrics assessing extent of fragmentation

From Zayad and Mark's original paper/outline we have a set of core quality metrics that assess fragmentation (could be expanded).

• https://scikit-image.org/docs/dev/api/skimage.measure.html#skimage.measure.regionprops to produce measures for core fragments as we did for labelled pores in thin sections
• API endpoint to return metrics based on estimated labels