angusg / deep-learning-dreissenid Goto Github PK

Remove 2 images that correspond to the same Analysis ID, and remove the ignore label=255, it should be background instead.

Run the model on a sequence of frames and generate a video.

Dominique:

With a bit of morphological operations (e.g. top-hat+watershed segmentation), I expect that individual mussels could be partitioned. (This could be one way to get the mussels size distribution deliverable.)

The previous code introduced a dependence on the batch_size

total_mussel = 0.
total_pixels = 0.
for idx, data in enumerate(val_loader):
    total_mussel += (data[1] == 1).sum().float().item()
    total_pixels += (data[1] == 0).sum().float().item()
    print('Batch %d of %d, pos_weight=%.4f' % (idx, len(val_loader), total_mussel / total_pixels))
#pos_weight = running_mean / len(val_loader)
print('pos_weight={:.4f}'.format(total_pixels / total_mussel))

Dominique

In fact, for validation I would compute contingency tables (a.k.a. confusion matrix) and note that false alarms could be in part explained by under-labelling tendencies.

Note, notebook for Tripod already created