def add_train_metrics(in_dict, result_dict):
train_probs = result_dict.pop('train_probs')
preds = train_probs > 0.5
if not np.any(preds):
print ">>> Predicted 0 pixels."
train_target = in_dict['y_full']
result_dict["train_iou"] = iou(preds, train_target)
def add_validate_metrics(in_dict, result_dict):
global best_valid_iou
global valid_iter
valid_iter += 1
valid_out = valid_fn(valid_chunk)
valid_out['window_max'] = valid_out['window_max'][0]
valid_out['window_min'] = valid_out['window_min'][0]
probs = valid_out.pop("valid_probs")
preds = probs > 0.5
result_dict.update(valid_out)
valid_target = valid_chunk['y_full']
this_iou = iou(preds, valid_target)
result_dict["valid_iou"] = this_iou
# Save an example image of the current best segmentation
# Let's also save one as a snapshot
if this_iou > best_valid_iou:
best_valid_iou = this_iou
store_image(
valid_chunk['x_full'][0, 0],
valid_chunk['y_full'][0, 0],
preds[0, 0],
filepath=trial.file_path('segmented_best.jpg'),
window_min=valid_out['window_min'],
window_max=valid_out['window_max'],
)
store_image(
valid_chunk['x_full'][0, 0],
valid_chunk['y_full'][0, 0],
preds[0, 0],
filepath=trial.file_path('segmentations/segmented_%04d.jpg' % valid_iter),
window_min=valid_out['window_min'],
window_max=valid_out['window_max'],
)
