def finalise(self,
full_res=False,
vote_miou=True,
ply_output="",
**kwargs):
per_class_iou = self._confusion_matrix.get_intersection_union_per_class(
)[0]
self._iou_per_class = {
self._dataset.INV_OBJECT_LABEL[k]: v
for k, v in enumerate(per_class_iou)
}
if vote_miou and self._test_area:
# Complete for points that have a prediction
self._test_area = self._test_area.to("cpu")
c = ConfusionMatrix(self._num_classes)
has_prediction = self._test_area.prediction_count > 0
gt = self._test_area.y[has_prediction].numpy()
pred = torch.argmax(self._test_area.votes[has_prediction],
1).numpy()
c.count_predicted_batch(gt, pred)
self._vote_miou = c.get_average_intersection_union() * 100
if full_res:
self._compute_full_miou()
if ply_output:
has_prediction = self._test_area.prediction_count > 0
self._dataset.to_ply(
self._test_area.pos[has_prediction].cpu(),
torch.argmax(self._test_area.votes[has_prediction],
1).cpu().numpy(),
ply_output,
)
def _compute_full_miou(self):
if self._full_vote_miou is not None:
return
has_prediction = self._test_area.prediction_count > 0
log.info(
"Computing full res mIoU, we have predictions for %.2f%% of the points."
% (torch.sum(has_prediction) /
(1.0 * has_prediction.shape[0]) * 100))
self._test_area = self._test_area.to("cpu")
# Full res interpolation
full_pred = knn_interpolate(
self._test_area.votes[has_prediction],
self._test_area.pos[has_prediction],
self._test_area.pos,
k=1,
)
# Full res pred
c = ConfusionMatrix(self._num_classes)
c.count_predicted_batch(self._test_area.y.numpy(),
torch.argmax(full_pred, 1).numpy())
self._full_vote_miou = c.get_average_intersection_union() * 100
def reset(self, stage="train"):
super().reset(stage=stage)
self._confusion_matrix = ConfusionMatrix(self._num_classes)
self._acc = 0
self._macc = 0
self._miou = 0
self._miou_per_class = {}
def reset(self, stage="train"):
super().reset(stage=stage)
self._full_confusion_matrix = ConfusionMatrix(self._num_classes)
self._raw_datas = {}
self._votes = {}
self._vote_counts = {}
self._full_preds = {}
self._full_acc = None
def reset(self, stage="train"):
super().reset(stage=stage)
self._confusion_matrix = ConfusionMatrix(self._num_classes)
self._ap_meter = APMeter()