def on_epoch_end(self, last_metrics, **kwargs):
idx_pairs = [(8, 9), (12, 13), (11, 14), (10, 15), (2, 3), (1, 4), (0, 5)]
accuracy = self.correct / self.total
pckh = [(accuracy[:, idx0] + accuracy[:, idx1]) / 2
for idx0, idx1
in idx_pairs]
# add upper body and total
pckh.extend([
self.correct[:, 8:].sum(dim=1) / self.total[:, 8:].sum(dim=1),
self.correct[:, self.all_idx].sum(dim=1) / self.total[:, self.all_idx].sum(dim=1)
])
# add multi-label classification accuracy, TP-accuracy, FN-accuracy
if self.acc_thresh:
pckh.extend([
self.mlc_correct / self.mlc_total if self.mlc_total else np.nan,
accuracy[16],
accuracy[17]
])
if self.niter > 1:
results = pckh[-1].tolist()
else:
results = [r.item() for r in pckh]
return add_metrics(last_metrics, results)
def on_epoch_end(self, last_metrics, **kwargs):
matrix = self.cm.matrix
objs = list(range(1, len(matrix)))
pa = (matrix[objs, objs].sum() / matrix[1:].sum())
gt_count = matrix[1:].sum(axis=1)
pred_count = matrix[1:, 1:].sum(axis=0)
intersection = matrix[objs, objs]
iou = intersection / (gt_count + pred_count - intersection)
miou = iou.mean()
fp = matrix[1:, self.obj].sum() - matrix[self.obj, self.obj]
fn = matrix[self.obj, :].sum() - matrix[self.obj, self.obj]
return add_metrics(last_metrics, [pa, miou, fp/fn])
def on_epoch_end(self, last_metrics, **kwargs):
results = self.metrics.avg()[:2]
results.extend(x.accuracy() for x in (self.overlap, self.no_class,
self.precision, self.recall))
return fv.add_metrics(last_metrics, results)
def on_epoch_end(self, last_metrics, **kwargs):
results = self.metrics.avg()
metrics = fv.add_metrics(last_metrics, results)
return metrics
def on_epoch_end(self, last_metrics, **kwargs):
return add_metrics(last_metrics, self.total_time / 60)