def test_compute():
acc = TopKCategoricalAccuracy(2)
y_pred = torch.FloatTensor([[0.2, 0.4, 0.6, 0.8], [0.8, 0.6, 0.4, 0.2]])
y = torch.ones(2).type(torch.LongTensor)
acc.update((y_pred, y))
assert acc.compute() == 0.5
acc.reset()
y_pred = torch.FloatTensor([[0.4, 0.8, 0.2, 0.6], [0.8, 0.6, 0.4, 0.2]])
y = torch.ones(2).type(torch.LongTensor)
acc.update((y_pred, y))
assert acc.compute() == 1.0
def test_compute():
acc = TopKCategoricalAccuracy(2)
y_pred = torch.FloatTensor([[0.2, 0.4, 0.6, 0.8], [0.8, 0.6, 0.4, 0.2]])
y = torch.ones(2).long()
acc.update((y_pred, y))
assert isinstance(acc.compute(), float)
assert acc.compute() == 0.5
acc.reset()
y_pred = torch.FloatTensor([[0.4, 0.8, 0.2, 0.6], [0.8, 0.6, 0.4, 0.2]])
y = torch.ones(2).long()
acc.update((y_pred, y))
assert isinstance(acc.compute(), float)
assert acc.compute() == 1.0
def _test_distrib_accumulator_device(device):
metric_devices = [torch.device("cpu")]
if device.type != "xla":
metric_devices.append(idist.device())
for metric_device in metric_devices:
acc = TopKCategoricalAccuracy(2, device=metric_device)
assert acc._device == metric_device
assert acc._num_correct.device == metric_device, "{}:{} vs {}:{}".format(
type(acc._num_correct.device), acc._num_correct.device,
type(metric_device), metric_device)
y_pred = torch.tensor([[0.2, 0.4, 0.6, 0.8], [0.8, 0.6, 0.4, 0.2]])
y = torch.ones(2).long()
acc.update((y_pred, y))
assert acc._num_correct.device == metric_device, "{}:{} vs {}:{}".format(
type(acc._num_correct.device), acc._num_correct.device,
type(metric_device), metric_device)
def evaluate(net, test_dataloader):
with torch.no_grad():
net.eval()
preds_all = torch.empty((len(test_dataloader), 256))
top_1 = TopKCategoricalAccuracy(k=1)
top_5 = TopKCategoricalAccuracy()
top_10 = TopKCategoricalAccuracy(k=10)
for i, data in enumerate(test_dataloader):
lidar, beams = data
lidar = lidar.cuda()
beams = beams.cuda()
preds = net(lidar)
preds = F.softmax(preds, dim=1)
preds_all[i, :] = preds
top_1.update((preds, torch.argmax(beams)))
top_5.update((preds, torch.argmax(beams)))
top_10.update((preds, torch.argmax(beams)))
net.train()
print("Top-1: {:.4f} Top-5: {:.4f} Top-10: {:.4f}".format(
top_1.compute(), top_5.compute(), top_10.compute()))
return preds_all
def test_topk_accuracy(self, k: int, y_pred: Tensor, y_true: Tensor,
score: float):
accuracy = TopKCategoricalAccuracy(k=k)
accuracy.update((y_pred, y_true))
self.assertEqual(score, accuracy.compute())