def _test(y_pred, y, batch_size):
re.reset()
assert re._updated is False
if batch_size > 1:
n_iters = y.shape[0] // batch_size + 1
for i in range(n_iters):
idx = i * batch_size
re.update(
(y_pred[idx:idx + batch_size], y[idx:idx + batch_size]))
else:
re.update((y_pred, y))
np_y_pred = to_numpy_multilabel(y_pred)
np_y = to_numpy_multilabel(y)
assert re._type == "multilabel"
assert re._updated is True
re_compute = re.compute() if average else re.compute().mean().item()
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(np_y, np_y_pred,
average="samples") == pytest.approx(re_compute)
re1 = Recall(is_multilabel=True, average=True)
re2 = Recall(is_multilabel=True, average=False)
assert re1._updated is False
assert re2._updated is False
re1.update((y_pred, y))
re2.update((y_pred, y))
assert re1._updated is True
assert re2._updated is True
assert re1.compute() == pytest.approx(re2.compute().mean().item())
assert re1._updated is True
assert re2._updated is True
def test_compute():
recall = Recall()
y_pred = torch.eye(4)
y = torch.ones(4).type(torch.LongTensor)
recall.update((y_pred, y))
result = list(recall.compute())
assert result[0] == 0.0
assert result[1] == 0.25
assert result[2] == 0.0
assert result[3] == 0.0
recall.reset()
y_pred = torch.eye(2)
y = torch.ones(2).type(torch.LongTensor)
recall.update((y_pred, y))
y = torch.zeros(2).type(torch.LongTensor)
recall.update((y_pred, y))
result = list(recall.compute())
assert result[0] == 0.5
assert result[1] == 0.5
def test_incorrect_type():
# Tests changing of type during training
def _test(average):
re = Recall(average=average)
y_pred = torch.softmax(torch.rand(4, 4), dim=1)
y = torch.ones(4).long()
re.update((y_pred, y))
y_pred = torch.zeros(4, )
y = torch.ones(4).long()
with pytest.raises(RuntimeError):
re.update((y_pred, y))
_test(average=True)
_test(average=False)
re1 = Recall(is_multilabel=True, average=True)
re2 = Recall(is_multilabel=True, average=False)
y_pred = torch.randint(0, 2, size=(10, 4, 20, 23))
y = torch.randint(0, 2, size=(10, 4, 20, 23)).long()
re1.update((y_pred, y))
re2.update((y_pred, y))
assert re1.compute() == pytest.approx(re2.compute().mean().item())
def test_multilabel_input_NCHW():
def _test(average):
re = Recall(average=average, is_multilabel=True)
y_pred = torch.randint(0, 2, size=(10, 5, 18, 16))
y = torch.randint(0, 2, size=(10, 5, 18, 16)).long()
re.update((y_pred, y))
np_y_pred = to_numpy_multilabel(y_pred)
np_y = to_numpy_multilabel(y)
assert re._type == 'multilabel'
re_compute = re.compute() if average else re.compute().mean().item()
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(np_y, np_y_pred, average='samples') == pytest.approx(re_compute)
re.reset()
y_pred = torch.randint(0, 2, size=(10, 4, 20, 23))
y = torch.randint(0, 2, size=(10, 4, 20, 23)).long()
re.update((y_pred, y))
np_y_pred = to_numpy_multilabel(y_pred)
np_y = to_numpy_multilabel(y)
assert re._type == 'multilabel'
re_compute = re.compute() if average else re.compute().mean().item()
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(np_y, np_y_pred, average='samples') == pytest.approx(re_compute)
# Batched Updates
re.reset()
y_pred = torch.randint(0, 2, size=(100, 5, 12, 14))
y = torch.randint(0, 2, size=(100, 5, 12, 14)).long()
batch_size = 16
n_iters = y.shape[0] // batch_size + 1
for i in range(n_iters):
idx = i * batch_size
re.update((y_pred[idx:idx + batch_size], y[idx:idx + batch_size]))
np_y = to_numpy_multilabel(y)
np_y_pred = to_numpy_multilabel(y_pred)
assert re._type == 'multilabel'
re_compute = re.compute() if average else re.compute().mean().item()
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(np_y, np_y_pred, average='samples') == pytest.approx(re_compute)
for _ in range(5):
_test(average=True)
_test(average=False)
re1 = Recall(is_multilabel=True, average=True)
re2 = Recall(is_multilabel=True, average=False)
y_pred = torch.randint(0, 2, size=(10, 4, 20, 23))
y = torch.randint(0, 2, size=(10, 4, 20, 23)).long()
re1.update((y_pred, y))
re2.update((y_pred, y))
assert re1.compute() == pytest.approx(re2.compute().mean().item())
def test_no_update():
recall = Recall()
with pytest.raises(NotComputableError):
recall.compute()
recall = Recall(is_multilabel=True, average=True)
with pytest.raises(NotComputableError):
recall.compute()
def test_compute_average():
recall = Recall(average=True)
y_pred = torch.eye(4)
y = torch.ones(4).type(torch.LongTensor)
recall.update((y_pred, y))
assert isinstance(recall.compute(), float)
assert recall.compute() == 0.0625
def test_no_update():
recall = Recall()
with pytest.raises(
NotComputableError,
match=
r"Recall must have at least one example before it can be computed"
):
recall.compute()
recall = Recall(is_multilabel=True, average=True)
with pytest.raises(NotComputableError):
recall.compute()
def _test(average):
re = Recall(average=average)
y_pred = torch.rand(10, 5, 18, 16)
y = torch.randint(0, 5, size=(10, 18, 16)).long()
re.update((y_pred, y))
num_classes = y_pred.shape[1]
np_y_pred = y_pred.argmax(dim=1).numpy().ravel()
np_y = y.numpy().ravel()
assert re._type == "multiclass"
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
sk_average_parameter = "macro" if average else None
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
sk_compute = recall_score(np_y, np_y_pred, labels=range(0, num_classes), average=sk_average_parameter)
assert sk_compute == pytest.approx(re_compute)
re.reset()
y_pred = torch.rand(10, 7, 20, 12)
y = torch.randint(0, 7, size=(10, 20, 12)).long()
re.update((y_pred, y))
num_classes = y_pred.shape[1]
np_y_pred = y_pred.argmax(dim=1).numpy().ravel()
np_y = y.numpy().ravel()
assert re._type == "multiclass"
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
sk_average_parameter = "macro" if average else None
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
sk_compute = recall_score(np_y, np_y_pred, labels=range(0, num_classes), average=sk_average_parameter)
assert sk_compute == pytest.approx(re_compute)
# Batched Updates
re.reset()
y_pred = torch.rand(100, 10, 12, 14)
y = torch.randint(0, 10, size=(100, 12, 14)).long()
batch_size = 16
n_iters = y.shape[0] // batch_size + 1
for i in range(n_iters):
idx = i * batch_size
re.update((y_pred[idx : idx + batch_size], y[idx : idx + batch_size]))
num_classes = y_pred.shape[1]
np_y = y.numpy().ravel()
np_y_pred = y_pred.argmax(dim=1).numpy().ravel()
assert re._type == "multiclass"
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
sk_average_parameter = "macro" if average else None
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
sk_compute = recall_score(np_y, np_y_pred, labels=range(0, num_classes), average=sk_average_parameter)
assert sk_compute == pytest.approx(re_compute)
def _test(average):
re = Recall(average=average)
y_pred = torch.rand(20, 6)
y = torch.randint(0, 5, size=(20, )).type(torch.LongTensor)
re.update((y_pred, y))
np_y_pred = y_pred.numpy().argmax(axis=1).ravel()
np_y = y.numpy().ravel()
assert re._type == 'multiclass'
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
sklearn_average_parameter = 'macro' if average else None
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(
np_y, np_y_pred,
average=sklearn_average_parameter) == pytest.approx(re_compute)
re.reset()
y_pred = torch.rand(10, 4)
y = torch.randint(0, 3, size=(10, 1)).type(torch.LongTensor)
re.update((y_pred, y))
np_y_pred = y_pred.numpy().argmax(axis=1).ravel()
np_y = y.numpy().ravel()
assert re._type == 'multiclass'
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
sklearn_average_parameter = 'macro' if average else None
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(
np_y, np_y_pred,
average=sklearn_average_parameter) == pytest.approx(re_compute)
# 2-classes
re.reset()
y_pred = torch.rand(10, 2)
y = torch.randint(0, 2, size=(10, 1)).type(torch.LongTensor)
re.update((y_pred, y))
np_y_pred = y_pred.numpy().argmax(axis=1).ravel()
np_y = y.numpy().ravel()
assert re._type == 'multiclass'
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
sklearn_average_parameter = 'macro' if average else None
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(
np_y, np_y_pred,
average=sklearn_average_parameter) == pytest.approx(re_compute)
def test_binary_shapes():
recall = Recall(average=True)
y = torch.LongTensor([1, 0])
y_pred = torch.FloatTensor([0.9, 0.2])
y_pred = y_pred.unsqueeze(1)
indices = torch.max(torch.cat([1.0 - y_pred, y_pred], dim=1), dim=1)[1]
recall.update((y_pred, y))
assert recall.compute() == pytest.approx(
recall_score(y.data.numpy(), indices.data.numpy(), average='macro'))
assert recall.compute() == 1.0
y = torch.LongTensor([[1], [0]])
y_pred = torch.FloatTensor([[0.9], [0.2]])
indices = torch.max(torch.cat([1.0 - y_pred, y_pred], dim=1), dim=1)[1]
recall.reset()
recall.update((y_pred, y))
assert recall.compute() == pytest.approx(
recall_score(y.data.numpy(), indices.data.numpy(), average='macro'))
assert recall.compute() == 1.0
def test_compute_all_wrong():
recall = Recall()
y_pred = torch.FloatTensor([[1.0, 0.0], [1.0, 0.0]])
y = torch.ones(2).type(torch.LongTensor)
recall.update((y_pred, y))
result = list(recall.compute())
assert result[0] == 0.0
assert result[1] == 0.0
def _test(average):
re = Recall(average=average)
y_pred = torch.randint(0, 2, size=(10, 12, 10))
y = torch.randint(0, 2, size=(10, 12, 10)).type(torch.LongTensor)
re.update((y_pred, y))
np_y = y.numpy().ravel()
np_y_pred = y_pred.numpy().ravel()
assert re._type == 'binary'
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
assert recall_score(np_y, np_y_pred,
average='binary') == pytest.approx(re_compute)
re.reset()
y_pred = torch.randint(0, 2, size=(10, 1, 12, 10))
y = torch.randint(0, 2, size=(10, 1, 12, 10)).type(torch.LongTensor)
re.update((y_pred, y))
np_y = y.numpy().ravel()
np_y_pred = y_pred.numpy().ravel()
assert re._type == 'binary'
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
assert recall_score(np_y, np_y_pred,
average='binary') == pytest.approx(re_compute)
re = Recall(average=average)
# Batched Updates
re.reset()
y_pred = torch.randint(0, 2, size=(100, 12, 10))
y = torch.randint(0, 2, size=(100, 1, 12, 10)).type(torch.LongTensor)
batch_size = 16
n_iters = y.shape[0] // batch_size + 1
for i in range(n_iters):
idx = i * batch_size
re.update((y_pred[idx:idx + batch_size], y[idx:idx + batch_size]))
np_y = y.numpy().ravel()
np_y_pred = y_pred.numpy().ravel()
assert re._type == 'binary'
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
assert recall_score(np_y, np_y_pred,
average='binary') == pytest.approx(re_compute)
def _test(average):
re = Recall(average=average)
# TODO: y_pred should be binary after 0.1.2 release
# y_pred = torch.randint(0, 2, size=(10, 12, 10)).type(torch.LongTensor)
y_pred = torch.rand(10, 12, 10)
y = torch.randint(0, 2, size=(10, 12, 10)).type(torch.LongTensor)
re.update((y_pred, y))
np_y = y.numpy().ravel()
# np_y_pred = y_pred.numpy().ravel()
np_y_pred = (y_pred.numpy().ravel() > 0.5).astype('int')
assert re._type == 'binary'
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
assert recall_score(np_y, np_y_pred,
average='binary') == pytest.approx(re_compute)
re.reset()
# TODO: y_pred should be binary after 0.1.2 release
# y_pred = torch.randint(0, 2, size=(10, 1, 12, 10)).type(torch.LongTensor)
y_pred = torch.rand(10, 1, 12, 10)
y = torch.randint(0, 2, size=(10, 1, 12, 10)).type(torch.LongTensor)
re.update((y_pred, y))
np_y = y.numpy().ravel()
# np_y_pred = y_pred.numpy().ravel()
np_y_pred = (y_pred.numpy().ravel() > 0.5).astype('int')
assert re._type == 'binary'
assert isinstance(re.compute(), float if average else torch.Tensor)
re_compute = re.compute() if average else re.compute().numpy()
assert recall_score(np_y, np_y_pred,
average='binary') == pytest.approx(re_compute)
def test_predict(model,dataloader_test,use_cuda):
if use_cuda:
model = model.cuda()
precision = Precision()
recall = Recall()
f1 = Fbeta(beta=1.0, average=True, precision=precision, recall=recall)
for i,(img, label) in enumerate(dataloader_test):
img, labels = Variable(img),Variable(label)
if use_cuda:
img = img.cuda()
label = label.cuda()
pred = model(img)
_,my_label = torch.max(label, dim=1)
precision.update((pred, my_label))
recall.update((pred, my_label))
f1.update((pred, my_label))
precision.compute()
recall.compute()
print("\tF1 Score: {:0.2f}".format(f1.compute()*100))
class FbetaScore(Metric):
def __init__(
self,
beta: int = 1,
output_transform: Callable = lambda x: x,
average: str = "macro",
is_multilabel: bool = False,
device: Optional[Union[str, torch.device]] = None,
):
self._beta = beta
self._average = average
_average_flag = self._average != "macro"
self._precision = Precision(
output_transform=output_transform,
average=_average_flag,
is_multilabel=is_multilabel,
device=device,
)
self._recall = Recall(
output_transform=output_transform,
average=_average_flag,
is_multilabel=is_multilabel,
device=device,
)
super(FbetaScore, self).__init__(
output_transform=output_transform, device=device
)
@reinit__is_reduced
def reset(self) -> None:
self._precision.reset()
self._recall.reset()
def compute(self) -> torch.Tensor:
precision_val = self._precision.compute()
recall_val = self._recall.compute()
fbeta_val = (
(1.0 + self._beta ** 2)
* precision_val
* recall_val
/ (self._beta ** 2 * precision_val + recall_val + 1e-15)
)
if self._average == "macro":
fbeta_val = torch.mean(fbeta_val).item()
return fbeta_val
@reinit__is_reduced
def update(self, output: Sequence[torch.Tensor]) -> None:
self._precision.update(output)
self._recall.update(output)
def _test(average, n_epochs, metric_device):
n_iters = 60
s = 16
n_classes = 7
offset = n_iters * s
y_true = torch.randint(0,
2,
size=(offset * idist.get_world_size(),
n_classes, 6, 8)).to(device)
y_preds = torch.randint(0,
2,
size=(offset * idist.get_world_size(),
n_classes, 6, 8)).to(device)
def update(engine, i):
return (
y_preds[i * s + rank * offset:(i + 1) * s + rank * offset,
...],
y_true[i * s + rank * offset:(i + 1) * s + rank * offset, ...],
)
engine = Engine(update)
re = Recall(average=average, is_multilabel=True, device=metric_device)
re.attach(engine, "re")
assert re._updated is False
data = list(range(n_iters))
engine.run(data=data, max_epochs=n_epochs)
assert "re" in engine.state.metrics
assert re._updated is True
res = engine.state.metrics["re"]
res2 = re.compute()
if isinstance(res, torch.Tensor):
res = res.cpu().numpy()
res2 = res2.cpu().numpy()
assert (res == res2).all()
else:
assert res == res2
np_y_preds = to_numpy_multilabel(y_preds)
np_y_true = to_numpy_multilabel(y_true)
assert re._type == "multilabel"
res = res if average else res.mean().item()
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(np_y_true, np_y_preds,
average="samples") == pytest.approx(res)
def evaluate_epoch(eval_dl, model, criterion, epoch, writer):
""" evaluation in a epoch
Args:
eval_dl (DataLoader): DataLoader of validation set
model (nn.Module): model in PyTorch
criterion (loss): PyTorch loss
epoch (int): epoch number
writer (SummaryWriter): instance of SummaryWriter for TensorBoard
Returns:
"""
print('\neval epoch {}'.format(epoch))
device = next(model.parameters()).device
model.eval()
recall = Recall(lambda x: (x[0], x[1]))
precision = Precision(lambda x: (x[0], x[1]))
mean_recall = []
mean_precision = []
mean_loss = []
with torch.no_grad():
for idx, (inputs, targets) in enumerate(eval_dl):
inputs, targets = inputs.to(device), targets.to(device)
outputs = model(inputs)
loss = criterion(outputs, targets)
preds = outputs.argmax(1)
precision.update((preds, targets))
recall.update((preds, targets))
mean_loss.append(loss.item())
mean_recall.append(recall.compute().item())
mean_precision.append(precision.compute().item())
# print('val-epoch:{} [{}/{}], loss: {:5.3}'.format(epoch, idx + 1, len(dataloader), loss.item()))
writer.add_scalar('test/loss', loss.item(), len(eval_dl) * epoch + idx)
mean_precision, mean_recall = np.array(mean_precision).mean(), np.array(mean_recall).mean()
f1 = mean_precision * mean_recall * 2 / (mean_precision + mean_recall + 1e-20)
print('precision: {:07.5}, recall: {:07.5}, f1: {:07.5}\n'.format(mean_precision, mean_recall, f1))
writer.add_scalar('test/epoch-loss', np.array(mean_loss).mean(), epoch)
writer.add_scalar('test/f1', f1, epoch)
writer.add_scalar('test/precision', mean_precision, epoch)
writer.add_scalar('test/recall', mean_recall, epoch)
def test_ner_example():
recall = Recall()
y = torch.Tensor([[0, 1, 1, 1, 1, 1, 1, 1], [2, 2, 2, 2, 2, 2, 2,
2]]).type(torch.LongTensor)
y_pred = torch.softmax(torch.rand(2, 3, 8), dim=1)
indices = torch.max(y_pred, dim=1)[1]
y_pred_labels = list(set(indices.view(-1).tolist()))
recall_sk = recall_score(y.view(-1).data.numpy(),
indices.view(-1).data.numpy(),
labels=y_pred_labels,
average=None)
recall.update((y_pred, y))
recall_ig = recall.compute().tolist()
recall_ig = [recall_ig[i] for i in y_pred_labels]
assert all([a == pytest.approx(b) for a, b in zip(recall_sk, recall_ig)])
def evalidation(epoch, dataloader, model, criterion, device, writer,
tb_test_imgs):
print('\neval epoch {}'.format(epoch))
model.eval()
recall = Recall(lambda x: (x[0], x[1]))
precision = Precision(lambda x: (x[0], x[1]))
mean_recall = []
mean_precision = []
mean_loss = []
with torch.no_grad():
for idx, (pre_img, post_img, targets) in enumerate(dataloader):
pre_img, post_img, targets = pre_img.to(device), post_img.to(
device), targets.to(device)
outputs = model(pre_img, post_img)
loss = criterion(outputs, targets)
preds = outputs.argmax(1)
precision.update((preds, targets))
recall.update((preds, targets))
mean_loss.append(loss.item())
mean_recall.append(recall.compute().item())
mean_precision.append(precision.compute().item())
# print('val-epoch:{} [{}/{}], loss: {:5.3}'.format(epoch, idx + 1, len(dataloader), loss.item()))
writer.add_scalar('test/loss', loss.item(),
len(dataloader) * epoch + idx)
if idx < tb_test_imgs:
writer.add_image('test/pre', pre_img[0], idx)
writer.add_image('test/post', post_img[0], idx)
writer.add_image('test/label', label[0], idx)
writer.add_image('test/pred', preds, idx)
mean_precision, mean_recall = np.array(mean_precision).mean(), np.array(
mean_recall).mean()
f1 = mean_precision * mean_recall * 2 / (mean_precision + mean_recall +
1e-20)
print('precision: {:07.5}, recall: {:07.5}, f1: {:07.5}\n'.format(
mean_precision, mean_recall, f1))
writer.add_scalar('test/epoch-loss', np.array(mean_loss).mean(), epoch)
writer.add_scalar('test/f1', f1, epoch)
writer.add_scalar('test/precision', mean_precision, epoch)
writer.add_scalar('test/recall', mean_recall, epoch)
def test_sklearn_compute():
recall = Recall(average=False)
y = torch.Tensor(range(5)).type(torch.LongTensor)
y_pred = torch.softmax(torch.rand(5, 5), dim=1)
indices = torch.max(y_pred, dim=1)[1]
recall.update((y_pred, y))
y_pred_labels = list(set(indices.tolist()))
recall_sk = recall_score(y.data.numpy(),
indices.data.numpy(),
labels=y_pred_labels,
average=None)
recall_ig = recall.compute().tolist()
recall_ig = [recall_ig[i] for i in y_pred_labels]
assert all([a == pytest.approx(b) for a, b in zip(recall_sk, recall_ig)])
def _test(average):
re = Recall(average=average, is_multilabel=True)
y_pred = torch.randint(0, 2, size=(10, 5, 10))
y = torch.randint(0, 2, size=(10, 5, 10)).long()
re.update((y_pred, y))
np_y_pred = to_numpy_multilabel(y_pred)
np_y = to_numpy_multilabel(y)
assert re._type == "multilabel"
re_compute = re.compute() if average else re.compute().mean().item()
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(np_y, np_y_pred,
average="samples") == pytest.approx(re_compute)
re.reset()
y_pred = torch.randint(0, 2, size=(15, 4, 10))
y = torch.randint(0, 2, size=(15, 4, 10)).long()
re.update((y_pred, y))
np_y_pred = to_numpy_multilabel(y_pred)
np_y = to_numpy_multilabel(y)
assert re._type == "multilabel"
re_compute = re.compute() if average else re.compute().mean().item()
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(np_y, np_y_pred,
average="samples") == pytest.approx(re_compute)
# Batched Updates
re.reset()
y_pred = torch.randint(0, 2, size=(100, 4, 12))
y = torch.randint(0, 2, size=(100, 4, 12)).long()
batch_size = 16
n_iters = y.shape[0] // batch_size + 1
for i in range(n_iters):
idx = i * batch_size
re.update((y_pred[idx:idx + batch_size], y[idx:idx + batch_size]))
np_y = to_numpy_multilabel(y)
np_y_pred = to_numpy_multilabel(y_pred)
assert re._type == "multilabel"
re_compute = re.compute() if average else re.compute().mean().item()
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(np_y, np_y_pred,
average="samples") == pytest.approx(re_compute)
def _test(average, n_epochs):
n_iters = 60
s = 16
n_classes = 7
offset = n_iters * s
y_true = torch.randint(0, 2, size=(offset * dist.get_world_size(), n_classes, 6, 8)).to(device)
y_preds = torch.randint(0, 2, size=(offset * dist.get_world_size(), n_classes, 6, 8)).to(device)
def update(engine, i):
return y_preds[i * s + rank * offset:(i + 1) * s + rank * offset, ...], \
y_true[i * s + rank * offset:(i + 1) * s + rank * offset, ...]
engine = Engine(update)
re = Recall(average=average, is_multilabel=True, device=device)
re.attach(engine, "re")
data = list(range(n_iters))
engine.run(data=data, max_epochs=n_epochs)
assert "re" in engine.state.metrics
res = engine.state.metrics['re']
res2 = re.compute()
if isinstance(res, torch.Tensor):
res = res.cpu().numpy()
res2 = res2.cpu().numpy()
assert (res == res2).all()
else:
assert res == res2
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
true_res = recall_score(to_numpy_multilabel(y_true),
to_numpy_multilabel(y_preds),
average='samples' if average else None)
assert pytest.approx(res) == true_res
def _test_distrib_integration_multilabel(device):
from ignite.engine import Engine
rank = idist.get_rank()
torch.manual_seed(12)
def _test(average, n_epochs, metric_device):
n_iters = 60
s = 16
n_classes = 7
offset = n_iters * s
y_true = torch.randint(0,
2,
size=(offset * idist.get_world_size(),
n_classes, 6, 8)).to(device)
y_preds = torch.randint(0,
2,
size=(offset * idist.get_world_size(),
n_classes, 6, 8)).to(device)
def update(engine, i):
return (
y_preds[i * s + rank * offset:(i + 1) * s + rank * offset,
...],
y_true[i * s + rank * offset:(i + 1) * s + rank * offset, ...],
)
engine = Engine(update)
re = Recall(average=average, is_multilabel=True, device=metric_device)
re.attach(engine, "re")
data = list(range(n_iters))
engine.run(data=data, max_epochs=n_epochs)
assert "re" in engine.state.metrics
res = engine.state.metrics["re"]
res2 = re.compute()
if isinstance(res, torch.Tensor):
res = res.cpu().numpy()
res2 = res2.cpu().numpy()
assert (res == res2).all()
else:
assert res == res2
np_y_preds = to_numpy_multilabel(y_preds)
np_y_true = to_numpy_multilabel(y_true)
assert re._type == "multilabel"
res = res if average else res.mean().item()
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
assert recall_score(np_y_true, np_y_preds,
average="samples") == pytest.approx(res)
metric_devices = ["cpu"]
if device.type != "xla":
metric_devices.append(idist.device())
for _ in range(2):
for metric_device in metric_devices:
_test(average=True, n_epochs=1, metric_device=metric_device)
_test(average=True, n_epochs=2, metric_device=metric_device)
_test(average=False, n_epochs=1, metric_device=metric_device)
_test(average=False, n_epochs=2, metric_device=metric_device)
re1 = Recall(is_multilabel=True, average=True)
re2 = Recall(is_multilabel=True, average=False)
y_pred = torch.randint(0, 2, size=(10, 4, 20, 23))
y = torch.randint(0, 2, size=(10, 4, 20, 23)).long()
re1.update((y_pred, y))
re2.update((y_pred, y))
assert re1.compute() == pytest.approx(re2.compute().mean().item())
def train_predict(dataloader_train,dataloader_val,model,epochs,learning_rate,use_cuda):
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
if use_cuda:
model = model.cuda()
model = model.train()
start.record()
train_loss_list=[]
val_loss_list=[]
train_f1=[]
val_f1=[]
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
precision = Precision()
recall = Recall()
f1 = Fbeta(beta=1.0, average=True, precision=precision, recall=recall)
for epoch in range(epochs):
print("Epoch: {}".format(epoch+1))
for i,(img, label) in enumerate(dataloader_train):
img, label = Variable(img),Variable(label)
if use_cuda:
img = img.cuda()
label = label.cuda()
optimizer.zero_grad()
pred = model.forward(img)
_,my_label = torch.max(label, dim=1)
loss = loss_fn(pred,my_label)
if i == len(dataloader_train)-1:
train_loss_list.append(loss.item())
loss.backward()
optimizer.step()
precision.update((pred, my_label))
recall.update((pred, my_label))
f1.update((pred, my_label))
print("\tTrain loss: {:0.2f}".format(train_loss_list[-1]))
precision.compute()
recall.compute()
train_f1.append(f1.compute()*100)
print("\tTrain F1 Score: {:0.2f}%".format(train_f1[-1]))
precision = Precision()
recall = Recall()
f1 = Fbeta(beta=1.0, average=True, precision=precision, recall=recall)
with torch.no_grad():
for i,(img, label) in enumerate(dataloader_val):
img, labels = Variable(img),Variable(label)
if use_cuda:
img = img.cuda()
label = label.cuda()
pred = model(img)
_,my_label = torch.max(label, dim=1)
loss = loss_fn(pred,my_label)
if i == len(dataloader_val)-1:
val_loss_list.append(loss.item())
precision.update((pred, my_label))
recall.update((pred, my_label))
f1.update((pred, my_label))
print("\n\tVal loss: {:0.2f}".format(val_loss_list[-1]))
precision.compute()
recall.compute()
val_f1.append(f1.compute()*100)
print("\tVal F1 Score: {:0.2f}%".format(val_f1[-1]))
end.record()
torch.cuda.synchronize()
time = start.elapsed_time(end)
return (train_loss_list,val_loss_list,train_f1,val_f1,time,model)
def _test_distrib_itegration_multilabel(device):
import torch.distributed as dist
from ignite.engine import Engine
rank = dist.get_rank()
torch.manual_seed(12)
def _test(average, n_epochs):
n_iters = 60
s = 16
n_classes = 7
offset = n_iters * s
y_true = torch.randint(0,
2,
size=(offset * dist.get_world_size(), n_classes,
6, 8)).to(device)
y_preds = torch.randint(0,
2,
size=(offset * dist.get_world_size(),
n_classes, 6, 8)).to(device)
def update(engine, i):
return (
y_preds[i * s + rank * offset:(i + 1) * s + rank * offset,
...],
y_true[i * s + rank * offset:(i + 1) * s + rank * offset, ...],
)
engine = Engine(update)
re = Recall(average=average, is_multilabel=True, device=device)
re.attach(engine, "re")
data = list(range(n_iters))
engine.run(data=data, max_epochs=n_epochs)
assert "re" in engine.state.metrics
res = engine.state.metrics["re"]
res2 = re.compute()
if isinstance(res, torch.Tensor):
res = res.cpu().numpy()
res2 = res2.cpu().numpy()
assert (res == res2).all()
else:
assert res == res2
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UndefinedMetricWarning)
true_res = recall_score(to_numpy_multilabel(y_true),
to_numpy_multilabel(y_preds),
average="samples" if average else None)
assert pytest.approx(res) == true_res
for _ in range(2):
_test(average=True, n_epochs=1)
_test(average=True, n_epochs=2)
with pytest.warns(
RuntimeWarning,
match=
"Precision/Recall metrics do not work in distributed setting when "
"average=False and is_multilabel=True",
):
re = Recall(average=False, is_multilabel=True, device=device)
y_pred = torch.randint(0, 2, size=(4, 3, 6, 8))
y = torch.randint(0, 2, size=(4, 3, 6, 8)).long()
re.update((y_pred, y))
re_compute1 = re.compute()
re_compute2 = re.compute()
assert len(re_compute1) == 4 * 6 * 8
assert (re_compute1 == re_compute2).all()
def test_no_update():
recall = Recall()
with pytest.raises(NotComputableError):
recall.compute()