def test_accuracy():
output = torch.rand((16, 4))
output_np = output.numpy()
target = torch.randint(0, 4, (16, ))
target_np = target.numpy()
expected = 100 * accuracy_score(target_np, np.argmax(output_np, 1))
result = Accuracy()(output, target).item()
assert np.allclose(expected, result)
# check that it also works in case of extra dim for target
result2 = Accuracy()(output, target.view(16, 1)).item()
assert np.allclose(expected, result2)
def test_accuracy_image():
output = torch.rand((16, 4, 20, 20))
target = torch.randint(0, 2, (16, 1, 20, 20))
expected = 100 * accuracy_score(target.flatten().numpy(),
output.argmax(1).flatten())
result = Accuracy()(output, target).item()
assert np.allclose(expected, result)
def test_accuracy():
output = torch.rand((16, 4))
output_np = output.numpy()
target = torch.randint(0, 4, (16, ))
target_np = target.numpy()
expected = 100 * accuracy_score(target_np, np.argmax(output_np, 1))
result = Accuracy()(output, target).flatten().numpy()
assert np.allclose(expected, result)
def test_binary_accuracy():
output = torch.rand((16, 1))
output_np = output.numpy()
target = torch.randint(0, 2, (16, ))
target_np = target.numpy()
expected = 100 * accuracy_score(target_np, output_np > 0)
result = Accuracy()(output, target).item()
assert np.allclose(expected, result)
def test_binary_accuracy_image():
output = torch.rand((16, 1, 20, 20)) - 0.5
output_np = output.numpy()
target = torch.randint(0, 2, (16, 1, 20, 20))
target_np = target.numpy()
expected = 100 * accuracy_score(target_np.flatten(),
output_np.flatten() > 0)
result = Accuracy()(output, target).item()
assert np.allclose(expected, result)
def test_accuracy_topk():
# fmt: off
output = torch.tensor([
[
10,
5,
-10,
],
[
10,
-10,
5,
],
[
5,
10,
-10,
],
[
-10,
5,
10,
],
[
-10,
5,
10,
],
])
# fmt: on
target = torch.tensor([1, 1, 1, 1, 1])
result = Accuracy(2)(output, target).item()
assert np.allclose(result, 4 / 5 * 100)
target2 = torch.tensor([2, 2, 2, 2, 2])
result2 = Accuracy(2)(output, target2).item()
assert np.allclose(result2, 3 / 5 * 100)
return self
def __len__(self):
return LOADER_LEN
def __next__(self):
img = torch.randn(BS, 3, IMG_SHAPE, IMG_SHAPE)
target = torch.randint(NUM_CLASSES, (BS, ))
return img.cuda(), target.cuda()
TestLoader = Loader()
TestModel = Model().cuda()
TestOptimizer = torch.optim.SGD(TestModel.parameters(), lr=1e-3)
TestCriterion = CrossEntropyLoss().cuda()
TestMetric = Accuracy()
TestModel, TestOptimizer = apex.amp.initialize(TestModel,
TestOptimizer,
verbosity=0)
def test_default():
runner = Runner(model=TestModel,
optimizer=TestOptimizer,
criterion=TestCriterion,
metrics=TestMetric,
callbacks=None)
runner.fit(TestLoader, epochs=2)
class SegmLoader(Loader):
def __next__(self):
img = torch.randn(BS, 3, IMG_SHAPE, IMG_SHAPE)
target = torch.randint(2, (BS, NUM_CLASSES, IMG_SHAPE, IMG_SHAPE))
return img.cuda(), target.cuda()
TEST_LOADER = Loader()
TEST_SEGM_LOADER = SegmLoader()
TEST_MODEL = Model().cuda()
TEST_SEGM_MODEL = SegmModel().cuda()
TEST_OPTIMIZER = torch.optim.SGD(TEST_MODEL.parameters(), lr=1e-3)
TEST_SEGM_OPTIMZER = torch.optim.SGD(TEST_SEGM_MODEL.parameters(), lr=1e-3)
TEST_CRITERION = CrossEntropyLoss().cuda()
TEST_METRIC = Accuracy()
TEST_MODEL, TEST_OPTIMIZER = apex.amp.initialize(TEST_MODEL, TEST_OPTIMIZER, verbosity=0)
TEST_SEGM_MODEL, TEST_SEGM_OPTIMZER = apex.amp.initialize(TEST_SEGM_MODEL, TEST_SEGM_OPTIMZER)
def test_default():
runner = Runner(
model=TEST_MODEL,
optimizer=TEST_OPTIMIZER,
criterion=TEST_CRITERION,
metrics=TEST_METRIC,
callbacks=None,
)
runner.fit(TEST_LOADER, epochs=2)