def test_backward(self):
x = torch.randn(3, 1025, requires_grad=True)
x.retain_grad()
min_value = torch.tensor(0.3)
max_value = torch.tensor(0.7)
min_value.requires_grad = True
max_value.requires_grad = True
min_value.retain_grad()
max_value.retain_grad()
clip_x = QF.clip(x, min_value, max_value)
clip_x.retain_grad()
labels = torch.randint(6, (3, )).type(torch.LongTensor).cuda()
criterion = torch.nn.CrossEntropyLoss()
loss = criterion(clip_x, labels)
loss.backward()
np.testing.assert_array_almost_equal(
clip_x.grad[x < min_value].sum().cpu().numpy(),
min_value.grad.cpu().numpy(),
decimal=6)
np.testing.assert_array_almost_equal(
clip_x.grad[x > max_value].sum().cpu().numpy(),
max_value.grad.cpu().numpy(),
decimal=6)
assert x.grad.cpu()[x < min_value].sum() == 0
assert x.grad.cpu()[x > max_value].sum() == 0
assert torch.equal(clip_x.grad[(x > min_value) & (x < max_value)],
x.grad[(x > min_value) & (x < max_value)])
def test_broadcast(self):
"""Test broadcast behavior by randomly picked shuffling of np.random.rand"""
x_np = np.random.rand(1023, 4, 5, 6).astype(np.float32) - 0.5
x_torch = torch.Tensor(x_np)
min_value = np.random.rand(1, 4, 1, 1).astype(np.float32) * 0.1 - 0.2
max_value = np.random.rand(1, 4, 1, 1).astype(np.float32) * 10 + 0.5
clip_x_np = np.clip(x_np, min_value, max_value)
clip_x_torch = QF.clip(x_torch, torch.tensor(min_value),
torch.tensor(max_value))
np.testing.assert_array_equal(clip_x_torch.cpu().numpy(), clip_x_np)
def test_raise(self):
x = torch.randn(3, 7, requires_grad=True)
min_value = torch.Tensor(3, 7)
max_value = torch.Tensor(3, 7)
min_value.requires_grad = True
max_value.requires_grad = True
clip_x = QF.clip(x, min_value, max_value)
labels = torch.randint(6, (3, )).type(torch.LongTensor).cuda()
criterion = torch.nn.CrossEntropyLoss()
loss = criterion(clip_x, labels)
with pytest.raises(ValueError, match="can only be scalar"):
loss.backward()
def test_simple_run(self):
x_np = np.random.rand(1023).astype(np.float32)
x_torch = torch.Tensor(x_np)
clip_x_np = np.clip(x_np, 0.3, 0.7)
clip_x_torch = QF.clip(x_torch, torch.tensor(0.3), torch.tensor(0.7))
np.testing.assert_array_equal(clip_x_torch.cpu().numpy(), clip_x_np)
def forward(self, inputs):
outputs = QF.clip(inputs, self.clip_value_min, self.clip_value_max)
return outputs