def test_torch(self):
"""Torch tensors will simply return their requires_grad attribute"""
t = torch.tensor([1.0, 2.0], requires_grad=True)
assert fn.requires_grad(t)
t = torch.tensor([1.0, 2.0], requires_grad=False)
assert not fn.requires_grad(t)
def test_autograd(self):
"""Autograd arrays will simply return their requires_grad attribute"""
t = np.array([1.0, 2.0], requires_grad=True)
assert fn.requires_grad(t)
t = np.array([1.0, 2.0], requires_grad=False)
assert not fn.requires_grad(t)
def test_tf(self):
"""TensorFlow tensors will True *if* they are being watched by a gradient tape"""
t1 = tf.Variable([1.0, 2.0])
t2 = tf.constant([1.0, 2.0])
assert not fn.requires_grad(t1)
assert not fn.requires_grad(t2)
with tf.GradientTape():
# variables are automatically watched within a context,
# but constants are not
assert fn.requires_grad(t1)
assert not fn.requires_grad(t2)
with tf.GradientTape() as tape:
# watching makes all tensors trainable
tape.watch([t1, t2])
assert fn.requires_grad(t1)
assert fn.requires_grad(t2)
def test_jax(self, t):
"""jax.DeviceArrays will always return True"""
assert fn.requires_grad(t)
def test_numpy(self, t):
"""Vanilla NumPy arrays, sequences, and lists will always return False"""
assert not fn.requires_grad(t)
