def test_smooth_l1_loss(self):
layer = tl.SmoothL1Loss()
sample_input = np.ones((2, 2))
sample_target = np.ones((2, 2))
sample_weights = np.ones((2, 2))
full_signature = shapes.signature(
[sample_input, sample_target, sample_weights])
layer.init(full_signature)
x = np.array([[1., 1.], [1., 2.]])
target = np.array([[1., 1.], [1., 0.]])
l1_dist = 2
weights = np.array([[1., 1.], [1., 0.]])
loss = layer((x, target, weights))
np.testing.assert_allclose(loss, 0.0)
weights = np.array([[1., 0.], [0., 1.]])
sum_weights = 2
loss = layer((x, target, weights))
np.testing.assert_allclose(loss, (l1_dist - 0.5) / sum_weights)
x = np.array([[1., 1.], [1., 1.5]])
target = np.array([[1., 1.], [1., 1.]])
l1_dist = 0.5
loss = layer((x, target, weights))
np.testing.assert_allclose(loss, 0.5 * l1_dist**2 / sum_weights)
def f(values, actions, returns, mask):
ind_0, ind_1 = np.indices(actions.shape)
# We calculate length using the shape of returns
# and adequatly remove a superflous slice of values.
# An analogous operation is done in value_batches_stream.
length = returns.shape[1]
values = values[:, :length, :]
selected_values = values[ind_0, ind_1, actions]
shapes.assert_same_shape(selected_values, returns)
shapes.assert_same_shape(selected_values, mask)
if self._smoothl1loss:
return tl.SmoothL1Loss().forward((selected_values, returns, mask))
else:
return tl.L2Loss().forward((selected_values, returns, mask))
def test_smooth_l1_loss(self):
layer = tl.SmoothL1Loss()
model_outputs = np.array([[1., 1.], [1., 2.]])
targets = np.array([[1., 1.], [1., 0.]])
l1_dist = 2
weights = np.array([[1., 1.], [1., 0.]])
loss = layer([model_outputs, targets, weights])
np.testing.assert_allclose(loss, 0.0)
weights = np.array([[1., 0.], [0., 1.]])
sum_weights = 2
loss = layer([model_outputs, targets, weights])
np.testing.assert_allclose(loss, (l1_dist - 0.5) / sum_weights)
model_outputs = np.array([[1., 1.], [1., 1.5]])
targets = np.array([[1., 1.], [1., 1.]])
l1_dist = 0.5
loss = layer([model_outputs, targets, weights])
np.testing.assert_allclose(loss, 0.5 * l1_dist**2 / sum_weights)