def test_square():
inp = np.ones((INT_OVERFLOW, 2))
inp[-1, -1] = 3
inp.attach_grad()
with mx.autograd.record():
out = np.square(inp)
out.backward()
assert out.shape == inp.shape
assert out[-1, -1] == 9
assert inp.grad.shape == inp.shape
assert inp.grad[-1, -1] == 6
def evaluator(network, inter_matrix, test_data, ctx):
scores = []
for values in inter_matrix:
feat = gluon.utils.split_and_load(values, ctx, even_split=False)
scores.extend([network(i).asnumpy() for i in feat])
recons = np.array([item for sublist in scores for item in sublist])
# Calculate the test RMSE.
rmse = np.sqrt(
np.sum(np.square(test_data - np.sign(test_data) * recons)) /
np.sum(np.sign(test_data)))
return float(rmse)
def forward(self, length_predictions, labels):
"""
Returns MSE loss.
:param length_predictions: Length predictions. Shape: (batch_size,).
:param labels: Targets. Shape: (batch_size,).
:return: MSE loss of length predictions of the batch.
"""
# (batch_size,)
loss = (self.weight / 2) * np.square(length_predictions - labels)
# (1,)
loss = np.sum(loss)
num_samples = np.sum(np.ones_like(length_predictions))
return loss, num_samples
def mse_loss(x1, x2):
loss = np.square(x1 - x2).mean()
if loss is None:
return 0
else:
return loss
def global_norm(ndarrays: List[np.ndarray]) -> float:
# accumulate in a list, as item() is blocking and this way we can run the norm calculation in parallel.
norms = [
np.square(np.linalg.norm(arr)) for arr in ndarrays if arr is not None
]
return sqrt(sum(norm.item() for norm in norms))
def style_loss(Y_hat, gram_Y):
return np.square(gram(Y_hat) - gram_Y).mean()
def content_loss(Y_hat, Y):
return np.square(Y_hat, Y).mean()
def style_loss(y_hat, gram_y):
return np.square(StyleTransferGF.gram(y_hat) - gram_y).mean()
def content_loss(y_hat, y):
return np.square(y_hat, y).mean()