def softmax_cross_entropy_with_label_smoothing(pred,
label,
label_smoothing=0.1):
'''
Defines softmax activation followed by Cross entropy loss and label smoothing.
Label smoothing loss is added by the following weight:
`(1 - label_smoothing) * xent_loss + label_smoothing * label_smoothing_loss`
Args:
pred (Variable): Logits with a shape of `(batch_size, num_classes)`.
label (Variable):
A class index for each example if a shape of `(batch_size, 1`) is
given, and a one-hot or probability over classes if
`(batch_size, num_classes)`.
label_smoothing (float):
Coefficient of label smoothing loss. If 0, it omits label
smoothing.
'''
logp = None
if label.shape[1] > 1:
# If mixup is enabled, we suppose the label shape is (batch_size, num_class)
logp = F.log_softmax(pred)
l = F.sum(-label * logp, axis=1, keepdims=True)
else:
l = F.softmax_cross_entropy(pred, label)
return apply_label_smoothing(l, pred, label_smoothing, logp)
def apply_label_smoothing(xent, pred, label_smoothing, logp=None):
if label_smoothing <= 0:
return xent
if logp is None:
logp = F.log_softmax(pred)
return (1 - label_smoothing) * xent - label_smoothing * F.mean(
logp, axis=1, keepdims=True)
def loss_function(pred, y_output, N):
# calculate cross entropy loss
Phi = F.sum(-y_output * F.log_softmax(pred, axis=1))
# calculate l2 norm of affine layer
l2 = 0
for param in nn.get_parameters().values():
l2 += F.sum(param**2)
loss = l2 * args.lmbd + Phi / N
return loss, Phi, l2
def test_obsolete_inplace_option(inplace, func, num_inputs):
'''
This test confirms the construction of graph.
Since F.log_softmax requires output for backward calculation, graph cannot be constructed if it is inplaced.
'''
x0 = nn.Variable((2, 3, 4, 5), need_grad=True)
x1 = nn.Variable((2, 3, 4, 5), need_grad=True)
if num_inputs == 1:
y = F.identity(x0)
y = F.log_softmax(y)
y = func(y, inplace=inplace)
y.forward()
y.backward()
elif num_inputs == 2:
y0 = F.identity(x0)
y1 = F.identity(x1)
y0 = F.log_softmax(y0)
y1 = F.log_softmax(y1)
y = func(y0, y1, inplace=inplace)
y.forward()
y.backward()
def label_smoothing_loss(pred, label, label_smoothing=0.1):
loss = F.softmax_cross_entropy(pred, label)
if label_smoothing <= 0:
return loss
return (1 - label_smoothing) * loss - label_smoothing \
* F.mean(F.log_softmax(pred), axis=1, keepdims=True)
def softmax_cross_entropy_loss_vlabel(pred, vlabel):
# The shape of vlabel is supposed to be (batch_size, n_class)
logp = F.log_softmax(pred)
loss = -1.0 * F.mean(F.sum(vlabel * logp, axis=1))
return loss
def loss_function(pred, label, label_smoothing=0.1):
l = F.softmax_cross_entropy(pred, label)
if label_smoothing <= 0:
return l
return (1 - label_smoothing) * l - label_smoothing * F.mean(
F.log_softmax(pred), axis=1, keepdims=True)
