def set_teacher_signal(self, y):
if isinstance(y, dict): y = y[P.KEY_LABEL_TARGETS]
if y is not None: y = utils.dense2onehot(y, self.NUM_CLASSES)
self.fc6.set_teacher_signal(y)
if y is None:
self.conv4.set_teacher_signal(y)
self.fc5.set_teacher_signal(y)
elif self.DEEP_TEACHER_SIGNAL:
# Extend teacher signal for deep layers
l4_knl_per_class = 240 // self.NUM_CLASSES
l5_knl_per_class = 4000 // self.NUM_CLASSES
if self.NUM_CLASSES <= 20:
self.conv4.set_teacher_signal(
torch.cat((
torch.ones(y.size(0),
self.conv4.weight.size(0) -
l4_knl_per_class * self.NUM_CLASSES,
device=y.device),
y.view(y.size(0), y.size(1), 1).repeat(
1, 1, l4_knl_per_class).view(y.size(0), -1),
),
dim=1))
self.fc5.set_teacher_signal(
torch.cat((
torch.ones(y.size(0),
self.fc5.weight.size(0) -
l5_knl_per_class * self.NUM_CLASSES,
device=y.device),
y.view(y.size(0), y.size(1), 1).repeat(
1, 1, l5_knl_per_class).view(y.size(0), -1),
),
dim=1))
def set_teacher_signal(self, y):
if y is not None: y = utils.dense2onehot(y, self.NUM_CLASSES)
self.fc10.set_teacher_signal(y)
if y is None:
self.conv8.set_teacher_signal(y)
self.fc9.set_teacher_signal(y)
elif self.DEEP_TEACHER_SIGNAL:
# Extend teacher signal for deep layers
l8_knl_per_class = 500 // self.NUM_CLASSES
l9_knl_per_class = 4000 // self.NUM_CLASSES
if self.NUM_CLASSES <= 20:
self.conv8.set_teacher_signal(
torch.cat((
torch.ones(y.size(0),
self.conv8.weight.size(0) -
l8_knl_per_class * self.NUM_CLASSES,
device=y.device),
y.view(y.size(0), y.size(1), 1).repeat(
1, 1, l8_knl_per_class).view(y.size(0), -1),
),
dim=1))
self.fc9.set_teacher_signal(
torch.cat((
torch.ones(y.size(0),
self.fc9.weight.size(0) -
l9_knl_per_class * self.NUM_CLASSES,
device=y.device),
y.view(y.size(0), y.size(1), 1).repeat(
1, 1, l9_knl_per_class).view(y.size(0), -1),
),
dim=1))
def set_teacher_signal(self, y):
if isinstance(y, dict): y = y[P.KEY_LABEL_TARGETS]
if y is not None: y = utils.dense2onehot(y, self.NUM_CLASSES)
self.fc2.set_teacher_signal(y)
def set_teacher_signal(self, y):
if y is not None: y = utils.dense2onehot(y, self.NUM_CLASSES)
self.fc2.set_teacher_signal(y)
def __call__(self, outputs, targets):
if isinstance(outputs, dict): outputs = outputs[P.KEY_CLASS_SCORES]
if isinstance(targets, dict): targets = targets[P.KEY_LABEL_TARGETS]
return self.mse_loss(outputs,
utils.dense2onehot(targets, outputs.size(1)))
