def adapt(self,
inputs,
targets,
is_classification_task=None,
num_adaptation_steps=1,
step_size=0.1,
first_order=False):
if is_classification_task is None:
is_classification_task = (not targets.dtype.is_floating_point)
params = None
results = {
'inner_losses': np.zeros((num_adaptation_steps, ),
dtype=np.float32)
}
for step in range(num_adaptation_steps):
logits = self.model(inputs, params=params)
inner_loss = self.loss_function(logits, targets)
results['inner_losses'][step] = inner_loss.item()
if (step == num_adaptation_steps - 1) and is_classification_task:
results['accuracy_before'] = compute_accuracy(logits, targets)
#print(step, logits,targets,results['accuracy_before'])
self.model.zero_grad()
params = gradient_update_parameters(
self.model,
inner_loss,
step_size=step_size,
params=params,
first_order=(not self.model.training) or first_order)
return params, results
def adapt(self,
inputs,
targets,
num_adaptation_steps=1,
step_size=0.1,
first_order=False):
params = None
results = {
'inner_losses': np.zeros((num_adaptation_steps, ),
dtype=np.float32)
}
for step in range(num_adaptation_steps):
logits = self.model(inputs, params=params)
inner_loss = self.loss_function(logits, targets)
results['inner_losses'][step] = inner_loss.item()
if step == 0:
results['accuracy_before'] = compute_accuracy(logits, targets)
self.model.zero_grad()
params = gradient_update_parameters(
self.model,
inner_loss,
step_size=step_size,
params=params,
first_order=(not self.model.training) or first_order)
return params, results
