def adapt(self, inputs, targets):
params = None
results = {
'inner_losses':
np.zeros((self.num_adaptation_steps, ), dtype=np.float32)
}
for step in range(self.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):
if self.is_classification_task:
accuracy_before = compute_accuracy(logits, targets)
results["accuracy_before"] = accuracy_before
else:
mse_before = inner_loss
results["mse_before"] = mse_before
self.model.zero_grad()
params = update_parameters(
self.model,
inner_loss,
step_size=self.step_size,
params=params,
first_order=(not self.model.training) or self.first_order,
freeze_visual_features=self.freeze_visual_features,
no_meta_learning=self.no_meta_learning)
return params, results
def adapt(self, inputs, targets):
results = {'inner_losses': np.zeros(
(self.num_adaptation_steps,), dtype=np.float32)}
params, params_masked, masks_logits, masks = self.init_params()
for step in range(self.num_adaptation_steps):
params_masked, masks_logits, reg = self.apply_masks(params, params_masked, masks_logits,
masks, regularize=True)
logits = self.model(inputs, params=params_masked)
inner_loss = self.loss_function(logits, targets) + reg
results['inner_losses'][step] = inner_loss.item()
if (step == 0) and self.is_classification_task:
results['accuracy_before'] = compute_accuracy(logits, targets)
self.model.zero_grad()
masks_logits = update_parameters(self.model, inner_loss,
step_size=self.step_size, params=masks_logits,
first_order=(not self.model.training) or self.first_order,
freeze_visual_features = self.freeze_visual_features,
no_meta_learning=self.no_meta_learning)
self.current_mask_stats = masks_logits
# final masking
params_masked, _ = self.apply_masks(params, params_masked, masks_logits, masks,
regularize=False, evaluate=(not self.model.training))
return params_masked, results