def test_step(self,
data_dict,
n_support,
n_classes,
n_query,
n_unlabeled=0,
n_episodes=1000):
accuracies = list()
losses = list()
self.eval()
for i in range(n_episodes):
episode = create_episode(data_dict=data_dict,
n_support=n_support,
n_classes=n_classes,
n_query=n_query,
n_unlabeled=n_unlabeled)
with torch.no_grad():
if n_unlabeled:
loss, loss_dict = self.loss_softkmeans(episode)
else:
loss, loss_dict = self.loss(episode)
accuracies.append(loss_dict["acc"])
losses.append(loss_dict["loss"])
return {"loss": np.mean(losses), "acc": np.mean(accuracies)}
def train_step(self, optimizer, data_dict: Dict[str, List[str]], n_support,
n_classes, n_query):
episode = create_episode(data_dict=data_dict,
n_support=n_support,
n_classes=n_classes,
n_query=n_query)
self.train()
optimizer.zero_grad()
torch.cuda.empty_cache()
loss, loss_dict = self.loss(episode)
loss.backward()
optimizer.step()
return loss, loss_dict
def test_step(self, data_dict, n_support, n_classes, n_query, n_episodes=1000):
metrics = collections.defaultdict(list)
self.eval()
for i in range(n_episodes):
episode = create_episode(
data_dict=data_dict,
n_support=n_support,
n_classes=n_classes,
n_query=n_query
)
with torch.no_grad():
loss, loss_dict = self.loss(episode)
for key, value in loss_dict["metrics"].items():
metrics[key].append(value)
return {
key: np.mean(value) for key, value in metrics.items()
}
def train_step(self, optimizer, data_dict: Dict[str,
List[str]], n_support: int,
n_classes: int, n_query: int, n_unlabeled: int):
episode = create_episode(data_dict=data_dict,
n_support=n_support,
n_classes=n_classes,
n_query=n_query,
n_unlabeled=n_unlabeled)
self.train()
optimizer.zero_grad()
torch.cuda.empty_cache()
if n_unlabeled:
loss, loss_dict = self.loss_softkmeans(episode)
else:
loss, loss_dict = self.loss(episode)
loss.backward()
optimizer.step()
return loss, loss_dict
