def test_loop(self, n_support, optimizer=None): # no optimizer needed for GP
inputs, targets = get_batch(test_people)
support_ind = list(np.random.choice(list(range(19)), replace=False, size=n_support))
query_ind = [i for i in range(19) if i not in support_ind]
x_all = inputs.cuda()
y_all = targets.cuda()
x_support = inputs[:,support_ind,:,:,:].cuda()
y_support = targets[:,support_ind].cuda()
x_query = inputs[:,query_ind,:,:,:]
y_query = targets[:,query_ind].cuda()
# choose a random test person
n = np.random.randint(0, len(test_people)-1)
z_support = self.feature_extractor(x_support[n]).detach()
self.model.set_train_data(inputs=z_support, targets=y_support[n], strict=False)
self.model.eval()
self.feature_extractor.eval()
self.likelihood.eval()
with torch.no_grad():
z_query = self.feature_extractor(x_all[n]).detach()
pred = self.likelihood(self.model(z_query))
lower, upper = pred.confidence_region() #2 standard deviations above and below the mean
mse = self.mse(pred.mean, y_all[n])
return mse
def train_loop(self, epoch, optimizer, params, results_logger):
if params.dataset != "sines":
batch, batch_labels = get_batch(train_people)
else:
batch, batch_labels, amp, phase = SinusoidalDataGenerator(
params.update_batch_size * 2, params.meta_batch_size,
params.num_tasks, params.multidimensional_amp,
params.multidimensional_phase, params.noise,
params.out_of_range).generate()
batch = torch.from_numpy(batch)
batch_labels = torch.from_numpy(batch_labels)
batch, batch_labels = batch.to(self.device), batch_labels.to(
self.device)
for inputs, labels in zip(batch, batch_labels):
optimizer.zero_grad()
output = self.model(self.feature_extractor(inputs))
loss = self.criterion(output, labels)
loss.backward()
optimizer.step()
if (epoch % 10 == 0):
print('[%d] - Loss: %.3f' % (epoch, loss.item()))
results_logger.log("epoch", epoch)
results_logger.log("loss", loss.item())
def test_loop(self, n_support,
optimizer): # we need optimizer to take one gradient step
inputs, targets = get_batch(test_people)
support_ind = list(
np.random.choice(list(range(19)), replace=False, size=n_support))
query_ind = [i for i in range(19) if i not in support_ind]
x_all = inputs.cuda()
y_all = targets.cuda()
x_support = inputs[:, support_ind, :, :, :].cuda()
y_support = targets[:, support_ind].cuda()
x_query = inputs[:, query_ind, :, :, :].cuda()
y_query = targets[:, query_ind].cuda()
# choose a random test person
n = np.random.randint(0, len(test_people) - 1)
optimizer.zero_grad()
z_support = self.feature_extractor(x_support[n]).detach()
output_support = self.model(z_support).squeeze()
loss = self.criterion(output_support, y_support[n])
loss.backward()
optimizer.step()
self.feature_extractor.eval()
self.model.eval()
z_all = self.feature_extractor(x_all[n]).detach()
output_all = self.model(z_all).squeeze()
return self.criterion(output_all, y_all[n])
def train_loop(self, epoch, optimizer, params, results_logger):
# print("NUM KERNEL PARAMS {}".format(sum([p.numel() for p in self.model.parameters() if p.requires_grad])))
# print("NUM TRANSFORM PARAMS {}".format(sum([p.numel() for p in self.feature_extractor.parameters() if p.requires_grad])))
self.model.train()
self.feature_extractor.train()
self.likelihood.train()
if self.is_flow:
self.cnf.train()
if self.dataset != "sines":
batch, batch_labels = get_batch(train_people)
else:
batch, batch_labels, amp, phase = SinusoidalDataGenerator(
params.update_batch_size * 2, params.meta_batch_size,
params.num_tasks, params.multidimensional_amp,
params.multidimensional_phase, params.noise,
params.out_of_range).generate()
if self.num_tasks == 1:
batch = torch.from_numpy(batch)
batch_labels = torch.from_numpy(batch_labels).view(
batch_labels.shape[0], -1)
else:
batch = torch.from_numpy(batch)
batch_labels = torch.from_numpy(batch_labels)
batch, batch_labels = batch.to(self.device), batch_labels.to(
self.device)
# print(batch.shape, batch_labels.shape)
for inputs, labels in zip(batch, batch_labels):
optimizer.zero_grad()
z = self.feature_extractor(inputs)
if self.add_noise:
labels = labels + torch.normal(0, 0.1,
size=labels.shape).to(labels)
if self.is_flow:
delta_log_py, labels, y = self.apply_flow(labels, z)
else:
y = labels
self.model.set_train_data(inputs=z, targets=y)
predictions = self.model(z)
loss = -self.mll(predictions, self.model.train_targets)
if self.is_flow:
loss = loss + torch.mean(delta_log_py)
loss.backward()
optimizer.step()
mse, _ = self.compute_mse(labels, predictions, z)
if epoch % 10 == 0:
print('[%d] - Loss: %.3f MSE: %.3f noise: %.3f' %
(epoch, loss.item(), mse.item(),
self.model.likelihood.noise.item()))
results_logger.log("epoch", epoch)
results_logger.log("loss", loss.item())
results_logger.log("MSE", mse.item())
results_logger.log("noise", self.model.likelihood.noise.item())
def get_support_query_qmul(self, n_support):
inputs, targets = get_batch(test_people)
support_ind = list(
np.random.choice(list(range(19)), replace=False, size=n_support))
query_ind = [i for i in range(19) if i not in support_ind]
x_all = inputs.to(self.device)
y_all = targets.to(self.device)
x_support = inputs[:, support_ind, :, :, :].to(self.device)
y_support = targets[:, support_ind].to(self.device)
x_query = inputs[:, query_ind, :, :, :]
y_query = targets[:, query_ind].to(self.device)
return x_all, x_support, y_all, y_support
def train_loop(self, epoch, optimizer):
batch, batch_labels = get_batch(train_people)
batch, batch_labels = batch.cuda(), batch_labels.cuda()
for inputs, labels in zip(batch, batch_labels):
optimizer.zero_grad()
output = self.model(self.feature_extractor(inputs))
loss = self.criterion(output, labels)
loss.backward()
optimizer.step()
if (epoch % 10 == 0):
print('[%d] - Loss: %.3f' % (epoch, loss.item()))
def train_loop(self, epoch, optimizer):
batch, batch_labels = get_batch(train_people)
batch, batch_labels = batch.cuda(), batch_labels.cuda()
for inputs, labels in zip(batch, batch_labels):
optimizer.zero_grad()
z = self.feature_extractor(inputs)
self.model.set_train_data(inputs=z, targets=labels)
predictions = self.model(z)
loss = -self.mll(predictions, self.model.train_targets)
loss.backward()
optimizer.step()
mse = self.mse(predictions.mean, labels)
if (epoch % 10 == 0):
print('[%d] - Loss: %.3f MSE: %.3f noise: %.3f' %
(epoch, loss.item(), mse.item(),
self.model.likelihood.noise.item()))
def test_loop(self, n_support, optimizer,
params): # we need optimizer to take one gradient step
if params.dataset != "sines":
inputs, targets = get_batch(train_people)
else:
batch, batch_labels, amp, phase = SinusoidalDataGenerator(
params.update_batch_size * 2, params.meta_batch_size,
params.num_tasks, params.multidimensional_amp,
params.multidimensional_phase, params.noise,
params.out_of_range).generate()
inputs = torch.from_numpy(batch)
targets = torch.from_numpy(batch_labels)
inputs, targets = inputs.to(self.device), targets.to(self.device)
support_ind = list(
np.random.choice(list(range(19)), replace=False, size=n_support))
query_ind = [i for i in range(19) if i not in support_ind]
x_all = inputs.to(self.device)
y_all = targets.to(self.device)
x_support = inputs[:, support_ind, :, :, :].to(self.device)
y_support = targets[:, support_ind].to(self.device)
x_query = inputs[:, query_ind, :, :, :].to(self.device)
y_query = targets[:, query_ind].to(self.device)
# choose a random test person
n = np.random.randint(0, len(test_people) - 1)
optimizer.zero_grad()
z_support = self.feature_extractor(x_support[n]).detach()
output_support = self.model(z_support).squeeze()
loss = self.criterion(output_support, y_support[n])
loss.backward()
optimizer.step()
self.feature_extractor.eval()
self.model.eval()
z_all = self.feature_extractor(x_all[n]).detach()
output_all = self.model(z_all).squeeze()
return self.criterion(output_all, y_all[n])
