def train_ensemble(model, acoustic_iterator, linguistic_iterator, optimizer, criterion, reg_ratio):
model.train()
epoch_loss = 0
conf_mat = ConfusionMatrix(np.zeros((NUM_CLASSES, NUM_CLASSES)))
assert len(acoustic_iterator) == len(linguistic_iterator)
for acoustic_tuple, linguistic_tuple in zip(acoustic_iterator(), linguistic_iterator()):
acoustic_batch = acoustic_tuple[0]
labels = acoustic_tuple[1]
linguistic_batch = linguistic_tuple[0]
optimizer.zero_grad()
predictions = model(acoustic_batch, linguistic_batch).squeeze(1)
loss = criterion(predictions, labels)
reg_loss = 0
for param in model.parameters():
reg_loss += param.norm(2)
total_loss = loss + reg_ratio*reg_loss
total_loss.backward()
optimizer.step()
epoch_loss += loss.item()
conf_mat += ConfusionMatrix.from_predictions(predictions, labels)
average_loss = epoch_loss / len(acoustic_iterator)
return average_loss, conf_mat
def train(model, iterator, optimizer, criterion, reg_ratio):
model.train()
epoch_loss = 0
conf_mat = ConfusionMatrix(np.zeros((NUM_CLASSES, NUM_CLASSES)))
for batch, labels in iterator():
optimizer.zero_grad()
predictions = model(batch).squeeze(1)
loss = criterion(predictions, labels)
reg_loss = 0
for param in model.parameters():
reg_loss += param.norm(2)
total_loss = loss + reg_ratio*reg_loss
total_loss.backward()
optimizer.step()
epoch_loss += loss.item()
conf_mat += ConfusionMatrix.from_predictions(predictions, labels)
average_loss = epoch_loss / len(iterator)
return average_loss, conf_mat
def eval_ensemble(ensemble_model, acoustic_model_iterator, linguistic_model_iterator, criterion):
epoch_losses = []
conf_mat = ConfusionMatrix(np.zeros((NUM_CLASSES, NUM_CLASSES)))
with torch.no_grad():
for ((acoustic_batch, labels), (linguistic_batch, _)) in zip(acoustic_model_iterator(), linguistic_model_iterator()):
predictions = ensemble_model(acoustic_batch, linguistic_batch)
predictions = torch.Tensor(predictions)
loss = criterion(predictions.float(), labels)
epoch_losses.append(loss.item())
conf_mat += ConfusionMatrix.from_predictions(predictions, labels)
average_loss = sum(epoch_losses) / len(acoustic_model_iterator)
return average_loss, conf_mat
def evaluate(model, iterator, criterion):
model.eval()
epoch_loss = 0
conf_mat = ConfusionMatrix(np.zeros((NUM_CLASSES, NUM_CLASSES)))
with torch.no_grad():
for batch, labels in iterator():
predictions = model(batch)
loss = criterion(predictions.float(), labels)
epoch_loss += loss.item()
conf_mat += ConfusionMatrix.from_predictions(predictions, labels)
average_loss = epoch_loss / len(iterator)
return average_loss, conf_mat
