def valid(weights):
running_loss = 0
correct = np.zeros(10)
pos = np.zeros(10)
total = 0
predictions = [np.array([])] * 10
actual = [np.array([])] * 10
v_neg_weights, v_class_weights = weights
v_neg_weights, v_class_weights = v_neg_weights.to(
args.device), v_class_weights.to(args.device)
with torch.no_grad():
for batch, labels, seq_len in val_loader:
# pass to GPU if available
batch, labels = batch.to(args.device), labels.to(args.device)
# run network
optimizer.zero_grad()
outputs = model(batch, seq_len, labels.shape[1])
multiplier = (((labels == 0).double() * v_neg_weights) +
(labels == 1).double()) * v_class_weights
mask = (multiplier > 0) * (labels <= 1) * (labels >= 0)
loss = torch.mean(
criterion(outputs.masked_select(mask),
labels.masked_select(mask)) *
multiplier.masked_select(mask))
running_loss += loss.cpu().data.numpy()
# Validation accuracy
for i in range(labels.shape[0]):
targets = labels.data[i][:int(seq_len[i])].cpu().numpy()
prob = outputs.data[i][:int(seq_len[i])].cpu().numpy()
prediction = np.zeros(targets.shape)
prediction[np.arange(prediction.shape[0]),
np.argmax(prob, axis=1)] = 1
match = (targets == prediction)
pos += np.sum(prediction, axis=0).astype(int)
correct += np.sum(match, axis=0).astype(int)
total += targets.shape[0]
# for AUC
for j in range(10):
actual[j] = np.concatenate(
(actual[j], labels[i][:int(seq_len[i]),
j].view(-1).cpu().numpy()))
predictions[j] = np.concatenate(
(predictions[j], outputs[i][:int(seq_len[i]),
j].view(-1).cpu().numpy()))
val_losses[epoch] = running_loss / len(val_loader)
val_acc[epoch] = correct / total
val_freq[epoch] = pos / sum(pos)
val_auc[epoch] = get_aucs(actual, predictions)
def train(weights):
running_loss = 0
correct = np.zeros((10))
pos = np.zeros((10))
total = 0
predictions = [np.array([])] * 10
actual = [np.array([])] * 10
t_neg_weights, t_class_weights = weights
t_neg_weights, t_class_weights = t_neg_weights.to(
args.device), t_class_weights.to(args.device)
for batch, labels, seq_len in train_loader:
# pass to GPU if available
batch, labels = batch.to(args.device), labels.to(args.device)
# run network
optimizer.zero_grad()
outputs = model(batch)
multiplier = (((labels == 0).double() * t_neg_weights) +
(labels == 1).double()) * t_class_weights
mask = (multiplier > 0) * (labels <= 1) * (labels >= 0)
loss = torch.mean(
criterion(outputs.masked_select(mask), labels.masked_select(mask))
* multiplier.masked_select(mask))
# adjust weights and record loss
loss.backward()
optimizer.step()
running_loss += loss.cpu().data.numpy()
# train accuracy
for i in range(labels.shape[0]):
targets = labels.data[i][:int(seq_len[i])].cpu().numpy()
prob = outputs.data[i][:int(seq_len[i])].cpu().numpy()
prediction = np.zeros(targets.shape)
prediction[np.arange(prediction.shape[0]),
np.argmax(prob, axis=1)] = 1
match = (targets == prediction)
pos += np.sum(prediction, axis=0).astype(int)
correct += np.sum(match, axis=0).astype(int)
total += targets.shape[0]
# for AUC
for j in range(10):
actual[j] = np.concatenate(
(actual[j], labels[i][:int(seq_len[i]),
j].view(-1).cpu().numpy()))
predictions[j] = np.concatenate(
(predictions[j],
outputs[i][:int(seq_len[i]),
j].detach().view(-1).cpu().numpy()))
train_losses[epoch] = running_loss / len(train_loader)
train_acc[epoch] = correct / total
train_freq[epoch] = pos / sum(pos)
train_auc[epoch] = get_aucs(actual, predictions)
