def discriminator_performance(x, y):
y, y_pred = to_Var(y), model.eval()(to_Var(x))
faux_pos = ((y_pred.max(1)[1] != y) * (y_pred.max(1)[1] == 0))
faux_pos = faux_pos.double().data.sum()
faux_neg = ((y_pred.max(1)[1] != y) * (y_pred.max(1)[1] == 1))
faux_neg = faux_neg.double().data.sum()
total = (y_pred.max(1)[1] != y).double().data.sum()
return (faux_pos, faux_neg, total)
def big_loss(images, labels):
data = TensorDataset(images, labels)
loader = DataLoader(data, batch_size=100, shuffle=False)
count = 0
for (x, y) in loader:
y, y_pred = to_Var(y), model.eval()(to_Var(x))
count += len(x) * loss_fn(y_pred, y).data.item()
return count / len(images)
def accuracy(images, labels, k=1):
data = TensorDataset(images, labels)
loader = DataLoader(data, batch_size=10, shuffle=False)
count = 0
for (x, y) in loader:
y, y_pred = to_Var(y), model.eval()(to_Var(x))
y_pred_k = y_pred.topk(k, 1, True, True)[1]
count += sum(sum((y_pred_k.t() == y).float())).data
# .double(): ByteTensor sums are limited at 256.
return 100 * count / len(images)
bar_format = left + " |{bar}| " + right
return tqdm(data, desc=epoch, ncols=74, unit='b', bar_format=bar_format)
train_accs, val_accs = [], []
train_losses, val_losses = [], []
try:
# Main loop over each epoch
for e in range(epochs):
# Secondary loop over each mini-batch
for (x, y) in bar(train_loader, e):
# Computes the network output
y_pred = model.train()(to_Var(x))
loss = loss_fn(y_pred, to_Var(y))
# Optimizer step
model.zero_grad()
loss.backward()
optimizer.step()
# Calculates accuracy and loss on the train database.
train_acc = accuracy(train_images, train_labels, k)
train_loss = big_loss(train_images, train_labels)
train_accs.append(train_acc)
train_losses.append(train_loss)
# Calculates accuracy and loss on the validation database.
val_acc = accuracy(val_images, val_labels, k)
def load_image(img_id):
return to_Var(images[img_id].view(1, 1, 28, 28))