def train_network(net, optimizer, loss_func, epochs, train_data, test_data,
with_preprocessing):
train_loss_per_epoch, test_loss_per_epoch, train_acc_per_epoch, test_acc_per_epoch = [], [], [], []
for epoch in range(epochs):
running_loss = 0.0
for i_batch, sample_batched in enumerate(train_data):
images, labels = sample_batched
images, labels = images.to(torch.device("cuda:0")), labels.to(
torch.device("cuda:0"))
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = net(images.float())
loss = loss_func(outputs, labels).cuda()
loss.backward()
optimizer.step()
# calculate train & test loss
epoch_train_loss = calc_dataset_loss(train_data, net, loss_func)
epoch_test_loss = calc_dataset_loss(test_data, net, loss_func)
# calculate train & test accuracy
epoch_train_acc = calc_dataset_acc(train_data, net)
epoch_test_acc = calc_dataset_acc(test_data, net)
print(
'[epoch %d]\ttrain_loss: %.3f\t test_loss: %.3f\ttrain_acc: %.2f%%\ttest_acc: %.2f%%'
% (epoch + 1, epoch_train_loss, epoch_test_loss,
100 * epoch_train_acc, 100 * epoch_test_acc))
train_loss_per_epoch.append(epoch_train_loss)
test_loss_per_epoch.append(epoch_test_loss)
train_acc_per_epoch.append(epoch_train_acc)
test_acc_per_epoch.append(epoch_test_acc)
np.save(
f"statistics/results/preprocessing/results/{with_preprocessing}_preprocessing_stats",
np.array([
train_loss_per_epoch, test_loss_per_epoch, train_acc_per_epoch,
test_acc_per_epoch
]))
print()
# Train Network
for epoch in range(epochs):
for sample_batched in train_dataloader:
images, labels = sample_batched
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = net(images.float())
loss = loss_fn(outputs, labels)
loss.backward()
optimizer.step()
# calculate train & test loss
epoch_train_loss = calc_dataset_loss(train_dataloader, net, loss_fn)
epoch_test_loss = calc_dataset_loss(test_dataloader, net, loss_fn)
# calculate train & test accuracy
epoch_train_acc = calc_dataset_acc(train_dataloader, net)
epoch_test_acc = calc_dataset_acc(test_dataloader, net)
print('[epoch %d]\ttrain_loss: %.3f\t test_loss: %.3f\ttrain_acc: %.2f%%\ttest_acc: %.2f%%' %
(epoch + 1, epoch_train_loss, epoch_test_loss, 100*epoch_train_acc, 100*epoch_test_acc))
train_loss_per_epoch.append(epoch_train_loss)
test_loss_per_epoch.append(epoch_test_loss)
train_acc_per_epoch.append(epoch_train_acc)
test_acc_per_epoch.append(epoch_test_acc)
np.save("statistics/results/widths/results/width_%d_stats" % network_width, np.array([train_loss_per_epoch, test_loss_per_epoch, train_acc_per_epoch, test_acc_per_epoch]))
print()