def cuda(self): """ If the trainset/valset are lists of tensors, this pushes them to cuda """ for attr in 'trainset', 'valset': if isinstance(getattr(self, attr), list): setattr(self, attr, [cudafy(_) for _ in getattr(self, attr)]) return
def test_validation(network, valset, loss_functional=None, use_cuda=False): """ Report the top1 accuracy of network over the if loss_functional is None: Returns top1 accuracy in range [0.0, 1.0] else: Returns average loss value over valset (Assumes devices are already set) """ total = 0 count_value = 0 for examples, labels in valset: if examples.dtype != network.dtype: examples = examples.type(network.dtype) if use_cuda: examples, labels = cudafy([examples, labels]) total += labels.numel() if loss_functional is None: xout = network(examples).max(1)[1] xout == labels count_value += (network(examples).max(1)[1] == labels).sum().item() else: count_value += loss_functional(examples, labels).data.item() return float(count_value) / total
def training_loop(network, train_params, epoch_start_no=0,
use_cuda=False, epoch_callback=None):
""" Trains a network over the trainset with given loss and optimizer
ARGS:
network: ReLUNet - network to train
train_params: TrainParameters - parameters object governing training
epoch_start_no: int - number to start epochs at for print purposes
use_cuda: bool - if True, we use CUDA to train the net. Everything
gets returned on CPU
epoch_callback: if not none, is a function that takes in arguments
{'network': network, 'epoch_no': epoch_no}
RETURNS:
None, but modifies network parameters
"""
use_cuda = torch.cuda.is_available() and use_cuda
if use_cuda:
network.cuda()
train_params.cuda()
else:
network.cpu()
train_params.cpu()
# Unpack the parameter object
if train_params.optimizer is None:
optimizer = optim.Adam(network.parameters(), lr=0.001, weight_decay=0)
else:
optimizer = train_params.optimizer
if train_params.test_after_epoch is False:
test_after_epoch = float('inf')
else:
test_after_epoch = int(train_params.test_after_epoch)
if train_params.loss_functional is None:
loss_functional = LossFunctional(regularizers=[XEntropyReg()])
else:
loss_functional = train_params.loss_functional
loss_functional.attach_network(network)
# Do the training loop
for epoch_no in range(epoch_start_no, epoch_start_no + train_params.num_epochs + 1):
if epoch_callback is not None:
epoch_callback(network=network, epoch_no=epoch_no)
for i, (examples, labels) in enumerate(train_params.trainset):
if examples.dtype != network.dtype:
examples = examples.type(network.dtype)
if use_cuda:
examples, labels = cudafy([examples, labels])
optimizer.zero_grad()
loss = loss_functional(examples, labels)
loss.backward()
optimizer.step()
if (epoch_no) % test_after_epoch == 0:
# If we run test accuracy this test
test_acc = test_validation(network, train_params.valset,
use_cuda=use_cuda)
test_acc_str = '| Accuracy: %.02f' % (test_acc * 100)
print(("Epoch %02d " % epoch_no) + test_acc_str)
if use_cuda:
network = network.cpu()
train_params.cpu()