def train(net, loader, ep, scheduler=None, writer=None):
global n_iter
if scheduler:
scheduler.step()
net.train()
loss_all, norm_all = [], []
train_iter = tqdm(loader)
for images, labels in train_iter:
n_iter += 1
images, labels = images.cuda(), labels.cuda()
embedding = net(images)
loss = criterion(embedding, labels)
loss_all.append(loss.item())
if writer:
writer.add_scalar('loss/train', loss.item(), n_iter)
print(cuda.memory_allocated(cuda.current_device()))
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_iter.set_description("[Train][Epoch %d] Loss: %.5f" % (ep, loss.item()))
print('[Epoch %d] Loss: %.5f\n' % (ep, torch.Tensor(loss_all).mean()))
def train(ep, writer=None):
global n_iter
lr_scheduler.step()
model.train()
sample_net.train()
loss_all = []
train_iter = tqdm(zip(loader_train_sample, loader_eval_sample),
total=len(loader_train_sample))
for (images, labels), (val_images, val_labels) in train_iter:
n_iter += 1
images, labels = images.cuda(), labels.cuda()
val_images, val_labels = val_images.cuda(), val_labels.cuda()
# weight = torch.ones((len(images), len(labels)), device=images.device, requires_grad=True)
embedding = model(images)
if opts.no_maml:
weight = torch.ones((len(images), len(labels)),
device=images.device)
else:
weight = sample_net(embedding)
loss = criterion(embedding, labels, weight)
optimizer.zero_grad()
model_gradients = grad(loss, model.parameters(), create_graph=True)
updated_param = maml_optimizer.update(model.parameters(),
model_gradients)
updated_param = model.format_parameters(updated_param)
updated_embedding = model(val_images, updated_param)
sample_loss = criterion(updated_embedding, val_labels)
sampler_optimizer.zero_grad()
sample_loss.backward(retain_graph=True)
sampler_optimizer.step()
weight = sample_net(embedding)
loss = criterion(embedding, labels, weight)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_all.append(loss.item())
if writer:
writer.add_scalar('loss/train', loss.item(), n_iter)
train_iter.set_description("[Train, Emnbedding][Epoch %d] Loss: %.5f" %
(ep, loss.item()))
print('[Epoch %d] Loss: %.5f\n' % (ep, torch.Tensor(loss_all).mean()))
def train(net, loader, ep, scheduler=None):
if scheduler is not None:
scheduler.step()
net.train()
loss_all, norm_all = [], []
train_iter = tqdm(loader)
for images, labels in train_iter:
images, labels = images.cuda(), labels.cuda()
loss = criterion(net(images), labels)
loss_all.append(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_iter.set_description("[Train][Epoch %d] Loss: %.5f" %
(ep, loss.item()))
print('[Epoch %d] Loss: %.5f\n' % (ep, torch.Tensor(loss_all).mean()))
def train(net, loader, ep):
lr_scheduler.step()
net.train()
loss_all, norm_all = [], []
train_iter = tqdm(loader, ncols=80)
for images, labels in train_iter:
images, labels = images.cuda(), labels.cuda()
pool, embedding = net(images)
loss = criterion(embedding, labels)
loss_all.append(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_iter.set_description("[Train][Epoch %d] Loss: %.5f" %
(ep, loss.item()))
print('[Epoch %d] MeanLoss: %.5f\n' % (ep, torch.Tensor(loss_all).mean()))
def train(net, loader, ep, scheduler=None, writer=None):
global n_iter
if scheduler:
scheduler.step()
beta_scheduler.step()
net.train()
loss_all, norm_all = [], []
train_iter = tqdm(loader)
for images, labels in train_iter:
n_iter += 1
images, labels = images.cuda(), labels.cuda()
embedding = net(images)
if embedding.dim() == 3:
loss = sum([
criterion(embedding[:, i], labels)
for i in range(embedding.size(1))
])
else:
loss = criterion(embedding, labels)
loss_all.append(loss.item())
if writer:
writer.add_scalar('loss/train', loss.item(), n_iter)
optimizer.zero_grad()
beta_optimizer.zero_grad()
loss.backward()
optimizer.step()
beta_optimizer.step()
train_iter.set_description("[Train][Epoch %d] Loss: %.5f" %
(ep, loss.item()))
print('[Epoch %d] Loss: %.5f\n' % (ep, torch.Tensor(loss_all).mean()))