#forward, loss, backward, step
output = encoder(images=images, rotation=rotation, mode=1)
loss_1 = noise_contrastive_estimator(representations,
output[1],
index,
memory,
negative_nb=negative_nb)
loss_2 = noise_contrastive_estimator(representations,
output[0],
index,
memory,
negative_nb=negative_nb)
loss = loss_weight * loss_1 + (1 - loss_weight) * loss_2
loss.backward()
optimizer.step()
#update representation memory
memory.update(index, output[0].detach())
# update metric and bar
train_loss.update(loss.item(), images.shape[0])
bar.update(step, values=[('train_loss', train_loss.return_avg())])
#save model if improved
checkpoint.save_model(encoder, optimizer, train_loss.return_avg(),
epoch)
logger.update(epoch, train_loss.return_avg())
optimizer.zero_grad()
#forward, loss, backward, step
output = net(images=images, patches=patches, mode=1)
loss_1 = noise_contrastive_estimator(representations,
output[1],
index,
memory,
negative_nb=negative_nb)
loss_2 = noise_contrastive_estimator(representations,
output[0],
index,
memory,
negative_nb=negative_nb)
loss = loss_weight * loss_1 + (1 - loss_weight) * loss_2
loss.backward()
optimizer.step()
# update representation memory
memory.update(index, output[0].detach().cpu().numpy())
# update metric and bar
train_loss.update(loss.item(), images.shape[0])
bar.update(step, values=[('train_loss', train_loss.return_avg())])
logger.update(epoch, train_loss.return_avg())
# save model if improved
checkpoint.save_model(net, train_loss.return_avg(), epoch)
