def train_epoch(epoch):
recons_batch_loss = []
transreg_batch_loss = []
for batch_idx, data in enumerate(train_loader):
x = Variable(data).to(device)
x_hat, z, trans = self.forward(x)
chamfer_loss = self.reconstruction_loss(x, x_hat)
trans_reg_loss = feature_transform_reguliarzer(trans)
loss = chamfer_loss + trans_reg_loss
recons_batch_loss.append(chamfer_loss.item())
transreg_batch_loss.append(trans_reg_loss.item())
opt.zero_grad()
loss.backward()
opt.step()
if batch_idx % 50 == 0 and True: #suppress iteration output now
print(
'Train Epoch: {} [{}/{} ({:.0f}%)]\t Chamfer Loss: {:.6f}; Trans Reg Loss: {:.6f}'
.format(epoch, batch_idx * len(data),
len(train_loader.dataset),
100. * batch_idx / len(train_loader),
chamfer_loss.item(), trans_reg_loss.item()))
recons_epoch_loss.append(np.mean(recons_batch_loss))
transreg_epoch_loss.append(np.mean(transreg_batch_loss))
return
def forward(self, pred, target, trans_feat):
# https://pytorch.org/docs/stable/generated/torch.nn.NLLLoss.html
loss = F.nll_loss(pred, target)
mat_diff_loss = feature_transform_reguliarzer(trans_feat)
total_loss = loss + mat_diff_loss * self.mat_diff_loss_scale
return total_loss
def forward(self, pred, target, trans_feat, weight):
loss = F.nll_loss(pred, target, weight=weight)
mat_diff_loss = feature_transform_reguliarzer(trans_feat)
total_loss = loss + mat_diff_loss * self.mat_diff_loss_scale
return total_loss
