Add ood test

models/test.py

@@ -100,7 +100,51 @@ def test_img_local(net_g, dataset, args, user_idx=-1, idxs=None, return_features
  return accuracy, test_loss, features, targets
  else:
  return accuracy, test_loss
+
+def ood_test_img_local(net_g, dataset, args, user_idx=-1, idxs=None, user_train_targets=None):
+ net_g.eval()
+ # testing
+ per_total = 0
+ per_correct = 0
+ ood_total = 0
+ ood_correct = 0
+ # data_loader = DataLoader(dataset, batch_size=args.bs)
+ data_loader = DataLoader(DatasetSplit(dataset, idxs), batch_size=args.bs, shuffle=False)
+ l = len(data_loader)
+
+ for idx, (data, target) in enumerate(data_loader):
+ if args.gpu != -1:
+ data, target = data.to(args.device), target.to(args.device)
+ user_train_targets = user_train_targets.to(args.device)
+ log_probs = net_g(data)
+ y_pred = log_probs.data.max(1, keepdim=True)[1]
+
+ # get the index of the max log-probability
+ target_dup = torch.cat([target.view(-1, 1)]*len(user_train_targets), dim=1)
+ user_train_targets_dup = torch.cat([user_train_targets.view(1, -1)]*len(target), dim=0)
+ per_ood = torch.sum(target_dup == user_train_targets_dup, dim=1)
+
+ per_idx = torch.where(per_ood == 1)
+ ood_idx = torch.where(per_ood == 0)
+
+ per_pred = y_pred[per_idx]
+ ood_pred = y_pred[ood_idx]
+
+ per_target = target[per_idx]
+ ood_target = target[ood_idx]
+
+ per_total += len(per_target)
+ ood_total += len(ood_target)
+
+ per_correct += per_pred.eq(per_target.data.view_as(per_pred)).long().cpu().sum()
+ ood_correct += ood_pred.eq(ood_target.data.view_as(ood_pred)).long().cpu().sum()
 
+ if args.verbose:
+ print('Local model {}: Average loss: {:.4f}, Accuracy: {}/{} ({:.2f}%)'.format(
+ user_idx, test_loss, correct, len(data_loader.dataset), accuracy))
+ else:
+ return per_correct.item()/per_total*100, ood_correct.item()/ood_total*100
+
 def distance_test_img_local(net_g, dataset_train, dataset_test, args, user_idx=-1, train_idxs=None, test_idxs=None):
  net_g.eval()