modified training

moco/main.py

@@ -37,9 +37,9 @@
 parser.add_argument('--mlp', default=True, type=bool, help='feature dimension')
 
 # misc.
-parser.add_argument('--data_root', default='../data', type=str, help='path to data')
-parser.add_argument('--logs_root', default='logs', type=str, help='path to logs')
-parser.add_argument('--check_point', default='check_point/moco.pth', type=str, help='path to model weights')
+parser.add_argument('--data_root', default='data', type=str, help='path to data')
+parser.add_argument('--logs_root', default='moco/logs', type=str, help='path to logs')
+parser.add_argument('--check_point', default='moco/check_point/moco.pth', type=str, help='path to model weights')
 
 args = parser.parse_args()
 
@@ -73,8 +73,8 @@
     test_data = CIFAR10(root=args.data_root, train=False, transform=test_transform, download=True)
     test_loader = DataLoader(test_data, batch_size=args.batch_size, shuffle=False, num_workers=28)
 
-    Path(args.check_point.split('/')[0]).mkdir(parents=True, exist_ok=True)
-    Path(args.logs_root).mkdir(parents=True, exist_ok=True)
+    Path(args.check_point.split('/')[1]).mkdir(parents=True, exist_ok=True)
+    Path(args.logs_root.split('/')[1]).mkdir(parents=True, exist_ok=True)
 
     f_q = MoCo(args).cuda()
     f_k = get_momentum_encoder(f_q)
@@ -104,7 +104,7 @@
             with torch.no_grad():
                 momentum_update(f_k, f_q, args.m)
             train_losses.append(loss.item())
-            pbar.set_postfix({'Loss': loss.item(), 'Learning Rate': scheduler.get_last_lr()})
+            pbar.set_postfix({'Loss': loss.item(), 'Learning Rate': scheduler.get_last_lr()[0]})
 
         writer.add_scalar('Train Loss', sum(train_losses) / len(train_losses), global_step=epoch)
         torch.save(f_q.state_dict(), args.check_point)

moco/utils.py

@@ -14,11 +14,12 @@ def __init__(self, T=0.07):
         self.criterion = nn.CrossEntropyLoss()
 
     def forward(self, q, k, memo_bank):
+        N = q.size(0)
         k = k.detach()
         pos_logits = torch.einsum('ij,ij->i', [q, k]).unsqueeze(-1)
-        neg_logits = torch.einsum('ij,jk->ik', [q, memo_bank.queue.clone()])
+        neg_logits = torch.einsum('ij,kj->ik', [q, memo_bank.queue.clone()])
         logits = torch.cat([pos_logits, neg_logits], dim=1)
-        labels = torch.zeros_like(logits, dtype=torch.long, device=logits.device)
+        labels = torch.zeros(N, dtype=torch.long, device=logits.device)
         return self.criterion(logits / self.T, labels)
 
 
@@ -75,7 +76,7 @@ def dequeue_data(self, K=None):
             return self.queue
 
     def dequeue_and_enqueue(self, k):
-        self.queue_data(k)
+        self.queue = self.queue_data(k)
         self.queue = self.dequeue_data()
 
 

simsiam/main.py

@@ -33,14 +33,10 @@
 parser.add_argument('--bottleneck_dim', default=512, type=int, help='bottleneck dimension')
 parser.add_argument('--num_encoder_fcs', default=2, type=int, help='number of layers of fcs for encoder')
 
-# knn monitor
-parser.add_argument('--knn-k', default=200, type=int, help='k in kNN monitor')
-parser.add_argument('--knn-t', default=0.1, type=float, help='softmax temperature in kNN monitor')
-
 # misc.
-parser.add_argument('--data_root', default='../data', type=str, help='path to data')
-parser.add_argument('--logs_root', default='logs', type=str, help='path to logs')
-parser.add_argument('--check_point', default='check_point/simsiam.pth', type=str, help='path to model weights')
+parser.add_argument('--data_root', default='data', type=str, help='path to data')
+parser.add_argument('--logs_root', default='simsiam/logs', type=str, help='path to logs')
+parser.add_argument('--check_point', default='simsiam/check_point/simsiam.pth', type=str, help='path to model weights')
 
 args = parser.parse_args()
 
@@ -82,8 +78,8 @@ def cosine_loss(p, z):
 
     writer = SummaryWriter(args.logs_root)
     model = SimSiam(args).to(device)
-    Path(args.check_point.split('/')[0]).mkdir(parents=True, exist_ok=True)
-    Path(args.logs_root).mkdir(parents=True, exist_ok=True)
+    Path(args.check_point.split('/')[1]).mkdir(parents=True, exist_ok=True)
+    Path(args.logs_root.split('/')[1]).mkdir(parents=True, exist_ok=True)
 
     optimizer = torch.optim.SGD(model.parameters(), lr=args.lr,
                                 momentum=args.momentum, weight_decay=args.wd)