Changed device assignments to be compatible with data-parallel. (#51)

modeling/models.py

@@ -364,7 +364,7 @@ def forward(self, x1, x2):
 
         if self.cfg.task == '2':
             # Concat. TPs
-            x = torch.cat([x1, x2], dim=1).to(self.cfg.device)
+            x = torch.cat([x1, x2], dim=1).to(x1.device)
             # x: (B, 2, SV, SV, SV)
         else:
             # Discard x2
@@ -581,9 +581,9 @@ def __init__(self, cfg):
 
     def forward(self, x):
         # Get position IDs and timepoint IDs
-        position_ids = torch.cat([torch.arange(self.cfg.num_patches)] * 2).unsqueeze(0).to(self.cfg.device)
+        position_ids = torch.cat([torch.arange(self.cfg.num_patches)] * 2).unsqueeze(0).to(x.device)
         # position_ids: (1, 2 * S)
-        timepoint_ids = torch.tensor([0] * self.cfg.num_patches + [1] * self.cfg.num_patches).unsqueeze(0).to(self.cfg.device)
+        timepoint_ids = torch.tensor([0] * self.cfg.num_patches + [1] * self.cfg.num_patches).unsqueeze(0).to(x.device)
         # timepoint_ids: (1, 2 * S)
 
         x, context_out, context_features = self.embeddings(x, position_ids, timepoint_ids)
@@ -632,7 +632,7 @@ def forward(self, x1, x2):
         # x2: (B, S, 1, P, P, P)
 
         # Concat. TPs
-        x = torch.cat([x1, x2], dim=1).to(self.cfg.device)
+        x = torch.cat([x1, x2], dim=1).to(x1.device)
         # x: (B, 2 * S, 1, P, P, P)
 
         x, attention_weights, context_out, context_features = self.transformer(x)