Remove self.device to prevent import requirement of torch for prediction (#617)

Author: darrylong

cornac/models/lightgcn/lightgcn.py

@@ -62,11 +62,10 @@ def forward(self, g, feat_dict):
 
 
 class Model(nn.Module):
-    def __init__(self, g, in_size, num_layers, lambda_reg, device=None):
+    def __init__(self, g, in_size, num_layers, lambda_reg):
         super(Model, self).__init__()
         self.norm_dict = dict()
         self.lambda_reg = lambda_reg
-        self.device = device
 
         self.layers = nn.ModuleList([GCNLayer() for _ in range(num_layers)])
 

cornac/models/lightgcn/recom_lightgcn.py

@@ -124,21 +124,21 @@ def fit(self, train_set, val_set=None):
         from .lightgcn import Model
         from .lightgcn import construct_graph
 
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         if self.seed is not None:
             torch.manual_seed(self.seed)
             if torch.cuda.is_available():
                 torch.cuda.manual_seed_all(self.seed)
 
         graph = construct_graph(train_set, self.total_users, self.total_items).to(
-            self.device
+            device
         )
         model = Model(
             graph,
             self.emb_size,
             self.num_layers,
             self.lambda_reg,
-        ).to(self.device)
+        ).to(device)
 
         optimizer = torch.optim.Adam(model.parameters(), lr=self.learning_rate)