memory usage reduced

ud.py

@@ -84,10 +84,13 @@ class UniversalDependenciesCausalPipeline(BellmanFordTokenClassificationPipeline
       m.append(e[x,:].sum(axis=0))
     m.append(e[self.tokenizer.sep_token_id,:])
     m.append(e[self.tokenizer.pad_token_id,:])
-    m=torch.stack(m)
+    m=torch.stack(m).to(self.device)
     k=list(range(len(d)+1))
+    e=[]
     with torch.no_grad():
-      e=self.model(inputs_embeds=torch.stack([m[k+list(range(i,len(d)))+[-1]*i,:] for i in range(len(d))]).to(self.device)).logits[:,-len(d):,:].cpu().numpy()
+      for i in range(len(d)):
+        e.append(self.model(inputs_embeds=torch.unsqueeze(m[k+list(range(i,len(d)))+[-1]*i,:],0)).logits[0,-len(d):,:])
+    e=torch.stack(e).cpu().numpy()
     for i in range(len(d)):
       for j in range(i):
         e[-j-1,-i-1],e[-i-1,-j-1]=e[-i-1,i-j]+self.left_arc,e[-i-1,i-j]+self.right_arc