CEL

Code/train.py

@@ -141,10 +141,10 @@ def train(model, name, seed,gradient_accumulation_steps,mixup, threshold, lambda
 #     return  -(target * logprobs).sum() / input.shape[0]
 
 #gives average #2
-def cross_entropy(preds, target):
-    log_preds = torch.log(preds)
-    loss = -torch.sum(target * log_preds) / preds.shape[0]
-    return loss
+# def cross_entropy(preds, target):
+#     log_preds = torch.log(preds)
+#     loss = -torch.sum(target * log_preds) / preds.shape[0]
+#     return loss
 
 #1
 # def cross_entropy(preds, target):
@@ -160,6 +160,14 @@ def cross_entropy(preds, target):
 #     return loss
 
 
+def cross_entropy(preds, target):
+    device = preds.device
+    target = torch.eye(preds.shape[1], device=device)[target.long()]
+    log_preds = torch.log(preds)
+    loss = -torch.mean(target * log_preds)
+    return loss
+
+
 def mixup_function(batch_of_matrices, batch_of_labels, l, t):
 	runs = math.floor(batch_of_matrices.size()[0]/2)
 	counter=0