try cel

Code/train.py

@@ -140,9 +140,19 @@ def train(model, name, seed,gradient_accumulation_steps,mixup, threshold, lambda
 	return evaluation_test, evaluation_train
 
 
+# def cross_entropy(logits, target):
+#     log_q = torch.log_softmax(logits, dim=1)
+#     return -torch.sum(log_q[range(log_q.shape[0]), target])
+
 def cross_entropy(logits, target):
+    # Calculate log_q
     log_q = torch.log_softmax(logits, dim=1)
-    return -torch.sum(log_q[range(log_q.shape[0]), target])
+	#define classes/options
+    target_class = (target == 0).float()
+    target_class2 = (target == 1).float()
+    target = target.float()
+	# calculate sum of losses of batch size
+    return -(target_class * log_q[:, 0] + target_class2 * log_q[:, 1] + (1 - target_class - target_class2) * (target * log_q[:, 1] + (1 - target) * log_q[:, 0]))
 
 
 def mixup_function(batch_of_matrices, batch_of_labels, l, t):