crossentropy

Code/train.py

@@ -103,10 +103,10 @@ def train(model, name, seed,gradient_accumulation_steps,mixup, threshold, lambda
 					#print("span output: ", span_output)
 					logits=model.classifier(span_output.detach()) #target_value?
 						
-					# print("logits: ", logits)
-					# print("logits shape: ", list(logits.shape))
+					print("logits: ", logits)
+					print("logits shape: ", list(logits.shape))
 					print("Newlabels: ", new_labels_batch)
-					# print("labels shape: ", list(new_labels_batch.shape))
+					print("labels shape: ", list(new_labels_batch.shape))
 
 					logits = logits.view(-1, 2).to("cuda")
 					target = new_labels_batch.view(-1).to("cuda")
@@ -137,22 +137,14 @@ def train(model, name, seed,gradient_accumulation_steps,mixup, threshold, lambda
 
 	return evaluation_test, evaluation_train
 
-#pytorch Forum, try with dim=1 for targets, changed from /logits.shape[0] zu .mean..
-# def cross_entropy(logits, target):
-# 	print("Cross Entropy Target: ", target)
-# 	logprobs = torch.nn.functional.log_softmax (logits, dim = 1)
-# 	return  -torch.mean(torch.sum(target * logprobs, dim=1))
-
+#-L(y, y') = -1/N * ∑( ∑(y'i * log(softmax(logits_i))) )
 def cross_entropy(logits, target):
     log_softmax = torch.nn.functional.log_softmax(logits, dim=1)
     loss = -torch.sum(target * log_softmax, dim=1)
     return loss.mean()
 
-#without softmax -> gives all values 0.0 (acc. around 70)
-# def cross_entropy(logits, target):
-#     log_probs = -torch.log(torch.exp(logits) / torch.sum(torch.exp(logits), dim=1, keepdim=True))
-#     loss = -torch.mean(torch.sum(target * log_probs, dim=1))
-#     return loss
+#def cross_entropy(logits, target):
+
 
 def mixup_function(batch_of_matrices, batch_of_labels, l, t):
 	runs = math.floor(batch_of_matrices.size()[0]/2)