fix cross_entropy bug of the axis parameter in log_softmax (#27311)

python/paddle/fluid/tests/unittests/test_cross_entropy_loss.py

@@ -26,7 +26,7 @@ def stable_softmax(x):
     return exps / np.sum(exps)
 
 
-def log_softmax(x, axis=-1):
+def log_softmax(x, axis=1):
     softmax_out = np.apply_along_axis(stable_softmax, axis, x)
     return np.log(softmax_out)
 

python/paddle/nn/functional/loss.py

@@ -1093,7 +1093,7 @@ def cross_entropy(input,
             " 'none', but received %s, which is not allowed." % reduction)
 
     #step 1. log_softmax
-    log_softmax_out = paddle.nn.functional.log_softmax(input)
+    log_softmax_out = paddle.nn.functional.log_softmax(input, axis=1)
     if weight is not None and not isinstance(weight, Variable):
         raise ValueError(
             "The weight' is not a Variable, please convert to Variable.")