fix bug for regression is_prediction ==true

fix using softmax whatever is_classify or is_regression.

finetune/classifier.py

@@ -92,8 +92,12 @@ def create_model(args,
             name=task_name + "_cls_out_b",
             initializer=fluid.initializer.Constant(0.)))
 
+    assert is_classify != is_regression, 'is_classify or is_regression must be true and only one of them can be true'
     if is_prediction:
-        probs = fluid.layers.softmax(logits)
+        if is_classify:
+            probs = fluid.layers.softmax(logits)
+        else:
+            probs = logits
         feed_targets_name = [
             src_ids.name, sent_ids.name, pos_ids.name, input_mask.name
         ]
@@ -101,7 +105,6 @@ def create_model(args,
             feed_targets_name += [task_ids.name]
         return pyreader, probs, feed_targets_name
 
-    assert is_classify != is_regression, 'is_classify or is_regression must be true and only one of them can be true'
     num_seqs = fluid.layers.create_tensor(dtype='int64')
     if is_classify:
         ce_loss, probs = fluid.layers.softmax_with_cross_entropy(