use `mod.predict` directly

cnocr/cn_ocr.py

@@ -252,9 +252,7 @@ class CnOcr(object):
         batch_size = len(img_list)
         img_list, img_widths = self._pad_arrays(img_list)
 
-        sample = SimpleBatch(data_names=['data'], data=[mx.nd.array(img_list)])
-
-        prob = self._predict(sample)
+        prob = self._predict(mx.nd.array(img_list))
         # [seq_len, batch_size, num_classes]
         prob = np.reshape(prob, (-1, batch_size, prob.shape[1]))
 
@@ -310,12 +308,9 @@ class CnOcr(object):
         return padded_img_list, img_widths
 
     def _predict(self, sample):
-        mod = self._mod
-        mod.forward(sample)
-        prob = mod.get_outputs()[0]
+        prob = self._mod.predict(sample)
         mx.nd.waitall()
-        prob = prob.asnumpy()
-        return prob
+        return prob.asnumpy()
 
     def _gen_line_pred_chars(self, line_prob, img_width, max_img_width):
         """