Support classifier's conversion

BERT/convert_params.py

@@ -134,6 +134,11 @@ def parse(init_checkpoint):
                 fluid_param_name = 'cls_squad_out_b'
             else:
                 print("ignored param: %s" % var_name)
+        else:
+            if var_name == 'output_weights':
+                fluid_param_name = 'cls_out_w'
+            elif var_name == 'output_bias':
+                fluid_param_name = 'cls_out_b'
             else:
                 print("ignored param: %s" % var_name)
 
@@ -172,6 +177,8 @@ def convert(args):
             value = np.transpose(value)
         if param == 'cls/squad/output_weights':
             value = np.transpose(value)
+        if param == 'output_weights':
+            value = np.transpose(value)
         fluid.global_scope().find_var(tf_fluid_param_name_map[
             param]).get_tensor().set(value, place)
         print(param, ' --> ', tf_fluid_param_name_map[param], '  ', value.shape)