Simplify arguments of sensitive API.

paddleslim/prune/sensitive.py

@@ -33,12 +33,14 @@ def sensitivity(program,
                 param_names,
                 eval_func,
                 sensitivities_file=None,
-                step_size=0.2,
-                max_pruned_times=None):
+                pruned_ratios=None):
     scope = fluid.global_scope()
     graph = GraphWrapper(program)
     sensitivities = _load_sensitivities(sensitivities_file)
 
+    if pruned_ratios is None:
+        pruned_ratios = np.arange(0.1, 1, step=0.1)
+
     for name in param_names:
         if name not in sensitivities:
             size = graph.var(name).shape()[0]
@@ -49,16 +51,9 @@ def sensitivity(program,
             }
     baseline = None
     for name in sensitivities:
-        ratio = step_size
-        pruned_times = 0
-        while ratio < 1:
-            if max_pruned_times is not None and pruned_times >= max_pruned_times:
-                break
-            ratio = round(ratio, 2)
+        for ratio in pruned_ratios:
             if ratio in sensitivities[name]['pruned_percent']:
                 _logger.debug('{}, {} has computed.'.format(name, ratio))
-                ratio += step_size
-                pruned_times += 1
                 continue
             if baseline is None:
                 baseline = eval_func(graph.program)
@@ -92,8 +87,6 @@ def sensitivity(program,
             for param_name in param_backup.keys():
                 param_t = scope.find_var(param_name).get_tensor()
                 param_t.set(param_backup[param_name], place)
-            ratio += step_size
-            pruned_times += 1
     return sensitivities