suppress based privacy model, 2021.2.22

mindarmour/privacy/sup_privacy/sup_ctrl/conctrl.py

@@ -134,6 +134,7 @@ class SuppressCtrl(Cell):
         self.mask_all_steps = (end_epoch - start_epoch + 1)*batch_num  # the amount of step contained in all suppress operation
         self.mask_step_interval = self.mask_all_steps/mask_times  # the amount of step contaied in one suppress operation
         self.mask_initialized = False  # flag means the initialization is done
+        self.grad_idx_map = []
 
         if self.lr > 0.5:
             msg = "learning rate should not be greater than 0.5, but got {}".format(self.lr)
@@ -210,6 +211,8 @@ class SuppressCtrl(Cell):
                 de_weight_cell.mask_able = False
                 self.de_weight_mask_list.append(de_weight_cell)
 
+                self.grad_idx_map.append(-1)
+
             m = 0
             for layer in networks.get_parameters(expand=True):
                 one_mask_layer = None
@@ -231,6 +234,7 @@ class SuppressCtrl(Cell):
                     de_weight_cell = DeWeightInCell(add_mask_array)
                     de_weight_cell.mask_able = True
                     self.de_weight_mask_list[one_mask_layer.grad_idx] = de_weight_cell
+                    self.grad_idx_map[m] = one_mask_layer.grad_idx
                     msg = "do mask {}, {}, {}".format(m, one_mask_layer.layer_name, one_mask_layer.grad_idx)
                     LOGGER.info(TAG, msg)
                 elif one_mask_layer is not None and one_mask_layer.inited:
@@ -294,7 +298,10 @@ class SuppressCtrl(Cell):
                           math.pow((1.0 - (cur_step + 0.0 - self.mask_start_step) / self.mask_all_steps), 3)
         m = 0
         for layer in networks.get_parameters(expand=True):
-            if self.grads_mask_list[m].mask_able:
+            grad_idx = self.grad_idx_map[m]
+            if grad_idx < 0:
+                continue
+            if self.grads_mask_list[grad_idx].mask_able:
                 weight_array = layer.data.asnumpy()
                 weight_avg = np.mean(weight_array)
                 weight_array_flat = weight_array.flatten()
@@ -307,14 +314,14 @@ class SuppressCtrl(Cell):
                     msg = "give up this masking .."
                     LOGGER.info(TAG, msg)
                     return
-                if self.grads_mask_list[m].min_num > 0:
+                if self.grads_mask_list[grad_idx].min_num > 0:
                     sparse_weight_thd, _, actual_stop_pos = self.calc_sparse_thd(weight_array_flat_abs,
-                                                                                 self.cur_sparse, m)
+                                                                                 self.cur_sparse, grad_idx)
                 else:
                     actual_stop_pos = int(len_array * self.cur_sparse)
                     sparse_weight_thd = weight_array_flat_abs[actual_stop_pos]
 
-                self.update_mask_layer(weight_array_flat, sparse_weight_thd, actual_stop_pos, weight_abs_max, m)
+                self.update_mask_layer(weight_array_flat, sparse_weight_thd, actual_stop_pos, weight_abs_max, grad_idx)
 
                 msg = "{} len={}, sparse={}, current sparse thd={}, max={}, avg={}, avg_abs={} \n".format(
                     layer.name, len_array, actual_stop_pos/len_array, sparse_weight_thd,
@@ -570,8 +577,12 @@ class MaskLayerDes:
 
     Args:
         layer_name (str): Layer name, get the name of one layer as following:
-            for layer in networks.get_parameters(expand=True):
-                if layer.name == "conv": ...
+
+            .. code-block::
+
+                for layer in networks.get_parameters(expand=True):
+                    if layer.name == "conv": ...
+
         grad_idx (int): Grad layer index, get mask layer's index in grad tuple.You can refer to the construct function
             of TrainOneStepCell in mindarmour/privacy/sup_privacy/train/model.py to get the index of some specified
             grad layers (print in PYNATIVE_MODE).