fix cbnet (#785)

Add `lr_multi_list` param for SSLD pretrained models' usage.

ppdet/modeling/backbones/cb_resnet.py

@@ -65,7 +65,8 @@ class CBResNet(object):
                  feature_maps=[2, 3, 4, 5],
                  dcn_v2_stages=[],
                  nonlocal_stages=[],
-                 repeat_num=2):
+                 repeat_num=2,
+                 lr_mult_list=[1., 1., 1., 1.]):
         super(CBResNet, self).__init__()
 
         if isinstance(feature_maps, Integral):
@@ -108,6 +109,9 @@ class CBResNet(object):
             200: 12,
         }
 
+        self.lr_mult_list = lr_mult_list
+        self.stage_num = -1
+
         self.stage_filters = [64, 128, 256, 512]
         self._c1_out_chan_num = 64
         self.na = NameAdapter(self)
@@ -143,6 +147,13 @@ class CBResNet(object):
                    act=None,
                    name=None,
                    dcn=False):
+
+        # need fine lr for distilled model, default as 1.0
+        lr_mult = 1.0
+        mult_idx = max(self.stage_num - 2, 0)
+        mult_idx = min(self.stage_num - 2, 3)
+        lr_mult = self.lr_mult_list[mult_idx]
+
         if not dcn:
             conv = fluid.layers.conv2d(
                 input=input,
@@ -153,7 +164,8 @@ class CBResNet(object):
                 groups=groups,
                 act=None,
                 param_attr=ParamAttr(
-                    name=name + "_weights_" + str(self.curr_level)),
+                    name=name + "_weights_" + str(self.curr_level),
+                    learning_rate=lr_mult),
                 bias_attr=False)
         else:
             offset_mask = self._conv_offset(
@@ -182,12 +194,13 @@ class CBResNet(object):
                 deformable_groups=1,
                 im2col_step=1,
                 param_attr=ParamAttr(
-                    name=name + "_weights_" + str(self.curr_level)),
+                    name=name + "_weights_" + str(self.curr_level),
+                    learning_rate=lr_mult),
                 bias_attr=False)
 
         bn_name = self.na.fix_conv_norm_name(name)
 
-        norm_lr = 0. if self.freeze_norm else 1.
+        norm_lr = 0. if self.freeze_norm else lr_mult
         norm_decay = self.norm_decay
         pattr = ParamAttr(
             name=bn_name + '_scale_' + str(self.curr_level),
@@ -315,6 +328,8 @@ class CBResNet(object):
         """
         assert stage_num in [2, 3, 4, 5]
 
+        self.stage_num = stage_num
+
         stages, block_func = self.depth_cfg[self.depth]
         count = stages[stage_num - 2]
 
@@ -422,6 +437,7 @@ class CBResNet(object):
                 res_endpoints.append(res)
 
         for num in range(1, self.repeat_num):
+            self.stage_num = -1
             self.curr_level = num
             res = self.c1_stage(input)
             for i in range(len(res_endpoints)):