dbg

ppcls/arch/gears/bnneck.py

@@ -8,8 +8,8 @@ class BNNeck(paddle.nn.Layer):
 
         self.bn = paddle.nn.BatchNorm1D(
             self.num_filters)
-        if not trainable:
-            self.bn.bias.trainable = False
+        # if not trainable:
+        #     self.bn.bias.trainable = False
 
     def forward(self, input, label=None):
         out = self.bn(input)

ppcls/arch/gears/fc.py

@@ -21,14 +21,14 @@ import paddle.nn as nn
 
 
 class FC(nn.Layer):
-    def __init__(self, embedding_size, class_num):
+    def __init__(self, embedding_size, class_num, bias_attr=None):
         super(FC, self).__init__()
         self.embedding_size = embedding_size
         self.class_num = class_num
         weight_attr = paddle.ParamAttr(
             initializer=paddle.nn.initializer.XavierNormal())
         self.fc = paddle.nn.Linear(
-            self.embedding_size, self.class_num, weight_attr=weight_attr)
+            self.embedding_size, self.class_num, weight_attr=weight_attr, bias_attr=bias_attr)
 
     def forward(self, input, label=None):
         out = self.fc(input)

ppcls/configs/Pedestrian/strong_baseline_m1.yaml

@@ -26,14 +26,14 @@ Arch:
     stem_act: null
   BackboneStopLayer:
     name: "flatten"
-  #Neck:
-  #  name: BNNeck
-  #  num_filters: 2048
-  #  trainable: false
+  Neck:
+    name: BNNeck
+    num_filters: 2048
   Head:
     name: "FC"
     embedding_size: 2048
     class_num: 751
+    bias_attr: false
 
 # loss function config for traing/eval process
 Loss:

ppcls/engine/evaluation/retrieval.py

@@ -125,7 +125,7 @@ def cal_feature(engine, name='gallery'):
         out = engine.model(batch[0], batch[1])
         if "Student" in out:
             out = out["Student"]
-        batch_feas = out["features"]
+        batch_feas = out["backbone"]
 
         # do norm
         if engine.config["Global"].get("feature_normalize", True):

ppcls/loss/triplet.py

@@ -24,7 +24,7 @@ class TripletLossV2(nn.Layer):
             inputs: feature matrix with shape (batch_size, feat_dim)
             target: ground truth labels with shape (num_classes)
         """
-        inputs = input["features"]
+        inputs = input["backbone"]
 
         if self.normalize_feature:
             inputs = 1. * inputs / (paddle.expand_as(