diff --git a/fastreid/config/defaults.py b/fastreid/config/defaults.py
index d724ce0..538941b 100644
--- a/fastreid/config/defaults.py
+++ b/fastreid/config/defaults.py
@@ -83,6 +83,7 @@ _C.MODEL.HEADS.CLS_LAYER = "Linear"  # ArcSoftmax" or "CircleSoftmax"
 # Margin and Scale for margin-based classification layer
 _C.MODEL.HEADS.MARGIN = 0.
 _C.MODEL.HEADS.SCALE = 1.
+_C.MODEL.HEADS.ARC_K = 1
 
 # ---------------------------------------------------------------------------- #
 # REID LOSSES options
diff --git a/fastreid/modeling/heads/clas_head.py b/fastreid/modeling/heads/clas_head.py
index 37a6399..c11b697 100644
--- a/fastreid/modeling/heads/clas_head.py
+++ b/fastreid/modeling/heads/clas_head.py
@@ -4,6 +4,7 @@
 @contact: sherlockliao01@gmail.com
 """
 
+import torch
 import torch.nn.functional as F
 
 from fastreid.config import configurable
@@ -22,19 +23,25 @@ class ClasHead(EmbeddingHead):
             self,
             *,
             return_embedding=False,
+            arc_k=1,
             **kwargs
     ):
         """
         NOTE: this interface is experimental.
         """
-
-        super(ClasHead, self).__init__(**kwargs)
+        self.num_classes = kwargs['num_classes']
         self.return_embedding = return_embedding
+        self.arc_k = arc_k
+
+        if arc_k > 1:
+            kwargs['num_classes'] = kwargs['num_classes'] * arc_k
+        super(ClasHead, self).__init__(**kwargs)
  
     @classmethod
     def from_config(cls, cfg):
         config_dict = super(ClasHead, cls).from_config(cfg)
         config_dict['return_embedding'] = cfg.MODEL.HEADS.RETURN_EMBEDDING
+        config_dict['arc_k'] = cfg.MODEL.HEADS.ARC_K
         return config_dict
 
     def forward(self, features, targets=None):
@@ -52,6 +59,9 @@ class ClasHead(EmbeddingHead):
             logits = F.linear(neck_feat, self.weight)
         else:
             logits = F.linear(F.normalize(neck_feat), F.normalize(self.weight))
+            if self.arc_k > 1:
+                logits = torch.reshape(logits, (-1, self.num_classes, self.arc_k))
+                logits = torch.max(logits, dim=2)[0]
 
         # Evaluation
         if not self.training: return logits.mul_(self.cls_layer.s)
diff --git a/projects/Shoe/configs/subceter_arc_k3.yaml b/projects/Shoe/configs/subceter_arc_k3.yaml
new file mode 100644
index 0000000..9e52730
--- /dev/null
+++ b/projects/Shoe/configs/subceter_arc_k3.yaml
@@ -0,0 +1,18 @@
+_BASE_: base-clas.yaml
+
+MODEL:
+  BACKBONE:
+    PRETRAIN: True
+
+  HEADS:
+    ARC_K: 3
+    EMBEDDING_DIM: 256
+    SCALE: 65.
+
+
+SOLVER:
+  MAX_EPOCH: 50
+  SCHED: MultiStepLR
+  STEPS: [20, 35, 45]
+
+OUTPUT_DIR: projects/Shoe/logs/subcenter-arc-k3