[Fix] Incorrect stage freeze on RIFormer Model (#1573)

* [Doc] RIFormer's README did not link to its paper properly

* Incorrect code for reproducing RIFormer 

the default value of frozen stage is set to 0, and the doc says that this will lead to no stage be frozen. But the actual case is the patch_embed will be freezed.

This may cause incorrect training, thus influencing the result.

I suggest a careful review.

configs/riformer/README.md

@@ -1,6 +1,6 @@
 # RIFormer
 
-> [RIFormer: Keep Your Vision Backbone Effective But Removing Token Mixer](https://arxiv.org/abs/xxxx.xxxxx)
+> [RIFormer: Keep Your Vision Backbone Effective But Removing Token Mixer](https://arxiv.org/abs/2304.05659)
 
 <!-- [ALGORITHM] -->
 

mmpretrain/models/backbones/riformer.py

@@ -202,7 +202,7 @@ class RIFormer(BaseBackbone):
             [stage1, downsampling, stage2, downsampling, stage3, downsampling, stage4]
             Defaults to -1, means the last stage.
         frozen_stages (int): Stages to be frozen (all param fixed).
-            Defaults to 0, which means not freezing any parameters.
+            Defaults to -1, which means not freezing any parameters.
         deploy (bool): Whether to switch the model structure to
             deployment mode. Default: False.
         init_cfg (dict, optional): Initialization config dict
@@ -259,7 +259,7 @@ class RIFormer(BaseBackbone):
                  drop_rate=0.,
                  drop_path_rate=0.,
                  out_indices=-1,
-                 frozen_stages=0,
+                 frozen_stages=-1,
                  init_cfg=None,
                  deploy=False):
 
@@ -366,7 +366,7 @@ class RIFormer(BaseBackbone):
             for param in self.patch_embed.parameters():
                 param.requires_grad = False
 
-        for i in range(self.frozen_stages):
+        for i in range(0, self.frozen_stages + 1):
             # Include both block and downsample layer.
             module = self.network[i]
             module.eval()