Support input of non three-channel image

Add in_chans augment and pass it to PatchEmbed layer.

mmpretrain/models/backbones/poolformer.py

@@ -220,6 +220,7 @@ class PoolFormer(BaseBackbone):
             Defaults to ``dict(type='LN2d', eps=1e-6)``.
         act_cfg (dict): The config dict for activation between pointwise
             convolution. Defaults to ``dict(type='GELU')``.
+        in_chans (int): The num of channels of input image.
         in_patch_size (int): The patch size of input image patch embedding.
             Defaults to 7.
         in_stride (int): The stride of input image patch embedding.
@@ -285,6 +286,7 @@ class PoolFormer(BaseBackbone):
                  pool_size=3,
                  norm_cfg=dict(type='GN', num_groups=1),
                  act_cfg=dict(type='GELU'),
+                 in_chans=3,
                  in_patch_size=7,
                  in_stride=4,
                  in_pad=2,
@@ -320,7 +322,7 @@ class PoolFormer(BaseBackbone):
             patch_size=in_patch_size,
             stride=in_stride,
             padding=in_pad,
-            in_chans=3,
+            in_chans=in_chans,
             embed_dim=embed_dims[0])
 
         # set the main block in network