1116 0

configs/vit_adapter/upernet_deit_adapter_tiny_512_160k_ade20k.yml

@@ -80,10 +80,11 @@ model:
     cffn_ratio: 0.25
     deform_ratio: 1.0
     interaction_indexes: [[0, 2], [3, 5], [6, 8], [9, 11]]
+    pretrained: pretrained_model/deit_tiny_patch16_224-a1311bcf_from_torch.pdparams
   backbone_indices: [0, 1, 2, 3]
   channels: 512
   pool_scales: [1, 2, 3, 6]
   dropout_ratio: 0.1
   aux_loss: True
   aux_channels: 256
-  pretrained: pretrained_model/upernet_deit_adapter_tiny_512_160_ade20k_from_torch.pdparams
+  #pretrained: pretrained_model/upernet_deit_adapter_tiny_512_160_ade20k_from_torch.pdparams

paddleseg/models/backbones/vit_adapter.py

@@ -256,20 +256,6 @@ def __init__(self,
     def init_weight(self):
         utils.load_pretrained_model(self, self.pretrained)
 
-    def forward_features(self, x):
-        x, H, W = self.patch_embed(x)
-        cls_token = self.cls_token.expand(
-            x.shape[0], -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
-        x = paddle.concat([cls_tokens, x], axis=1)
-        x = self.pos_drop(x + self.pos_embed)
-        for blk in self.blocks:
-            x = blk(x, H, W)
-        return x
-
-    def forward(self, x):
-        x = self.forward_features(x)
-        return x
-
 
 @manager.BACKBONES.add_component
 class ViTAdapter(VisionTransformer):

paddleseg/models/layers/vit_adapter_layers.py

@@ -156,11 +156,11 @@ def _reset_parameters(self):
         for i in range(self.n_points):
             grid_init[:, :, i, :] *= i + 1
 
-        with paddle.no_grad():
-            grid_init = grid_init.reshape([-1])
-            self.sampling_offsets.bias = self.create_parameter(
-                shape=grid_init.shape,
-                default_initializer=paddle.nn.initializer.Assign(grid_init))
+        grid_init = grid_init.reshape([-1])
+        self.sampling_offsets.bias = self.create_parameter(
+            shape=grid_init.shape,
+            default_initializer=paddle.nn.initializer.Assign(grid_init))
+        self.sampling_offsets.bias.stop_gradient = True
 
         constant_init(self.attention_weights.weight, value=0.)
         constant_init(self.attention_weights.bias, value=0.)