Add backbone_hidden_size

src/transformers/models/zoedepth/configuration_zoedepth.py

@@ -63,6 +63,8 @@ class ZoeDepthConfig(PretrainedConfig):
             - "project" passes information to the other tokens by concatenating the readout to all other tokens before
               projecting the
             representation to the original feature dimension D using a linear layer followed by a GELU non-linearity.
+        backbone_hidden_size (`int`, *optional*, defaults to 1024):
+            The hidden size of the backbone.
         reassemble_factors (`List[int]`, *optional*, defaults to `[4, 2, 1, 0.5]`):
             The up/downsampling factors of the reassemble layers.
         neck_hidden_sizes (`List[str]`, *optional*, defaults to `[96, 192, 384, 768]`):
@@ -133,6 +135,7 @@ def __init__(
         hidden_act="gelu",
         initializer_range=0.02,
         readout_type="project",
+        backbone_hidden_size=1024,
         reassemble_factors=[4, 2, 1, 0.5],
         neck_hidden_sizes=[96, 192, 384, 768],
         fusion_hidden_size=256,
@@ -198,6 +201,7 @@ def __init__(
         self.hidden_act = hidden_act
         self.use_pretrained_backbone = use_pretrained_backbone
         self.initializer_range = initializer_range
+        self.backbone_hidden_size = backbone_hidden_size
         self.readout_type = readout_type
         self.reassemble_factors = reassemble_factors
         self.neck_hidden_sizes = neck_hidden_sizes

src/transformers/models/zoedepth/modeling_zoedepth.py

@@ -109,7 +109,7 @@ def __init__(self, config):
 
         if config.readout_type == "project":
             self.readout_projects = nn.ModuleList()
-            hidden_size = config.backbone_config.hidden_size
+            hidden_size = config.backbone_hidden_size
             for _ in range(len(config.neck_hidden_sizes)):
                 self.readout_projects.append(
                     nn.Sequential(nn.Linear(2 * hidden_size, hidden_size), ACT2FN[config.hidden_act])
@@ -156,7 +156,7 @@ class ZoeDepthReassembleLayer(nn.Module):
     def __init__(self, config, channels, factor):
         super().__init__()
         # projection
-        hidden_size = config.backbone_config.hidden_size
+        hidden_size = config.backbone_hidden_size
         self.projection = nn.Conv2d(in_channels=hidden_size, out_channels=channels, kernel_size=1)
 
         # up/down sampling depending on factor
@@ -1302,7 +1302,7 @@ def forward(
         hidden_states = outputs.feature_maps
 
         _, _, height, width = pixel_values.shape
-        patch_size = self.config.backbone_config.patch_size
+        patch_size = self.backbone.config.patch_size
         patch_height = height // patch_size
         patch_width = width // patch_size
 

tests/models/zoedepth/test_modeling_zoedepth.py

@@ -56,6 +56,7 @@ def __init__(
         out_features=["stage1", "stage2"],
         apply_layernorm=False,
         reshape_hidden_states=False,
+        backbone_hidden_size=4,
         neck_hidden_sizes=[2, 2],
         fusion_hidden_size=6,
         bottleneck_features=6,
@@ -76,6 +77,7 @@ def __init__(
         self.use_labels = use_labels
         self.num_labels = num_labels
         self.is_training = is_training
+        self.backbone_hidden_size = backbone_hidden_size
         self.neck_hidden_sizes = neck_hidden_sizes
         self.fusion_hidden_size = fusion_hidden_size
         self.bottleneck_features = bottleneck_features
@@ -98,6 +100,7 @@ def get_config(self):
         return ZoeDepthConfig(
             backbone_config=self.get_backbone_config(),
             backbone=None,
+            backbone_hidden_size=self.backbone_hidden_size,
             neck_hidden_sizes=self.neck_hidden_sizes,
             fusion_hidden_size=self.fusion_hidden_size,
             bottleneck_features=self.bottleneck_features,