argus.py: reflect-padded DPT convs, drop post-hoc border crop

argus.py

@@ -1429,9 +1429,9 @@ N_PREFIX_TOKENS = 5  # 1 CLS + 4 register/storage tokens
 class _ResidualConvUnit(nn.Module):
     def __init__(self, dim: int):
         super().__init__()
-        self.conv1 = nn.Conv2d(dim, dim, 3, padding=1, bias=False)
+        self.conv1 = nn.Conv2d(dim, dim, 3, padding=1, padding_mode="reflect", bias=False)
         self.bn1 = nn.BatchNorm2d(dim)
-        self.conv2 = nn.Conv2d(dim, dim, 3, padding=1, bias=False)
+        self.conv2 = nn.Conv2d(dim, dim, 3, padding=1, padding_mode="reflect", bias=False)
         self.bn2 = nn.BatchNorm2d(dim)
         self.act = nn.GELU()
 
@@ -1463,7 +1463,7 @@ class _DPTReassemble(nn.Module):
         ])
         self.refine = nn.ModuleList([
             nn.Sequential(
-                nn.Conv2d(out_dim, out_dim, 3, padding=1, bias=False),
+                nn.Conv2d(out_dim, out_dim, 3, padding=1, padding_mode="reflect", bias=False),
                 nn.BatchNorm2d(out_dim),
                 nn.GELU(),
             )
@@ -1502,7 +1502,7 @@ class DPTDepthDecoder(nn.Module):
             _FeatureFusionBlock(decoder_dim, has_skip=False),
         ])
         self.head = nn.Sequential(
-            nn.Conv2d(decoder_dim, decoder_dim, 3, padding=1, bias=False),
+            nn.Conv2d(decoder_dim, decoder_dim, 3, padding=1, padding_mode="reflect", bias=False),
             nn.BatchNorm2d(decoder_dim),
             nn.GELU(),
             nn.Conv2d(decoder_dim, n_bins, 1),
@@ -1790,13 +1790,8 @@ class Argus(PreTrainedModel):
             depth_b = self.depth_head(inter_list, H, W)
             std_b = None
 
-        # Crop the DPT fusion border artifact (zero-padding in the conv chain
-        # produces systematically wrong edge values that compound across 4 stages)
-        crop = max(4, depth_b.shape[2] // 13)
-        depth_b = depth_b[:, :, crop:-crop, crop:-crop]
         depth_b = F.interpolate(depth_b, size=(resolution, resolution), mode="bilinear", align_corners=False)
         if std_b is not None:
-            std_b = std_b[:, :, crop:-crop, crop:-crop]
             std_b = F.interpolate(std_b, size=(resolution, resolution), mode="bilinear", align_corners=False)
 
         depth_squeezed = depth_b[:, 0].float()