Add intern300m vit w/ converted timm weights. Fix #2300

timm/models/davit.py

@@ -710,10 +710,12 @@ def checkpoint_filter_fn(state_dict, model):
 def _create_davit(variant, pretrained=False, **kwargs):
     default_out_indices = tuple(i for i, _ in enumerate(kwargs.get('depths', (1, 1, 3, 1))))
     out_indices = kwargs.pop('out_indices', default_out_indices)
-    strict = True
+
+    strict = kwargs.pop('pretrained_strict', True)
     if variant.endswith('_fl'):
         # FIXME cleaner approach to missing head norm?
         strict = False
+
     model = build_model_with_cfg(
         DaVit,
         variant,

timm/models/vision_transformer.py

@@ -438,6 +438,7 @@ class VisionTransformer(nn.Module):
             no_embed_class: bool = False,
             reg_tokens: int = 0,
             pre_norm: bool = False,
+            final_norm: bool = True,
             fc_norm: Optional[bool] = None,
             dynamic_img_size: bool = False,
             dynamic_img_pad: bool = False,
@@ -471,7 +472,9 @@ class VisionTransformer(nn.Module):
             class_token: Use class token.
             no_embed_class: Don't include position embeddings for class (or reg) tokens.
             reg_tokens: Number of register tokens.
-            fc_norm: Pre head norm after pool (instead of before), if None, enabled when global_pool == 'avg'.
+            pre_norm: Enable norm after embeddings, before transformer blocks (standard in CLIP ViT).
+            final_norm: Enable norm after transformer blocks, before head (standard in most ViT).
+            fc_norm: Move final norm after pool (instead of before), if None, enabled when global_pool == 'avg'.
             drop_rate: Head dropout rate.
             pos_drop_rate: Position embedding dropout rate.
             attn_drop_rate: Attention dropout rate.
@@ -554,7 +557,7 @@ class VisionTransformer(nn.Module):
             for i in range(depth)])
         self.feature_info = [
             dict(module=f'blocks.{i}', num_chs=embed_dim, reduction=reduction) for i in range(depth)]
-        self.norm = norm_layer(embed_dim) if not use_fc_norm else nn.Identity()
+        self.norm = norm_layer(embed_dim) if final_norm and not use_fc_norm else nn.Identity()
 
         # Classifier Head
         if global_pool == 'map':
@@ -566,7 +569,7 @@ class VisionTransformer(nn.Module):
             )
         else:
             self.attn_pool = None
-        self.fc_norm = norm_layer(embed_dim) if use_fc_norm else nn.Identity()
+        self.fc_norm = norm_layer(embed_dim) if final_norm and use_fc_norm else nn.Identity()
         self.head_drop = nn.Dropout(drop_rate)
         self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
 
@@ -2051,6 +2054,12 @@ default_cfgs = {
     'vit_so150m_patch16_reg4_map_256.untrained': _cfg(
         input_size=(3, 256, 256)),
 
+    'vit_intern300m_patch14_448.ogvl_dist': _cfg(
+        hf_hub_id='timm/',
+        mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD,
+        input_size=(3, 448, 448), crop_pct=1.0, num_classes=0,
+    ),
+
     'test_vit.r160_in1k': _cfg(
         hf_hub_id='timm/',
         input_size=(3, 160, 160), crop_pct=0.95),
@@ -2091,7 +2100,7 @@ def _create_vision_transformer(variant: str, pretrained: bool = False, **kwargs)
         _filter_fn = checkpoint_filter_fn
 
     # FIXME attn pool (currently only in siglip) params removed if pool disabled, is there a better soln?
-    strict = True
+    strict = kwargs.pop('pretrained_strict', True)
     if 'siglip' in variant and kwargs.get('global_pool', None) != 'map':
         strict = False
 
@@ -3298,6 +3307,17 @@ def vit_so150m_patch16_reg4_gap_256(pretrained: bool = False, **kwargs) -> Visio
     return model
 
 
+@register_model
+def vit_intern300m_patch14_448(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    model_args = dict(
+        patch_size=14, embed_dim=1024, depth=24, num_heads=16,
+        init_values=0.1, final_norm=False, dynamic_img_size=True,
+    )
+    model = _create_vision_transformer(
+        'vit_intern300m_patch14_448', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
+
+
 @register_model
 def test_vit(pretrained: bool = False, **kwargs) -> VisionTransformer:
     """ ViT Test