Add the 256x256 in1k ft of the so150m, add an alternate so150m def

huggingface · Jan 18, 2025 · 3677f67 · 3677f67
1 parent 2a84d68
commit 3677f67
Showing 1 changed file with 22 additions and 2 deletions.
diff --git a/timm/models/vision_transformer.py b/timm/models/vision_transformer.py
@@ -2152,15 +2152,20 @@ def _cfg(url: str = '', **kwargs) -> Dict[str, Any]:
     'vit_base_patch16_reg4_gap_256.untrained': _cfg(
         input_size=(3, 256, 256)),
 
-    'vit_so150m_patch16_reg4_gap_384.sbb_e250_in12k_ft_in1k': _cfg(
+    'vit_so150m_patch16_reg4_gap_256.sbb_e250_in12k_ft_in1k': _cfg(
         hf_hub_id='timm/',
-        input_size=(3, 384, 384), crop_pct=1.0),
+        input_size=(3, 256, 256), crop_pct=0.95),
     'vit_so150m_patch16_reg4_gap_256.sbb_e250_in12k': _cfg(
         hf_hub_id='timm/',
         num_classes=11821,
         input_size=(3, 256, 256), crop_pct=0.95),
+    'vit_so150m_patch16_reg4_gap_384.sbb_e250_in12k_ft_in1k': _cfg(
+        hf_hub_id='timm/',
+        input_size=(3, 384, 384), crop_pct=1.0),
     'vit_so150m_patch16_reg4_map_256.untrained': _cfg(
         input_size=(3, 256, 256)),
+    'vit_so150m2_patch16_reg1_gap_256.untrained': _cfg(
+        input_size=(3, 256, 256), crop_pct=0.95),
 
     'vit_intern300m_patch14_448.ogvl_dist': _cfg(
         hf_hub_id='timm/',
@@ -3467,6 +3472,7 @@ def vit_base_patch16_reg4_gap_256(pretrained: bool = False, **kwargs) -> VisionT
 
 @register_model
 def vit_so150m_patch16_reg4_map_256(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    """ SO150M (shape optimized, but diff than paper def, optimized for GPU) """
     model_args = dict(
         patch_size=16, embed_dim=896, depth=18, num_heads=14, mlp_ratio=2.572,
         class_token=False, reg_tokens=4, global_pool='map',
@@ -3478,6 +3484,7 @@ def vit_so150m_patch16_reg4_map_256(pretrained: bool = False, **kwargs) -> Visio
 
 @register_model
 def vit_so150m_patch16_reg4_gap_256(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    """ SO150M (shape optimized, but diff than paper def, optimized for GPU) """
     model_args = dict(
         patch_size=16, embed_dim=896, depth=18, num_heads=14, mlp_ratio=2.572,
         class_token=False, reg_tokens=4, global_pool='avg', fc_norm=False,
@@ -3489,6 +3496,7 @@ def vit_so150m_patch16_reg4_gap_256(pretrained: bool = False, **kwargs) -> Visio
 
 @register_model
 def vit_so150m_patch16_reg4_gap_384(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    """ SO150M (shape optimized, but diff than paper def, optimized for GPU) """
     model_args = dict(
         patch_size=16, embed_dim=896, depth=18, num_heads=14, mlp_ratio=2.572,
         class_token=False, reg_tokens=4, global_pool='avg', fc_norm=False,
@@ -3498,6 +3506,18 @@ def vit_so150m_patch16_reg4_gap_384(pretrained: bool = False, **kwargs) -> Visio
     return model
 
 
+@register_model
+def vit_so150m2_patch16_reg1_gap_256(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    """ SO150M v2 (shape optimized, but diff than paper def, optimized for GPU) """
+    model_args = dict(
+        patch_size=16, embed_dim=896, depth=20, num_heads=14, mlp_ratio=2.429, init_values=1e-5,
+        qkv_bias=False, class_token=False, reg_tokens=1, global_pool='avg',
+    )
+    model = _create_vision_transformer(
+        'vit_so150m2_patch16_reg1_gap_256', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
+
+
 @register_model
 def vit_intern300m_patch14_448(pretrained: bool = False, **kwargs) -> VisionTransformer:
     model_args = dict(