Tweak so150m2 def

rwightman · rwightman · commit aa333079daf7 · 2025-01-19T13:40:53.000-08:00
diff --git a/timm/models/vision_transformer.py b/timm/models/vision_transformer.py
@@ -3510,7 +3510,7 @@ def vit_so150m_patch16_reg4_gap_384(pretrained: bool = False, **kwargs) -> Visio
 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,
+        patch_size=16, embed_dim=832, depth=21, num_heads=13, mlp_ratio=34/13, init_values=1e-5,
         qkv_bias=False, class_token=False, reg_tokens=1, global_pool='avg',
     )
     model = _create_vision_transformer(

Original file line number	Diff line number	Diff line change
`@@ -3510,7 +3510,7 @@ def vit_so150m_patch16_reg4_gap_384(pretrained: bool = False, **kwargs) -> Visio`
`3510`	`3510`	`def vit_so150m2_patch16_reg1_gap_256(pretrained: bool = False, **kwargs) -> VisionTransformer:`
`3511`	`3511`	`""" SO150M v2 (shape optimized, but diff than paper def, optimized for GPU) """`
`3512`	`3512`	`model_args = dict(`
`3513`		`- patch_size=16, embed_dim=896, depth=20, num_heads=14, mlp_ratio=2.429, init_values=1e-5,`
	`3513`	`+ patch_size=16, embed_dim=832, depth=21, num_heads=13, mlp_ratio=34/13, init_values=1e-5,`
`3514`	`3514`	`qkv_bias=False, class_token=False, reg_tokens=1, global_pool='avg',`
`3515`	`3515`	`)`
`3516`	`3516`	`model = _create_vision_transformer(`