mirrors/pytorch-image-models
1
0
Fork 0
mirror of https://github.com/huggingface/pytorch-image-models.git synced 2025-06-03 15:01:08 +08:00
Code Issues Projects Releases Wiki Activity

Add explicit GAP (avg pool) variants of other SigLIP models.

Browse Source
This commit is contained in:
Ross Wightman 2024-05-15 07:53:19 -07:00
parent f04802fb8d
commit 83aee5c28c
1 changed files with 99 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

99
timm/models/vision_transformer.py
View File

@ -1778,6 +1778,35 @@ default_cfgs = {
input_size=(3, 384, 384), input_size=(3, 384, 384),
num_classes=0), num_classes=0),
'vit_base_patch16_siglip_gap_224.webli': _cfg(
hf_hub_id='timm/ViT-B-16-SigLIP',
hf_hub_filename='open_clip_pytorch_model.bin',
num_classes=0),
'vit_base_patch16_siglip_gap_256.webli': _cfg(
hf_hub_id='timm/ViT-B-16-SigLIP-256',
hf_hub_filename='open_clip_pytorch_model.bin',
input_size=(3, 256, 256),
num_classes=0),
'vit_base_patch16_siglip_gap_384.webli': _cfg(
hf_hub_id='timm/ViT-B-16-SigLIP-384',
hf_hub_filename='open_clip_pytorch_model.bin',
input_size=(3, 384, 384),
num_classes=0),
'vit_base_patch16_siglip_gap_512.webli': _cfg(
hf_hub_id='timm/ViT-B-16-SigLIP-512',
hf_hub_filename='open_clip_pytorch_model.bin',
input_size=(3, 512, 512),
num_classes=0),
'vit_large_patch16_siglip_gap_256.webli': _cfg(
hf_hub_id='timm/ViT-L-16-SigLIP-256',
hf_hub_filename='open_clip_pytorch_model.bin',
input_size=(3, 256, 256),
num_classes=0),
'vit_large_patch16_siglip_gap_384.webli': _cfg(
hf_hub_id='timm/ViT-L-16-SigLIP-384',
hf_hub_filename='open_clip_pytorch_model.bin',
input_size=(3, 384, 384),
num_classes=0),
'vit_so400m_patch14_siglip_gap_224.webli': _cfg( 'vit_so400m_patch14_siglip_gap_224.webli': _cfg(
hf_hub_id='timm/ViT-SO400M-14-SigLIP', hf_hub_id='timm/ViT-SO400M-14-SigLIP',
hf_hub_filename='open_clip_pytorch_model.bin', hf_hub_filename='open_clip_pytorch_model.bin',
@ -2803,8 +2832,75 @@ def vit_so400m_patch14_siglip_384(pretrained: bool = False, **kwargs) -> VisionT
return model return model
@register_model
def vit_base_patch16_siglip_gap_224(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict(
patch_size=16, embed_dim=768, depth=12, num_heads=12, class_token=False, global_pool='avg', fc_norm=False,
)
model = _create_vision_transformer(
'vit_base_patch16_siglip_gap_224', pretrained=pretrained, **dict(model_args, **kwargs))
return model
@register_model
def vit_base_patch16_siglip_gap_256(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict(
patch_size=16, embed_dim=768, depth=12, num_heads=12, class_token=False, global_pool='avg', fc_norm=False,
)
model = _create_vision_transformer(
'vit_base_patch16_siglip_gap_256', pretrained=pretrained, **dict(model_args, **kwargs))
return model
@register_model
def vit_base_patch16_siglip_gap_384(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict(
patch_size=16, embed_dim=768, depth=12, num_heads=12, class_token=False, global_pool='avg', fc_norm=False,
)
model = _create_vision_transformer(
'vit_base_patch16_siglip_gap_384', pretrained=pretrained, **dict(model_args, **kwargs))
return model
@register_model
def vit_base_patch16_siglip_gap_512(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict(
patch_size=16, embed_dim=768, depth=12, num_heads=12, class_token=False, global_pool='avg', fc_norm=False,
)
model = _create_vision_transformer(
'vit_base_patch16_siglip_gap_512', pretrained=pretrained, **dict(model_args, **kwargs))
return model
@register_model
def vit_large_patch16_siglip_gap_256(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict(
patch_size=16, embed_dim=1024, depth=24, num_heads=16, class_token=False, global_pool='avg', fc_norm=False,
)
model = _create_vision_transformer(
'vit_large_patch16_siglip_gap_256', pretrained=pretrained, **dict(model_args, **kwargs))
return model
@register_model
def vit_large_patch16_siglip_gap_384(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict(
patch_size=16, embed_dim=1024, depth=24, num_heads=16, class_token=False, global_pool='avg', fc_norm=False,
)
model = _create_vision_transformer(
'vit_large_patch16_siglip_gap_384', pretrained=pretrained, **dict(model_args, **kwargs))
return model
@register_model @register_model
def vit_so400m_patch14_siglip_gap_224(pretrained: bool = False, **kwargs) -> VisionTransformer: def vit_so400m_patch14_siglip_gap_224(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict( model_args = dict(
patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362, patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362,
class_token=False, global_pool='avg', fc_norm=False, class_token=False, global_pool='avg', fc_norm=False,
@ -2816,6 +2912,7 @@ def vit_so400m_patch14_siglip_gap_224(pretrained: bool = False, **kwargs) -> Vis
@register_model @register_model
def vit_so400m_patch14_siglip_gap_384(pretrained: bool = False, **kwargs) -> VisionTransformer: def vit_so400m_patch14_siglip_gap_384(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict( model_args = dict(
patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362, patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362,
class_token=False, global_pool='avg', fc_norm=False, class_token=False, global_pool='avg', fc_norm=False,
@ -2827,6 +2924,7 @@ def vit_so400m_patch14_siglip_gap_384(pretrained: bool = False, **kwargs) -> Vis
@register_model @register_model
def vit_so400m_patch14_siglip_gap_448(pretrained: bool = False, **kwargs) -> VisionTransformer: def vit_so400m_patch14_siglip_gap_448(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict( model_args = dict(
patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362, patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362,
class_token=False, global_pool='avg', fc_norm=False, class_token=False, global_pool='avg', fc_norm=False,
@ -2838,6 +2936,7 @@ def vit_so400m_patch14_siglip_gap_448(pretrained: bool = False, **kwargs) -> Vis
@register_model @register_model
def vit_so400m_patch14_siglip_gap_896(pretrained: bool = False, **kwargs) -> VisionTransformer: def vit_so400m_patch14_siglip_gap_896(pretrained: bool = False, **kwargs) -> VisionTransformer:
""" A SigLIP variant of ViT with global average pooling (GAP) instead of attention pooling (MAP)."""
model_args = dict( model_args = dict(
patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362, patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362,
class_token=False, global_pool='avg', fc_norm=False, class_token=False, global_pool='avg', fc_norm=False,