Merge pull request #2178 from huggingface/pali_siglip

Support loading of PaliGemma weights into GAP variants of SigLIP ViT.

Author: rwightman

timm/models/_builder.py

@@ -10,7 +10,8 @@
 from timm.models._features import FeatureListNet, FeatureDictNet, FeatureHookNet, FeatureGetterNet
 from timm.models._features_fx import FeatureGraphNet
 from timm.models._helpers import load_state_dict
-from timm.models._hub import has_hf_hub, download_cached_file, check_cached_file, load_state_dict_from_hf
+from timm.models._hub import has_hf_hub, download_cached_file, check_cached_file, load_state_dict_from_hf,\
+    load_custom_from_hf
 from timm.models._manipulate import adapt_input_conv
 from timm.models._pretrained import PretrainedCfg
 from timm.models._prune import adapt_model_from_file
@@ -185,7 +186,12 @@ def load_pretrained(
     elif load_from == 'hf-hub':
         _logger.info(f'Loading pretrained weights from Hugging Face hub ({pretrained_loc})')
         if isinstance(pretrained_loc, (list, tuple)):
-            state_dict = load_state_dict_from_hf(*pretrained_loc)
+            custom_load = pretrained_cfg.get('custom_load', False)
+            if isinstance(custom_load, str) and custom_load == 'hf':
+                load_custom_from_hf(*pretrained_loc, model)
+                return
+            else:
+                state_dict = load_state_dict_from_hf(*pretrained_loc)
         else:
             state_dict = load_state_dict_from_hf(pretrained_loc)
     else:

timm/models/_hub.py

@@ -190,6 +190,13 @@ def load_state_dict_from_hf(model_id: str, filename: str = HF_WEIGHTS_NAME):
     return torch.load(cached_file, map_location='cpu')
 
 
+def load_custom_from_hf(model_id: str, filename: str, model: torch.nn.Module):
+    assert has_hf_hub(True)
+    hf_model_id, hf_revision = hf_split(model_id)
+    cached_file = hf_hub_download(hf_model_id, filename=filename, revision=hf_revision)
+    return model.load_pretrained(cached_file)
+
+
 def save_config_for_hf(
         model,
         config_path: str,

timm/models/vision_transformer.py

@@ -845,7 +845,9 @@ def _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str =
     """
     import numpy as np
 
-    def _n2p(w, t=True):
+    def _n2p(w, t=True, idx=None):
+        if idx is not None:
+            w = w[idx]
         if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1:
             w = w.flatten()
         if t:
@@ -955,21 +957,28 @@ def _n2p(w, t=True):
 
     mha_sub, b_sub, ln1_sub = (0, 0, 1) if big_vision else (1, 3, 2)
     for i, block in enumerate(model.blocks.children()):
-        block_prefix = f'{prefix}Transformer/encoderblock_{i}/'
+        if f'{prefix}Transformer/encoderblock/LayerNorm_0/scale' in w:
+            block_prefix = f'{prefix}Transformer/encoderblock/'
+            idx = i
+        else:
+            block_prefix = f'{prefix}Transformer/encoderblock_{i}/'
+            idx = None
         mha_prefix = block_prefix + f'MultiHeadDotProductAttention_{mha_sub}/'
-        block.norm1.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/scale']))
-        block.norm1.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/bias']))
+        block.norm1.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/scale'], idx=idx))
+        block.norm1.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/bias'], idx=idx))
         block.attn.qkv.weight.copy_(torch.cat([
-            _n2p(w[f'{mha_prefix}{n}/kernel'], t=False).flatten(1).T for n in ('query', 'key', 'value')]))
+            _n2p(w[f'{mha_prefix}{n}/kernel'], t=False, idx=idx).flatten(1).T for n in ('query', 'key', 'value')]))
         block.attn.qkv.bias.copy_(torch.cat([
-            _n2p(w[f'{mha_prefix}{n}/bias'], t=False).reshape(-1) for n in ('query', 'key', 'value')]))
-        block.attn.proj.weight.copy_(_n2p(w[f'{mha_prefix}out/kernel']).flatten(1))
-        block.attn.proj.bias.copy_(_n2p(w[f'{mha_prefix}out/bias']))
-        block.norm2.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_{ln1_sub}/scale']))
-        block.norm2.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_{ln1_sub}/bias']))
+            _n2p(w[f'{mha_prefix}{n}/bias'], t=False, idx=idx).reshape(-1) for n in ('query', 'key', 'value')]))
+        block.attn.proj.weight.copy_(_n2p(w[f'{mha_prefix}out/kernel'], idx=idx).flatten(1))
+        block.attn.proj.bias.copy_(_n2p(w[f'{mha_prefix}out/bias'], idx=idx))
+        block.norm2.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_{ln1_sub}/scale'], idx=idx))
+        block.norm2.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_{ln1_sub}/bias'], idx=idx))
         for r in range(2):
-            getattr(block.mlp, f'fc{r + 1}').weight.copy_(_n2p(w[f'{block_prefix}MlpBlock_{b_sub}/Dense_{r}/kernel']))
-            getattr(block.mlp, f'fc{r + 1}').bias.copy_(_n2p(w[f'{block_prefix}MlpBlock_{b_sub}/Dense_{r}/bias']))
+            getattr(block.mlp, f'fc{r + 1}').weight.copy_(
+                _n2p(w[f'{block_prefix}MlpBlock_{b_sub}/Dense_{r}/kernel'], idx=idx))
+            getattr(block.mlp, f'fc{r + 1}').bias.copy_(
+                _n2p(w[f'{block_prefix}MlpBlock_{b_sub}/Dense_{r}/bias'], idx=idx))
 
 
 def _convert_openai_clip(
@@ -1769,6 +1778,44 @@ def _cfg(url: str = '', **kwargs) -> Dict[str, Any]:
         input_size=(3, 384, 384),
         num_classes=0),
 
+    'vit_so400m_patch14_siglip_gap_224.webli': _cfg(
+        hf_hub_id='timm/ViT-SO400M-14-SigLIP',
+        hf_hub_filename='open_clip_pytorch_model.bin',
+        num_classes=0),
+    'vit_so400m_patch14_siglip_gap_224.pali_mix': _cfg(
+        hf_hub_id='google/paligemma-3b-mix-224-jax',
+        hf_hub_filename='paligemma-3b-mix-224.npz',
+        custom_load='hf',
+        num_classes=0),
+    'vit_so400m_patch14_siglip_gap_224.pali_pt': _cfg(
+        hf_hub_id='google/paligemma-3b-pt-224-jax',
+        hf_hub_filename='paligemma-3b-pt-224.npz',
+        custom_load='hf',
+        num_classes=0),
+    'vit_so400m_patch14_siglip_gap_384.webli': _cfg(
+        hf_hub_id='timm/ViT-SO400M-14-SigLIP-384',
+        hf_hub_filename='open_clip_pytorch_model.bin',
+        input_size=(3, 384, 384), crop_pct=1.0,
+        num_classes=0),
+    'vit_so400m_patch14_siglip_gap_448.pali_mix': _cfg(
+        hf_hub_id='google/paligemma-3b-mix-448-jax',
+        hf_hub_filename='paligemma-3b-mix-448.npz',
+        custom_load='hf',
+        input_size=(3, 448, 448), crop_pct=1.0,
+        num_classes=0),
+    'vit_so400m_patch14_siglip_gap_448.pali_pt': _cfg(
+        hf_hub_id='google/paligemma-3b-pt-448-jax',
+        hf_hub_filename='paligemma-3b-pt-448.npz',
+        custom_load='hf',
+        input_size=(3, 448, 448), crop_pct=1.0,
+        num_classes=0),
+    'vit_so400m_patch14_siglip_gap_896.pali_pt': _cfg(
+        hf_hub_id='google/paligemma-3b-pt-896-jax',
+        hf_hub_filename='paligemma-3b-pt-896.npz',
+        custom_load='hf',
+        input_size=(3, 896, 896), crop_pct=1.0,
+        num_classes=0),
+
     'vit_xsmall_patch16_clip_224.tinyclip_yfcc15m': _cfg(
         hf_hub_id='timm/',
         hf_hub_filename='open_clip_pytorch_model.bin',
@@ -2756,15 +2803,48 @@ def vit_so400m_patch14_siglip_384(pretrained: bool = False, **kwargs) -> VisionT
     return model
 
 
-# @register_model
-# def vit_medium_patch16_reg4_256(pretrained: bool = False, **kwargs) -> VisionTransformer:
-#     model_args = dict(
-#         patch_size=16, embed_dim=512, depth=12, num_heads=8, class_token=True,
-#         no_embed_class=True, reg_tokens=4,
-#     )
-#     model = _create_vision_transformer(
-#         'vit_medium_patch16_reg4_256', pretrained=pretrained, **dict(model_args, **kwargs))
-#     return model
+@register_model
+def vit_so400m_patch14_siglip_gap_224(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    model_args = dict(
+        patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362,
+        class_token=False, global_pool='avg', fc_norm=False,
+    )
+    model = _create_vision_transformer(
+        'vit_so400m_patch14_siglip_gap_224', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
+
+
+@register_model
+def vit_so400m_patch14_siglip_gap_384(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    model_args = dict(
+        patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362,
+        class_token=False, global_pool='avg', fc_norm=False,
+    )
+    model = _create_vision_transformer(
+        'vit_so400m_patch14_siglip_gap_384', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
+
+
+@register_model
+def vit_so400m_patch14_siglip_gap_448(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    model_args = dict(
+        patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362,
+        class_token=False, global_pool='avg', fc_norm=False,
+    )
+    model = _create_vision_transformer(
+        'vit_so400m_patch14_siglip_gap_448', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
+
+
+@register_model
+def vit_so400m_patch14_siglip_gap_896(pretrained: bool = False, **kwargs) -> VisionTransformer:
+    model_args = dict(
+        patch_size=14, embed_dim=1152, depth=27, num_heads=16, mlp_ratio=3.7362,
+        class_token=False, global_pool='avg', fc_norm=False,
+    )
+    model = _create_vision_transformer(
+        'vit_so400m_patch14_siglip_gap_896', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
 
 
 @register_model