Adding support for fine-tune CLIP LAION-2B image tower weights for B/32, L/14, H/14 and g/14. Still WIP

tests/test_models.py

@@ -38,7 +38,7 @@ if 'GITHUB_ACTIONS' in os.environ:
     EXCLUDE_FILTERS = [
         '*efficientnet_l2*', '*resnext101_32x48d', '*in21k', '*152x4_bitm', '*101x3_bitm', '*50x3_bitm',
         '*nfnet_f3*', '*nfnet_f4*', '*nfnet_f5*', '*nfnet_f6*', '*nfnet_f7*', '*efficientnetv2_xl*',
-        '*resnetrs350*', '*resnetrs420*', 'xcit_large_24_p8*', 'vit_huge*', 'vit_gi*', 'swin*huge*',
+        '*resnetrs350*', '*resnetrs420*', 'xcit_large_24_p8*', 'vit_huge*', 'vit_g*', 'swin*huge*',
         'swin*giant*']
     NON_STD_EXCLUDE_FILTERS = ['vit_huge*', 'vit_gi*', 'swin*giant*']
 else:

timm/models/helpers.py

@@ -138,6 +138,9 @@ def _resolve_pretrained_source(pretrained_cfg):
             # hf-hub available as alternate weight source in default_cfg
             load_from = 'hf-hub'
             pretrained_loc = hf_hub_id
+    if load_from == 'hf-hub' and 'hf_hub_filename' in pretrained_cfg:
+        # if a filename override is set, return tuple for location w/ (hub_id, filename)
+        pretrained_loc = pretrained_loc, pretrained_cfg['hf_hub_filename']
     return load_from, pretrained_loc
 
 
@@ -246,7 +249,10 @@ def load_pretrained(
             pretrained_loc, map_location='cpu', progress=_DOWNLOAD_PROGRESS, check_hash=_CHECK_HASH)
     elif load_from == 'hf-hub':
         _logger.info(f'Loading pretrained weights from Hugging Face hub ({pretrained_loc})')
-        state_dict = load_state_dict_from_hf(pretrained_loc)
+        if isinstance(pretrained_loc, (list, tuple)):
+            state_dict = load_state_dict_from_hf(*pretrained_loc)
+        else:
+            state_dict = load_state_dict_from_hf(pretrained_loc)
     else:
         _logger.warning("No pretrained weights exist or were found for this model. Using random initialization.")
         return

timm/models/hub.py

@@ -55,7 +55,7 @@ def download_cached_file(url, check_hash=True, progress=False):
 
 def has_hf_hub(necessary=False):
     if not _has_hf_hub and necessary:
-        # if no HF Hub module installed and it is necessary to continue, raise error
+        # if no HF Hub module installed, and it is necessary to continue, raise error
         raise RuntimeError(
             'Hugging Face hub model specified but package not installed. Run `pip install huggingface_hub`.')
     return _has_hf_hub
@@ -78,7 +78,7 @@ def load_cfg_from_json(json_file: Union[str, os.PathLike]):
 
 def _download_from_hf(model_id: str, filename: str):
     hf_model_id, hf_revision = hf_split(model_id)
-    return hf_hub_download(hf_model_id, filename, revision=hf_revision, cache_dir=get_cache_dir('hf'))
+    return hf_hub_download(hf_model_id, filename, revision=hf_revision)
 
 
 def load_model_config_from_hf(model_id: str):
@@ -91,9 +91,9 @@ def load_model_config_from_hf(model_id: str):
     return pretrained_cfg, model_name
 
 
-def load_state_dict_from_hf(model_id: str):
+def load_state_dict_from_hf(model_id: str, filename: str = 'pytorch_model.bin'):
     assert has_hf_hub(True)
-    cached_file = _download_from_hf(model_id, 'pytorch_model.bin')
+    cached_file = _download_from_hf(model_id, filename)
     state_dict = torch.load(cached_file, map_location='cpu')
     return state_dict
 

timm/models/layers/patch_embed.py

@@ -15,7 +15,16 @@ from .trace_utils import _assert
 class PatchEmbed(nn.Module):
     """ 2D Image to Patch Embedding
     """
-    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True):
+    def __init__(
+            self,
+            img_size=224,
+            patch_size=16,
+            in_chans=3,
+            embed_dim=768,
+            norm_layer=None,
+            flatten=True,
+            bias=True,
+    ):
         super().__init__()
         img_size = to_2tuple(img_size)
         patch_size = to_2tuple(patch_size)
@@ -25,7 +34,7 @@ class PatchEmbed(nn.Module):
         self.num_patches = self.grid_size[0] * self.grid_size[1]
         self.flatten = flatten
 
-        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
         self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
 
     def forward(self, x):

timm/models/vision_transformer.py

@@ -25,6 +25,7 @@ from functools import partial
 from collections import OrderedDict
 from typing import Optional
 
+import huggingface_hub.file_download
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -106,7 +107,7 @@ default_cfgs = {
     'vit_large_patch14_224': _cfg(url=''),
     'vit_huge_patch14_224': _cfg(url=''),
     'vit_giant_patch14_224': _cfg(url=''),
-    'vit_gigantic_patch14_224': _cfg(url=''),
+    'vit_gee_patch14_224': _cfg(url=''),
 
 
     # patch models, imagenet21k (weights from official Google JAX impl)
@@ -177,6 +178,20 @@ default_cfgs = {
     'vit_small_patch16_36x1_224': _cfg(url=''),
     'vit_small_patch16_18x2_224': _cfg(url=''),
     'vit_base_patch16_18x2_224': _cfg(url=''),
+
+    'vit_base_patch32_224_clip_laion2b': _cfg(
+        hf_hub_id='',
+        num_classes=512),
+    'vit_large_patch14_224_clip_laion2b': _cfg(
+        hf_hub_id='',
+        num_classes=768),
+    'vit_huge_patch14_224_clip_laion2b': _cfg(
+        hf_hub_id='',
+        num_classes=1024),
+    'vit_giant_patch14_224_clip_laion2b': _cfg(
+        hf_hub_id='',
+        num_classes=1024),
+
 }
 
 
@@ -221,8 +236,18 @@ class LayerScale(nn.Module):
 class Block(nn.Module):
 
     def __init__(
-            self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0., init_values=None,
+            self,
-            drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+            dim,
+            num_heads,
+            mlp_ratio=4.,
+            qkv_bias=False,
+            drop=0.,
+            attn_drop=0.,
+            init_values=None,
+            drop_path=0.,
+            act_layer=nn.GELU,
+            norm_layer=nn.LayerNorm
+    ):
         super().__init__()
         self.norm1 = norm_layer(dim)
         self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)
@@ -244,8 +269,18 @@ class Block(nn.Module):
 class ResPostBlock(nn.Module):
 
     def __init__(
-            self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0., init_values=None,
+            self,
-            drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+            dim,
+            num_heads,
+            mlp_ratio=4.,
+            qkv_bias=False,
+            drop=0.,
+            attn_drop=0.,
+            init_values=None,
+            drop_path=0.,
+            act_layer=nn.GELU,
+            norm_layer=nn.LayerNorm
+    ):
         super().__init__()
         self.init_values = init_values
 
@@ -274,8 +309,19 @@ class ResPostBlock(nn.Module):
 class ParallelBlock(nn.Module):
 
     def __init__(
-            self, dim, num_heads, num_parallel=2, mlp_ratio=4., qkv_bias=False, init_values=None,
+            self,
-            drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+            dim,
+            num_heads,
+            num_parallel=2,
+            mlp_ratio=4.,
+            qkv_bias=False,
+            init_values=None,
+            drop=0.,
+            attn_drop=0.,
+            drop_path=0.,
+            act_layer=nn.GELU,
+            norm_layer=nn.LayerNorm
+    ):
         super().__init__()
         self.num_parallel = num_parallel
         self.attns = nn.ModuleList()
@@ -320,10 +366,31 @@ class VisionTransformer(nn.Module):
     """
 
     def __init__(
-            self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, global_pool='token',
+            self,
-            embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, init_values=None,
+            img_size=224,
-            class_token=True, no_embed_class=False, fc_norm=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0.,
+            patch_size=16,
-            weight_init='', embed_layer=PatchEmbed, norm_layer=None, act_layer=None, block_fn=Block):
+            in_chans=3,
+            num_classes=1000,
+            global_pool='token',
+            embed_dim=768,
+            depth=12,
+            num_heads=12,
+            mlp_ratio=4.,
+            qkv_bias=True,
+            init_values=None,
+            class_token=True,
+            no_embed_class=False,
+            pre_norm=False,
+            fc_norm=None,
+            drop_rate=0.,
+            attn_drop_rate=0.,
+            drop_path_rate=0.,
+            weight_init='',
+            embed_layer=PatchEmbed,
+            norm_layer=None,
+            act_layer=None,
+            block_fn=Block,
+    ):
         """
         Args:
             img_size (int, tuple): input image size
@@ -362,19 +429,34 @@ class VisionTransformer(nn.Module):
         self.grad_checkpointing = False
 
         self.patch_embed = embed_layer(
-            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+            bias=not pre_norm,  # disable bias if pre-norm is used (e.g. CLIP)
+        )
         num_patches = self.patch_embed.num_patches
 
         self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if class_token else None
         embed_len = num_patches if no_embed_class else num_patches + self.num_prefix_tokens
         self.pos_embed = nn.Parameter(torch.randn(1, embed_len, embed_dim) * .02)
         self.pos_drop = nn.Dropout(p=drop_rate)
+        self.norm_pre = norm_layer(embed_dim) if pre_norm else nn.Identity()
 
         dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
         self.blocks = nn.Sequential(*[
             block_fn(
-                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, init_values=init_values,
+                dim=embed_dim,
-                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, act_layer=act_layer)
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                init_values=init_values,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[i],
+                norm_layer=norm_layer,
+                act_layer=act_layer
+            )
             for i in range(depth)])
         self.norm = norm_layer(embed_dim) if not use_fc_norm else nn.Identity()
 
@@ -445,6 +527,7 @@ class VisionTransformer(nn.Module):
     def forward_features(self, x):
         x = self.patch_embed(x)
         x = self._pos_embed(x)
+        x = self.norm_pre(x)
         if self.grad_checkpointing and not torch.jit.is_scripting():
             x = checkpoint_seq(self.blocks, x)
         else:
@@ -623,6 +706,40 @@ def resize_pos_embed(posemb, posemb_new, num_prefix_tokens=1, gs_new=()):
     return posemb
 
 
+def _convert_openai_clip(state_dict, model):
+    out_dict = {}
+    swaps = [
+        ('visual.', ''), ('conv1', 'patch_embed.proj'), ('positional_embedding', 'pos_embed'),
+        ('transformer.resblocks.', 'blocks.'), ('ln_pre', 'norm_pre'), ('ln_post', 'norm'), ('ln_', 'norm'),
+        ('in_proj_', 'qkv.'), ('out_proj', 'proj'), ('mlp.c_fc', 'mlp.fc1'), ('mlp.c_proj', 'mlp.fc2'),
+    ]
+    for k, v in state_dict.items():
+        if not k.startswith('visual.'):
+            continue
+        for sp in swaps:
+            k = k.replace(sp[0], sp[1])
+
+        if k == 'proj':
+            k = 'head.weight'
+            v = v.transpose(0, 1)
+            out_dict['head.bias'] = torch.zeros(v.shape[0])
+        elif k == 'class_embedding':
+            k = 'cls_token'
+            v = v.unsqueeze(0).unsqueeze(1)
+        elif k == 'pos_embed':
+            v = v.unsqueeze(0)
+            if v.shape[1] != model.pos_embed.shape[1]:
+                # To resize pos embedding when using model at different size from pretrained weights
+                v = resize_pos_embed(
+                    v,
+                    model.pos_embed,
+                    0 if getattr(model, 'no_embed_class') else getattr(model, 'num_prefix_tokens', 1),
+                    model.patch_embed.grid_size
+                )
+        out_dict[k] = v
+    return out_dict
+
+
 def checkpoint_filter_fn(state_dict, model, adapt_layer_scale=False):
     """ convert patch embedding weight from manual patchify + linear proj to conv"""
     import re
@@ -631,6 +748,9 @@ def checkpoint_filter_fn(state_dict, model, adapt_layer_scale=False):
         # For deit models
         state_dict = state_dict['model']
 
+    if 'visual.class_embedding' in state_dict:
+        return _convert_openai_clip(state_dict, model)
+
     for k, v in state_dict.items():
         if 'patch_embed.proj.weight' in k and len(v.shape) < 4:
             # For old models that I trained prior to conv based patchification
@@ -833,7 +953,7 @@ def vit_huge_patch14_224(pretrained=False, **kwargs):
 
 @register_model
 def vit_giant_patch14_224(pretrained=False, **kwargs):
-    """ ViT-Giant model (ViT-g/14) from `Scaling Vision Transformers` - https://arxiv.org/abs/2106.04560
+    """ ViT-Giant (little-g) model (ViT-g/14) from `Scaling Vision Transformers` - https://arxiv.org/abs/2106.04560
     """
     model_kwargs = dict(patch_size=14, embed_dim=1408, mlp_ratio=48/11, depth=40, num_heads=16, **kwargs)
     model = _create_vision_transformer('vit_giant_patch14_224', pretrained=pretrained, **model_kwargs)
@@ -841,11 +961,12 @@ def vit_giant_patch14_224(pretrained=False, **kwargs):
 
 
 @register_model
-def vit_gigantic_patch14_224(pretrained=False, **kwargs):
+def vit_gee_patch14_224(pretrained=False, **kwargs):
-    """ ViT-Gigantic model (ViT-G/14) from `Scaling Vision Transformers` - https://arxiv.org/abs/2106.04560
+    """ ViT-GEE (big-G) model (ViT-G/14) from `Scaling Vision Transformers` - https://arxiv.org/abs/2106.04560
+    As per https://twitter.com/wightmanr/status/1570549064667889666
     """
     model_kwargs = dict(patch_size=14, embed_dim=1664, mlp_ratio=64/13, depth=48, num_heads=16, **kwargs)
-    model = _create_vision_transformer('vit_gigantic_patch14_224', pretrained=pretrained, **model_kwargs)
+    model = _create_vision_transformer('vit_gee_patch14_224', pretrained=pretrained, **model_kwargs)
     return model
 
 
@@ -1085,3 +1206,44 @@ def vit_base_patch16_18x2_224(pretrained=False, **kwargs):
         patch_size=16, embed_dim=768, depth=18, num_heads=12, init_values=1e-5, block_fn=ParallelBlock, **kwargs)
     model = _create_vision_transformer('vit_base_patch16_18x2_224', pretrained=pretrained, **model_kwargs)
     return model
+
+
+@register_model
+def vit_base_patch32_224_clip_laion2b(pretrained=False, **kwargs):
+    """ ViT-B/32
+    Pretrained weights from CLIP image tower trained on LAION-2B image-text pairs.
+    """
+    model_kwargs = dict(patch_size=32, embed_dim=768, depth=12, num_heads=12, pre_norm=True, **kwargs)
+    model = _create_vision_transformer('vit_base_patch32_224_clip_laion2b', pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_large_patch14_224_clip_laion2b(pretrained=False, **kwargs):
+    """ ViT-Large model (ViT-L/14)
+    Pretrained weights from CLIP image tower trained on LAION-2B image-text pairs.
+    """
+    model_kwargs = dict(patch_size=14, embed_dim=1024, depth=24, num_heads=16, pre_norm=True, **kwargs)
+    model = _create_vision_transformer('vit_large_patch14_224_clip_laion2b', pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_huge_patch14_224_clip_laion2b(pretrained=False, **kwargs):
+    """ ViT-Huge model (ViT-H/14) from original paper (https://arxiv.org/abs/2010.11929).
+    Pretrained weights from CLIP image tower trained on LAION-2B image-text pairs.
+    """
+    model_kwargs = dict(patch_size=14, embed_dim=1280, depth=32, num_heads=16, pre_norm=True, **kwargs)
+    model = _create_vision_transformer('vit_huge_patch14_224_clip_laion2b', pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_giant_patch14_224_clip_laion2b(pretrained=False, **kwargs):
+    """ ViT-Giant (little-g) model (ViT-g/14) from `Scaling Vision Transformers` - https://arxiv.org/abs/2106.04560
+    Pretrained weights from CLIP image tower trained on LAION-2B image-text pairs.
+    """
+    model_kwargs = dict(
+        patch_size=14, embed_dim=1408, mlp_ratio=48/11, depth=40, num_heads=16, pre_norm=True, **kwargs)
+    model = _create_vision_transformer('vit_giant_patch14_224_clip_laion2b', pretrained=pretrained, **model_kwargs)
+    return model

timm/models/vision_transformer_hybrid.py

@@ -101,7 +101,16 @@ class HybridEmbed(nn.Module):
     """ CNN Feature Map Embedding
     Extract feature map from CNN, flatten, project to embedding dim.
     """
-    def __init__(self, backbone, img_size=224, patch_size=1, feature_size=None, in_chans=3, embed_dim=768):
+    def __init__(
+            self,
+            backbone,
+            img_size=224,
+            patch_size=1,
+            feature_size=None,
+            in_chans=3,
+            embed_dim=768,
+            bias=True,
+    ):
         super().__init__()
         assert isinstance(backbone, nn.Module)
         img_size = to_2tuple(img_size)
@@ -130,7 +139,7 @@ class HybridEmbed(nn.Module):
         assert feature_size[0] % patch_size[0] == 0 and feature_size[1] % patch_size[1] == 0
         self.grid_size = (feature_size[0] // patch_size[0], feature_size[1] // patch_size[1])
         self.num_patches = self.grid_size[0] * self.grid_size[1]
-        self.proj = nn.Conv2d(feature_dim, embed_dim, kernel_size=patch_size, stride=patch_size)
+        self.proj = nn.Conv2d(feature_dim, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
 
     def forward(self, x):
         x = self.backbone(x)