vit-zoo 0.2.0

Vision Transformers Zoo

A clean, extensible factory for creating HuggingFace-based Vision Transformer models (ViT, DeiT, DINO, DINOv2, DINOv3, CLIP) with flexible heads and easy backbone freezing.

Installation

pip install vit_zoo

From source:

git clone https://github.com/jbindaAI/vit_zoo.git
cd vit_zoo
pip install -e .

For development: pip install -e ".[dev]"

Quick start

from vit_zoo.factory import build_model

model = build_model("facebook/dinov2-base", head=10, freeze_backbone=True)
outputs = model(images)
logits = outputs["predictions"]  # (batch_size, 10)

Basic usage

from vit_zoo.factory import build_model

# Simple classification - pass any HuggingFace model ID
model = build_model("google/vit-base-patch16-224", head=10, freeze_backbone=True)
outputs = model(images)
predictions = outputs["predictions"]  # Shape: (batch_size, 10)

Custom MLP Head

from vit_zoo.factory import build_model
from vit_zoo.components import MLPHead

mlp_head = MLPHead(
    input_dim=768,
    hidden_dims=[512, 256],
    output_dim=100,
    dropout=0.1,
    activation="gelu"  # or 'relu', 'tanh', or nn.Module
)

model = build_model("facebook/dinov2-base", head=mlp_head)

Embedding Extraction

from transformers import CLIPVisionModel
model = build_model("openai/clip-vit-base-patch16", backbone_cls=CLIPVisionModel, head=None)
outputs = model(images, output_embeddings=True)
hidden_states = outputs["last_hidden_state"]  # (batch_size, seq_len, embedding_dim)
cls_embedding = hidden_states[:, 0, :]  # (batch_size, embedding_dim)
predictions = outputs["predictions"]  # same as cls_embedding when head=None (IdentityHead)

Attention Weights

model = build_model(
    "google/vit-base-patch16-224",
    head=10,
    config_kwargs={"attn_implementation": "eager"}
)
outputs = model(images, output_attentions=True)
attentions = outputs["attentions"]

Custom Head

from vit_zoo.components import BaseHead
import torch.nn as nn

class CustomHead(BaseHead):
    def __init__(self, input_dim: int, num_classes: int):
        super().__init__()
        self._input_dim = input_dim
        self.fc = nn.Linear(input_dim, num_classes)
    
    @property
    def input_dim(self) -> int:
        return self._input_dim
    
    def forward(self, embeddings):
        return self.fc(embeddings)

head = CustomHead(input_dim=768, num_classes=10)
model = build_model("google/vit-base-patch16-224", head=head)

Multi-modal Models (CLIP)

For CLIP and other multi-modal models, pass backbone_cls to load only the vision encoder (AutoModel would load the full model):

from transformers import CLIPVisionModel
model = build_model("openai/clip-vit-base-patch16", backbone_cls=CLIPVisionModel, head=10)

Any HuggingFace Model

build_model uses AutoModel to auto-detect the model type from the HuggingFace Hub. Any ViT-compatible model works:

model = build_model("google/vit-large-patch16-224", head=10)
model = build_model("facebook/deit-base-distilled-patch16-224", head=10)
model = build_model("facebook/dinov2-with-registers-base", head=10)

API Reference

build_model()

build_model(
    model_name: str,
    head: Optional[Union[int, BaseHead]] = None,
    backbone_cls: Optional[Type] = None,
    freeze_backbone: bool = False,
    load_pretrained: bool = True,
    backbone_dropout: float = 0.0,
    config_kwargs: Optional[Dict[str, Any]] = None,
) -> ViTModel

Parameters:

• model_name: HuggingFace model identifier (e.g., "google/vit-base-patch16-224", "facebook/dinov2-base", "openai/clip-vit-base-patch16"). Uses AutoModel to auto-detect model type.
• head: int (creates LinearHead), BaseHead instance, or None (embedding extraction)
• backbone_cls: Optional HuggingFace model class (e.g., CLIPVisionModel). Use for multi-modal models where AutoModel loads the full model.
• freeze_backbone: Freeze all backbone parameters
• config_kwargs: Extra config options (e.g., {"attn_implementation": "eager"})

Usage:

• build_model("google/vit-base-patch16-224", head=10)
• build_model("facebook/dinov2-base", head=None) (embedding extraction)

ViTModel.forward()

forward(
    pixel_values: torch.Tensor,
    output_attentions: bool = False,
    output_embeddings: bool = False,
) -> Dict[str, Any]

Always returns a dict. Keys:

• "predictions": head output tensor (always present)
• "attentions": optional, when output_attentions=True
• "last_hidden_state": optional, when output_embeddings=True; shape (batch_size, seq_len, embedding_dim)

ViTModel.freeze_backbone()

model.freeze_backbone(freeze: bool = True)  # Freeze/unfreeze backbone

Supported Models

Any ViT-compatible model on the HuggingFace Hub works. Examples:

• google/vit-base-patch16-224, google/vit-large-patch16-224 (ViT)
• facebook/deit-base-distilled-patch16-224 (DeiT)
• facebook/dino-vitb16 (DINO)
• facebook/dinov2-base, facebook/dinov2-with-registers-base (DINOv2)
• facebook/dinov3-vitb16-pretrain-lvd1689m (DINOv3)
• openai/clip-vit-base-patch16 (CLIP Vision; pass backbone_cls=CLIPVisionModel)

Browse the HuggingFace Hub for more models.

Import Patterns

from vit_zoo import ViTModel
from vit_zoo.factory import build_model
from vit_zoo.components import ViTBackbone, BaseHead, LinearHead, MLPHead, IdentityHead

Available Heads

• LinearHead: Simple linear layer (auto-created when head=int)
• MLPHead: Multi-layer perceptron with configurable depth, activation, dropout
• IdentityHead: Returns embeddings unchanged

All heads must implement input_dim property. Custom heads by subclassing BaseHead.

License

GPL-3.0