Model card for gemma4_vit_167m_enc.gemma4_e4b_it

A Gemma 4 ViT image feature encoder extracted from Google's Gemma 4 E4B-it multimodal language model. This is the native VLM vision encoder with a spatial k×k soft-token pool + √D scale (bit-perfect with the HuggingFace Gemma4VisionModel reference implementation). The architecture is a custom ViT with 2D RoPE, gated MLP (GELU-tanh), RMS normalization on Q/K/V, and a 4-norm sandwich block layout.

Model Notes

• Weights were imported directly from the source VLM's vision tower (Gemma 4 E4B-it); outputs match the transformers Gemma4VisionModel implementation bit-for-bit on matching inputs.
• Gemma 4 expects raw [0, 1]-range pixel tensors — the model maps them internally to [-1, 1] via 2 * (x - 0.5). The pretrained_cfg therefore declares mean=(0, 0, 0) / std=(1, 1, 1) to disable normalization in timm's standard transform pipeline and avoid double-normalization.
• The encoder supports NaFlex-style variable-resolution inputs and integrates with timm's NaFlex data pipeline (see timm.data.naflex_loader / --naflex-loader in train.py). Both raw (B, C, H, W) images and pre-patchified (B, N, ...) tensors (plus patch_coord / patch_valid) are accepted — so the same model can be trained or fine-tuned with dynamic batch sizing that preserves native aspect ratios at flexible resolutions, or run at a fixed square input size.

Model Details

• Model Type: Image Feature Encoder
• Model Stats:
  • Params (M): 167.4
  • GMACs: 516.4
  • Activations (M): 1421.2
  • Image size: 768 x 768
• Original: https://huggingface.co/google/gemma-4-E4B-it
• License: Apache 2.0
• Papers:
  • Gemma 4: https://ai.google.dev/gemma/docs/core/model_card_4
  • PyTorch Image Models: https://github.com/huggingface/pytorch-image-models

Model Usage

Feature Map Extraction

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model(
    'gemma4_vit_167m_enc.gemma4_e4b_it',
    pretrained=True,
    features_only=True,
)
model = model.eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # unsqueeze single image into batch of 1

for o in output:
    # print shape of each feature map in output
    # e.g.:
    #  torch.Size([1, 768, 48, 48])
    #  torch.Size([1, 768, 48, 48])
    #  torch.Size([1, 768, 48, 48])

    print(o.shape)

Image Embeddings

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model(
    'gemma4_vit_167m_enc.gemma4_e4b_it',
    pretrained=True,
    num_classes=0,  # remove classifier nn.Linear
)
model = model.eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # output is (batch_size, num_features) shaped tensor

# or equivalently (without needing to set num_classes=0)

output = model.forward_features(transforms(img).unsqueeze(0))
# output is unpooled, a (1, 2304, 768) shaped tensor

output = model.forward_head(output, pre_logits=True)
# output is a (1, num_features) shaped tensor

Citation

@misc{gemma4_2025,
  title={Gemma 4},
  author={{Gemma Team, Google DeepMind}},
  year={2025},
  howpublished={\url{https://ai.google.dev/gemma/docs/core/model_card_4}}
}

@misc{rw2019timm,
  author = {Ross Wightman},
  title = {PyTorch Image Models},
  year = {2019},
  publisher = {GitHub},
  journal = {GitHub repository},
  doi = {10.5281/zenodo.4414861},
  howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
}