mobilevitv2_175_cvnets_in22k_ft_in1k_384

See our collection for all versions of MobileViT-V2.

Run MobileViT-V2 with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/mobilevitv2_175_cvnets_in22k_ft_in1k_384

Paper: Separable Self-attention for Mobile Vision Transformers (arXiv:2206.02680) · HF Papers

MobileViTV2 replaces MHSA with separable self-attention (O(k)) and scales width via a single multiplier. Classification at 256/384; DeepLabV3 segmentation is a separate script/collection.

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of timm/mobilevitv2_175.cvnets_in22k_ft_in1k_384 for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an image-classification / backbone checkpoint (MobileViTV2ImageClassify / MobileViTV2Model).

✨ Quick start

import os

os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from zeromodels.models.mobilevitv2 import MobileViTV2ImageClassify, MobileViTV2Model, MobileViTV2ImageProcessor

model = MobileViTV2ImageClassify.from_weights("zeromodels/mobilevitv2_175_cvnets_in22k_ft_in1k_384")
processor = MobileViTV2ImageProcessor.from_weights("zeromodels/mobilevitv2_175_cvnets_in22k_ft_in1k_384")

image = Image.open("your_image.jpg").convert("RGB")
pixels = processor(image)  # resize + normalize (normalization lives in the processor)
logits = model(pixels, training=False)
print(logits.shape)  # (1, num_classes)

# Feature extraction: the backbone without the classifier head
backbone = MobileViTV2Model.from_weights("zeromodels/mobilevitv2_175_cvnets_in22k_ft_in1k_384", as_backbone=True)
features = backbone(pixels, training=False)

Load any MobileViT-V2 variant the same way with from_weights("zeromodels/<variant>"):

Tips

• Set KERAS_BACKEND before importing Keras / zeromodels.
• MobileViTV2ImageClassify returns class logits; MobileViTV2Model returns features (as_backbone=True for multi-scale stages).
• See docs and Loading Weights.
• Upstream / timm checkpoints: MobileViTV2ImageClassify.from_weights("hf:timm/mobilevitv2_175.cvnets_in22k_ft_in1k_384").

Special Thanks

A huge thank you to the MobileViT-V2 authors and the timm / Hub communities for creating and releasing these models.

License: see YAML license (usually matches the upstream checkpoint).