See our collection for all versions of LeViT.

Run LeViT with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/levit-192

Paper: LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference (arXiv:2104.01136) · HF Papers

LeViT is a hybrid convolution/transformer image classifier built for fast inference: a four-layer conv stem downsamples the image 16x, then three attention stages (each adding a learnable 2D relative-position bias) run over the tokens, with a BatchNorm fused into every linear layer and Hardswish activations. The released checkpoints are distilled - a second classification head is averaged with the first at inference. Wider LeViT (hidden sizes 192/288/384).

For more details on the model, please go to Meta's original model card.

Pure-Keras 3 conversion of facebook/levit-192 for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

✨ Quick start

import os

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

from PIL import Image
from zeromodels.models.levit import LevitImageClassify, LevitModel, LevitImageProcessor

model = LevitImageClassify.from_weights("zeromodels/levit-192")
processor = LevitImageProcessor.from_weights("zeromodels/levit-192")

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 = LevitModel.from_weights("zeromodels/levit-192")
features = backbone(pixels, training=False)

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

Variant	Hub
`levit-128S`	`zeromodels/levit-128S`
`levit-128`	`zeromodels/levit-128`
`levit-192`	`zeromodels/levit-192`
`levit-256`	`zeromodels/levit-256`
`levit-384`	`zeromodels/levit-384`

Tips

Set KERAS_BACKEND before importing Keras / zeromodels.
ImageNet normalization is baked into the model, so pass raw [0, 255] pixels.
Preprocess by resizing the shortest edge to 256 and center-cropping 224 (shown above) to match the reference; a plain resize((224, 224)) is close and also works.
LevitImageClassify averages the two distillation heads internally; LevitModel.from_weights(...) gives the backbone (the final token sequence, no head).
See Classification backbones and Loading Weights.
Community / upstream safetensors still work via the hf: prefix, e.g. LevitImageClassify.from_weights("hf:facebook/levit-192").