See our collection for all versions of DeiT.

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

zeromodels/deit3_large_patch16_384_fb_in22k_ft_in1k

Paper: Training data-efficient image transformers and distillation through attention (arXiv:2012.12877) · HF Papers

DeiT / DeiT3 are data-efficient ViT variants (distillation token in DeiT; improved training recipe in DeiT3). Same ImageClassify / Model API.

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

Pure-Keras 3 conversion of timm/deit3_large_patch16_384.fb_in22k_ft_in1k for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an image-classification / backbone checkpoint (DeiTImageClassify / DeiTModel).

✨ Quick start

import os

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

from PIL import Image
from zeromodels.models.deit import DeiTImageClassify, DeiTModel, DeiTImageProcessor

model = DeiTImageClassify.from_weights("zeromodels/deit3_large_patch16_384_fb_in22k_ft_in1k")
processor = DeiTImageProcessor.from_weights("zeromodels/deit3_large_patch16_384_fb_in22k_ft_in1k")

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 = DeiTModel.from_weights("zeromodels/deit3_large_patch16_384_fb_in22k_ft_in1k", as_backbone=True)
features = backbone(pixels, training=False)

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

Variant	Hub
`deit3_base_patch16_224_fb_in1k`	`zeromodels/deit3_base_patch16_224_fb_in1k`
`deit3_base_patch16_224_fb_in22k_ft_in1k`	`zeromodels/deit3_base_patch16_224_fb_in22k_ft_in1k`
`deit3_base_patch16_384_fb_in1k`	`zeromodels/deit3_base_patch16_384_fb_in1k`
`deit3_base_patch16_384_fb_in22k_ft_in1k`	`zeromodels/deit3_base_patch16_384_fb_in22k_ft_in1k`
`deit3_huge_patch14_224_fb_in1k`	`zeromodels/deit3_huge_patch14_224_fb_in1k`
`deit3_huge_patch14_224_fb_in22k_ft_in1k`	`zeromodels/deit3_huge_patch14_224_fb_in22k_ft_in1k`
`deit3_large_patch16_224_fb_in1k`	`zeromodels/deit3_large_patch16_224_fb_in1k`
`deit3_large_patch16_224_fb_in22k_ft_in1k`	`zeromodels/deit3_large_patch16_224_fb_in22k_ft_in1k`
`deit3_large_patch16_384_fb_in1k`	`zeromodels/deit3_large_patch16_384_fb_in1k`
`deit3_large_patch16_384_fb_in22k_ft_in1k`	`zeromodels/deit3_large_patch16_384_fb_in22k_ft_in1k`
`deit3_medium_patch16_224_fb_in1k`	`zeromodels/deit3_medium_patch16_224_fb_in1k`
`deit3_medium_patch16_224_fb_in22k_ft_in1k`	`zeromodels/deit3_medium_patch16_224_fb_in22k_ft_in1k`
`deit3_small_patch16_224_fb_in1k`	`zeromodels/deit3_small_patch16_224_fb_in1k`
`deit3_small_patch16_224_fb_in22k_ft_in1k`	`zeromodels/deit3_small_patch16_224_fb_in22k_ft_in1k`
`deit3_small_patch16_384_fb_in1k`	`zeromodels/deit3_small_patch16_384_fb_in1k`
`deit3_small_patch16_384_fb_in22k_ft_in1k`	`zeromodels/deit3_small_patch16_384_fb_in22k_ft_in1k`
`deit_base_distilled_patch16_224_fb_in1k`	`zeromodels/deit_base_distilled_patch16_224_fb_in1k`
`deit_base_distilled_patch16_384_fb_in1k`	`zeromodels/deit_base_distilled_patch16_384_fb_in1k`
`deit_base_patch16_224_fb_in1k`	`zeromodels/deit_base_patch16_224_fb_in1k`
`deit_base_patch16_384_fb_in1k`	`zeromodels/deit_base_patch16_384_fb_in1k`
`deit_small_distilled_patch16_224_fb_in1k`	`zeromodels/deit_small_distilled_patch16_224_fb_in1k`
`deit_small_patch16_224_fb_in1k`	`zeromodels/deit_small_patch16_224_fb_in1k`
`deit_tiny_distilled_patch16_224_fb_in1k`	`zeromodels/deit_tiny_distilled_patch16_224_fb_in1k`
`deit_tiny_patch16_224_fb_in1k`	`zeromodels/deit_tiny_patch16_224_fb_in1k`

Tips

Set KERAS_BACKEND before importing Keras / zeromodels.
DeiTImageClassify returns class logits; DeiTModel returns features (as_backbone=True for multi-scale stages).
See docs and Loading Weights.
Upstream / timm checkpoints: DeiTImageClassify.from_weights("hf:timm/deit3_large_patch16_384.fb_in22k_ft_in1k").