zeromodels/pit_b_distilled_224_in1k

See our collection for all versions of PiT.

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

zeromodels/pit_b_distilled_224_in1k

Paper: Rethinking Spatial Dimensions of Vision Transformers (arXiv:2103.16427) · HF Papers

Pooling-based Vision Transformer (PiT) reshapes spatial dimensions across stages like a CNN. Classifier or multi-stage backbone.

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

Pure-Keras 3 conversion of timm/pit_b_distilled_224.in1k for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an image-classification / backbone checkpoint (PiTImageClassify / PiTModel).

✨ Quick start

import os

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

from PIL import Image
from zeromodels.models.pit import PiTImageClassify, PiTModel, PiTImageProcessor

model = PiTImageClassify.from_weights("zeromodels/pit_b_distilled_224_in1k")
processor = PiTImageProcessor.from_weights("zeromodels/pit_b_distilled_224_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 = PiTModel.from_weights("zeromodels/pit_b_distilled_224_in1k", as_backbone=True)
features = backbone(pixels, training=False)

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

Variant	Hub
pit_b_224_in1k	zeromodels/pit_b_224_in1k
pit_b_distilled_224_in1k	zeromodels/pit_b_distilled_224_in1k
pit_s_224_in1k	zeromodels/pit_s_224_in1k
pit_s_distilled_224_in1k	zeromodels/pit_s_distilled_224_in1k
pit_ti_224_in1k	zeromodels/pit_ti_224_in1k
pit_ti_distilled_224_in1k	zeromodels/pit_ti_distilled_224_in1k
pit_xs_224_in1k	zeromodels/pit_xs_224_in1k
pit_xs_distilled_224_in1k	zeromodels/pit_xs_distilled_224_in1k

Tips

• Set KERAS_BACKEND before importing Keras / zeromodels.
• PiTImageClassify returns class logits; PiTModel returns features (as_backbone=True for multi-scale stages).
• See docs and Loading Weights.
• Upstream / timm checkpoints: PiTImageClassify.from_weights("hf:timm/pit_b_distilled_224.in1k").

Special Thanks

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

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