zeromodels/pvt-tiny-224

See our collection for all PVT and PVTv2 versions.

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

zeromodels/pvt-tiny-224

Paper: Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions (arXiv:2102.12122) · HF Papers

PVT is a hierarchical vision transformer: four pyramid stages with spatial-reduction attention over non-overlapping patches and learned position embeddings. Use PvtImageClassify for logits or PvtModel for tokens / per-stage features via as_backbone=True.

• Parameters: ~13.2M
• ImageNet-1k top-1: 75.1%

For more details on the model, see the upstream model card.

Pure-Keras 3 conversion of Zetatech/pvt-tiny-224 for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an image-classification / backbone checkpoint (PvtImageClassify / PvtModel).

✨ Quick start

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

from PIL import Image
import numpy as np
from zeromodels.models.pvt import PvtImageClassify, PvtModel

model = PvtImageClassify.from_weights("zeromodels/pvt-tiny-224")
backbone = PvtModel.from_weights("zeromodels/pvt-tiny-224", as_backbone=True)

image = Image.open("your_image.jpg").convert("RGB").resize((224, 224))
x = np.asarray(image, dtype="float32")[None]  # (1, H, W, 3), raw [0, 255]
print(model(x).shape)  # (1, num_classes)
feats = backbone(x)
print(len(feats), [tuple(f.shape) for f in feats])  # 4-stage feature pyramid

Normalization is baked into the graph, so pass raw [0, 255] pixels. Load any PVT variant the same way with from_weights("zeromodels/<variant>"):

Variant	ImageNet-1k top-1	Hub
pvt-tiny-224	75.1%	zeromodels/pvt-tiny-224
pvt-small-224	79.8%	zeromodels/pvt-small-224
pvt-medium-224	81.2%	zeromodels/pvt-medium-224
pvt-large-224	81.7%	zeromodels/pvt-large-224

Tips

• Set KERAS_BACKEND before importing Keras / zeromodels.
• PvtImageClassify returns class logits; PvtModel returns features (as_backbone=True for the four-stage pyramid).
• Both the model and its data format (channels_last / channels_first) are supported and bit-exact.
• See the docs and Loading Weights.
• Upstream checkpoints load directly: PvtImageClassify.from_weights("hf:Zetatech/pvt-tiny-224").

Special Thanks

A huge thank you to the PVT authors (whai362/PVT) and the Hugging Face community for creating and releasing these models.

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