PyTorch state_dict and JAX pytree weight checkpoints for VGG16 (Simonyan & Zisserman 2014), converted from the canonical tf.keras.applications.VGG16 weights. Hosted as a one-file-each shortcut so xaitalk users can run cross-framework XAI without re-converting the TF Keras weights themselves.

If you want the original Keras weights, install tensorflow and use tf.keras.applications.VGG16(weights='imagenet'). This repo is for the PyTorch + JAX equivalents only.

Files

File	Framework	Size
`pytorch_vgg16_from_tf.pth`	PyTorch state_dict	~528 MB
`vgg16_jax_weights.pkl`	JAX pytree (pickle)	~528 MB

Both files load weights numerically equivalent to the TF Keras original. Verified by xaitalk's cross-framework comparison suite (see Performance below).

Architecture

Property	Value
Layers	13 conv (3×3) + 3 dense
Parameters	138 M
Input	224×224 RGB image
Output	1000-class ImageNet logits
Activation	ReLU
Pooling	Max (2×2)

Standard VGG16. Reference: Simonyan & Zisserman, Very Deep Convolutional Networks for Large-Scale Image Recognition, ICLR 2015.

Cross-framework verification

These weights are validated by the xaitalk method × architecture matrix. All XAI methods produce identical attributions (Pearson r=1.0 across float32 numerical precision) when run on the PyTorch / TF Keras / JAX models with this checkpoint set.

Methods tested	PASS at r ≥ 0.95	Min(min_r)
8 (gradient family + LRP + smoothgrad variants)	8/8	1.0000

Full per-method results: see findings/method_matrix/ in the xaitalk repo.

Usage

python

# Anonymous download — no HF token needed:
from xaitalk.hub import ensure_model

pt_path  = ensure_model('vgg16/pt')   # → local path to .pth
jax_path = ensure_model('vgg16/jax')  # → local path to .pkl

# Load PyTorch model
import torch
import torchvision.models as M
model_pt = M.vgg16()
model_pt.load_state_dict(torch.load(pt_path, weights_only=True))
model_pt.eval()

# Run XAI cross-framework
import xaitalk
explanation = xaitalk.explain(model_pt, x, method='lrp_epsilon')

License

VGG16 weights inherit the license of the original Oxford VGG release. Format-converted files in this repo are provided as-is under the same constraints — academic / research use, no commercial restriction. xaitalk's conversion code is Apache 2.0 (see github.com/alexanderfpaul/xaitalk-framework).

Citation

If you use these weights, please cite the original VGG paper:

bibtex

@inproceedings{simonyan2015vgg,
  author = {Simonyan, Karen and Zisserman, Andrew},
  title = {Very Deep Convolutional Networks for Large-Scale Image Recognition},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year = {2015}
}

And, if you use the cross-framework XAI infrastructure that produced these conversions, please also cite xaitalk:

bibtex

@software{paul2026xaitalk,
  author = {Paul, Alexander},
  title = {xaitalk: Cross-Framework Explainable AI Library},
  year = {2026},
  url = {https://xaitalk.com}
}

xaitalk/vgg16-imagenet

VGG16 (ImageNet) — cross-framework weights for xaitalk