MobileNetV1 transfer reconstruction for 25 Oxford-IIIT Pet dog breeds.

Frozen backbone: Keras Applications MobileNetV1 alpha=0.5, ImageNet pretrained, no classification top, global average pooling.
Official weight source: https://storage.googleapis.com/tensorflow/keras-applications/mobilenet/mobilenet_5_0_224_tf_no_top.h5
Architecture source: https://github.com/keras-team/keras/blob/master/keras/src/applications/mobilenet.py
Legacy distribution and MIT licence: https://github.com/keras-team/keras-applications
TensorFlow.js runtime: https://github.com/tensorflow/tfjs (Apache-2.0).
Associated source licence texts are retained alongside this notice.

Changes: frozen Keras weights serialized as little-endian float32, minimal layer metadata; ImageNet top omitted; separate new linear SVM head trained on the attributed Oxford-IIIT Pet mirror. No ImageNet image is redistributed.

Dataset source: https://www.robots.ox.ac.uk/~vgg/data/pets/
Authors: O. M. Parkhi, A. Vedaldi, A. Zisserman, C. V. Jawahar (CVPR 2012).
Mirror: https://huggingface.co/datasets/enterprise-explorers/oxford-pets
Oxford dataset licence: CC BY-SA 4.0, original image authors retain copyright.
The three displayed image adaptations and their exact attributions are in /demos/santevet/ATTRIBUTION.txt.

Limitations: closed set of 25 dog breeds; no animal-presence or out-of-domain detector. Scores are uncalibrated SVM margins, not probabilities. No client data or client model. External ImageNet pretraining overlap with Oxford was not audited. The random held-out split is not the official Oxford split and not external production validation.