Khalyie/lfw-face-recognition

LFW Face Recognition (InceptionResnetV1, fine-tuned)

facenet-pytorch InceptionResnetV1 (VGGFace2-pretrained), fine-tuned on the bitmind/lfw dataset. Only the final embedding layer (last_linear, last_bn) and a classifier head were trained; the rest of the backbone is frozen.

• Identities in the recognition database: 5749
• Top-1 recognition accuracy (held-out test images): 89.31%
• Embedding dimension: 512 (L2-normalized)

Files

• inceptionresnet_lfw.pt — model checkpoint (state_dict + identity name lists)
• database_embeddings.npy — mean L2-normalized embedding per known identity
• database_names.json — identity names, row-aligned with database_embeddings.npy
• metadata.json — dataset/split metadata (full identity list, train/val/test indices)

Usage

See predict_from_hub.py (companion script) for a ready-to-run example: it downloads this checkpoint + database, aligns a new photo with MTCNN, embeds it, and returns the nearest identity by cosine similarity.

The full aligned face tensors, per-split embeddings, and test-set predictions used to produce the accuracy above are in the companion dataset repo: Khalyie/lfw-face-recognition-data.