How novendrastywn/facelivt works
FaceLiVT is a face-recognition model repository with two CNN–Transformer generations. Its user-facing workflow aligns a face image, preprocesses it, and runs a selected backbone to produce an embedding; training and benchmark evaluation are major supporting workflows. The graph also preserves the documented v1/v2 model families and training configurations. Some subsystem connections rely on README or SOURCE INDEX references; unsampled implementation details are not inferred.
Main entry points: train_v2.py Distributed training · mtcnn.py MTCNN detector · __init__.py Model selection · inference.py Embedding inference
Components 26
Recognition models
- Embedding inference
inference.py - Model selection
backbones/__init__.py - FaceLiVT v1
backbones/facelivt.py - FaceLiVT v2
backbones/facelivtv2.py - Face embedding
Training pipeline
- Model configurations
configs - Training dataset
dataset.py - Distributed sampler
utils/utils_distributed_sampler.py - Distributed training
train_v2.py - Recognition losses
losses.py - Partial FC
partial_fc_v2.py - Learning-rate scheduler
lr_scheduler.py - Training callbacks
utils/utils_callbacks.py
Face alignment
- Aligned face
face_alignment/align.py - MTCNN detector
face_alignment/mtcnn.py - Detection stages
face_alignment/mtcnn_pytorch/src/first_stage.py - PNet RNet ONet
face_alignment/mtcnn_pytorch/src/get_nets.py - Box operations
face_alignment/mtcnn_pytorch/src/box_utils.py - Face transform
face_alignment/mtcnn_pytorch/src/align_trans.py - Similarity transform
face_alignment/mtcnn_pytorch/src/matlab_cp2tform.py
Evaluation and export
- Verification metrics
eval/verification.py - IJB-C evaluation
eval_ijbc.py - ONNX utilities
onnx_helper.py - ONNX IJB-C
onnx_ijbc.py
Other components
- Face recognition user
- Face image
Connections 28
- Face recognition user → Face image: provides
- Face recognition user → Aligned face: requests alignment
- Face image → Aligned face: aligns
- Aligned face → MTCNN detector: detects and aligns
face_alignment/align.py - MTCNN detector → PNet RNet ONet: runs networks
face_alignment/mtcnn.py - MTCNN detector → Detection stages: runs detection
face_alignment/mtcnn.py - MTCNN detector → Box operations: uses box operations
face_alignment/mtcnn.py - MTCNN detector → Face transform: warps face
face_alignment/mtcnn.py - Face transform → Similarity transform: computes transform
face_alignment/mtcnn_pytorch/src/align_trans.py - Face recognition user → Embedding inference: requests embedding
- Embedding inference → Model selection: selects model
inference.py - Model selection → FaceLiVT v1: exposes model
backbones/__init__.py - Model selection → FaceLiVT v2: exposes model
backbones/__init__.py - Embedding inference → Face embedding: produces
inference.py - Aligned face → Embedding inference: preprocesses input
- Distributed training → Model configurations: loads configuration
train_v2.py - Distributed training → Training dataset: loads batches
train_v2.py - Distributed training → Distributed sampler: uses sampler
train_v2.py - Training dataset → Distributed sampler: samples indices
dataset.py - Distributed training → Model selection: creates backbone
train_v2.py - Distributed training → Recognition losses: computes loss
train_v2.py - Distributed training → Partial FC: uses classifier
train_v2.py - Distributed training → Learning-rate scheduler: schedules learning rate
train_v2.py - Distributed training → Training callbacks: runs callbacks
train_v2.py - Training callbacks → Verification metrics: evaluates checkpoints
utils/utils_callbacks.py - IJB-C evaluation → Model selection: loads backbone
eval_ijbc.py - IJB-C evaluation → ONNX utilities: uses inference helpers
- ONNX IJB-C → ONNX utilities: uses utilities
onnx_ijbc.py