tomas-gajarsky/facetorch-detector-retinaface

Face Detection

Exported PyTorch model (.pt2) for use with facetorch.

Model Details

Task	Face Detection
Architecture	RetinaFace with ResNet-50 backbone
Format	torch.export (.pt2) — no model source code needed
Dynamic shapes	Batch (1-32), Height and Width (multiples of 32, 64-2048)
Input	RGB image, spatial dims must be multiples of 32
Output	(bbox_regressions, classifications, landmark_regressions)

Original Work

This model is based on biubug6/Pytorch_Retinaface. Weights converted and exported by facetorch.

Dynamic Shape Export

The model is exported with derived dimensions using torch.export.Dim:

• Batch: 1-32
• Height: 32 * h_base where h_base in [2, 64] (i.e., 64-2048 in steps of 32)
• Width: 32 * w_base where w_base in [2, 64] (i.e., 64-2048 in steps of 32)

The multiples-of-32 constraint matches the model stride chain (8, 16, 32) ensuring all feature map dimensions are integral.

Usage

import torch

# Load — no model class needed
ep = torch.export.load("model.pt2")
model = ep.module()

# Inference (spatial dims must be multiples of 32)
x = torch.randn(1, 3, 640, 480)
output = model(x)

Or via facetorch:

from facetorch import FaceAnalyzer
from omegaconf import OmegaConf

cfg = OmegaConf.load("conf/config.yaml")
analyzer = FaceAnalyzer(cfg.analyzer)
response = analyzer.run(path_image="face.jpg")