RetinaFace (fastforks)

RetinaFace (Deng et al., CVPR 2020) in the facexlib / Pytorch_Retinaface version: single-stage face detection with five facial landmarks (eyes, nose, mouth corners), the detector of the CodeFormer / GFPGAN face restoration pipelines. A ResNet-50 or MobileNet-0.25 backbone, an FPN, SSH context modules and per-anchor face, box and landmark heads on the stride 8 / 16 / 32 maps. Takes the whole image at its own size (no resize, as facexlib's default use_origin_size=True); vendored from facexlib 0.3.0 (no facexlib / torchvision dependency). The portable module() maps the image to the raw head maps (anchor decoding and NMS stay outside).

This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/xinntao/facexlib at 260620ae9399.

Task: pose. Tier: stable. Outputs: keypoints, keypoint_scores, boxes, scores. Backends: portable, kernels.

Input: one RGB image and a box per person. Output: keypoints with confidence scores for every person.

Variants

Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.

variant weights (author link) SHA-256 weights licence commercial use
retinaface-mobilenet0.25@1 detection_mobilenet0.25_Final.pth 2979b33ffafd MIT yes
retinaface-resnet50@1 detection_Resnet50_Final.pth 6d1de9c2944f MIT yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: MIT. Licence files: LICENSE.

Training data: WIDER FACE (32k images, 394k labelled faces) (WIDER FACE train with the RetinaFace five-landmark annotations; the ResNet-50 backbone starts from ImageNet classification weights).

Licence evidence: https://github.com/xinntao/facexlib/blob/260620ae93990a300f4b16448df9bb459f1caba9/README.md

Input size

any H x W >= 32, padded to a multiple of 32 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): retinaface-mobilenet0.25 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); retinaface-resnet50 3000x4000 (1.6 GiB peak allocated, 1.8 reserved)

Speed and memory

Pose, synthetic, 1080x1920

model time (ms): original fp32 original portable kernels kernels speed-up: vs original vs portable peak memory (MiB, allocated / reserved): original portable kernels
retinaface-mobilenet0.25 75.9 77.4 3.44 2.74 27.7x 1.3x 87 / 110 86 / 108 105 / 128
retinaface-resnet50 81.0 79.4 24.3 11.7 6.8x 2.1x 371 / 490 405 / 504 379 / 464

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
retinaface-mobilenet0.25 0.13 0.71 1.49 1.37 0.68 0.06
retinaface-resnet50 0.43 0.71 1.55 0.31 0.11 0.10

torch.compile and TensorRT

model input time: portable compile TensorRT fp16 kernels preparation: compile (cold first call) ONNX + engine build peak MiB (alloc / reserved): portable compile kernels TensorRT (engine + torch)
retinaface-mobilenet0.25 1080x1920 3.44 ms 2.16 ms 0.982 ms (network only) 2.74 ms 8.87 s 53.5 s 86 / 108 125 / 148 105 / 128 133 / 144
retinaface-resnet50 1080x1920 24.3 ms 16.5 ms 9.10 ms (network only) 11.7 ms 9.56 s 20.4 s 405 / 504 374 / 444 379 / 464 451 / 462

Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the original code named above at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.

Accuracy

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 0.001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.016 and p99 <= 0.04 (absolute error of the output keypoints).

Limitations

  • module() exports the raw per-level head maps; anchor decoding and NMS remain outside the exported graph.
  • No per-landmark confidence: keypoint_scores repeats the face score.
  • The image is not resized: large photos cost time and memory in proportion to their pixels; downscale them first for big faces.

Usage

from fastforks import pose

model = pose.load("retinaface-mobilenet0.25")  # mode="auto": the preferred backend from recorded measurements
result = model("photo.jpg", boxes=[[x0, y0, x1, y1]])  # one box per person, or a fastforks.detection result
result.keypoints, result.keypoint_scores  # [P, K, 2] input pixels, [P, K]
result.backend  # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')

Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: retinaface-mobilenet0.25, retinaface-resnet50.

Kernels backend

pose.load("retinaface-mobilenet0.25", mode="kernels") runs our CUDA kernels (cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Collection including fastforks/retinaface