FocalNet (fastforks)

FocalNet (Yang et al. 2022): a hierarchical classifier that replaces self-attention by focal modulation (stacked depthwise convolutions with gated aggregation).

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/huggingface/pytorch-image-models at 0df212b369a5.

Author weight repos: https://huggingface.co/timm/focalnet_base_lrf.ms_in1k, https://huggingface.co/timm/focalnet_base_srf.ms_in1k, https://huggingface.co/timm/focalnet_small_lrf.ms_in1k, https://huggingface.co/timm/focalnet_small_srf.ms_in1k, https://huggingface.co/timm/focalnet_tiny_lrf.ms_in1k, https://huggingface.co/timm/focalnet_tiny_srf.ms_in1k.

Task: classification. Tier: experimental. Outputs: logits. Backends: portable.

Input: one RGB image. Output: ImageNet-1k class logits.

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
focalnet_base_lrf.ms_in1k@1 focalnet_base_lrf.ms_in1k.safetensors e1c36016eaef MIT yes
focalnet_base_srf.ms_in1k@1 focalnet_base_srf.ms_in1k.safetensors 05a7ff7cda7e MIT yes
focalnet_small_lrf.ms_in1k@1 focalnet_small_lrf.ms_in1k.safetensors c7d963802b47 MIT yes
focalnet_small_srf.ms_in1k@1 focalnet_small_srf.ms_in1k.safetensors 2590faabe2cd MIT yes
focalnet_tiny_lrf.ms_in1k@1 focalnet_tiny_lrf.ms_in1k.safetensors f5f73a0b0430 MIT yes
focalnet_tiny_srf.ms_in1k@1 focalnet_tiny_srf.ms_in1k.safetensors fc5ae22ca975 MIT yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: Apache-2.0.

Training data: ImageNet-1k (ILSVRC 2012 classification).

Licence evidence: https://huggingface.co/timm/focalnet_tiny_srf.ms_in1k/blob/de75c3a73e9ed84fee554a7138d4bac119751645/README.md

Input size

fixed 224x224: the input is resized (short side 248) and centre-cropped to it Tested up to (portable, fp16, RTX 5060 Ti 16 GB): focalnet_base_lrf.ms_in1k 224x224 (0.2 GiB peak allocated, 0.2 reserved); focalnet_base_srf.ms_in1k 224x224 (0.2 GiB peak allocated, 0.2 reserved); focalnet_small_lrf.ms_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved); focalnet_small_srf.ms_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved); focalnet_tiny_lrf.ms_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved); focalnet_tiny_srf.ms_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved)

Speed and memory

Classification, synthetic

model input time (ms): original fp32 original portable kernels kernels speed-up: vs original vs portable peak memory (MiB, allocated / reserved): original portable kernels
focalnet_base_lrf.ms_in1k 224x224 18.0 22.1 15.5 – – – 546 / 560 212 / 234 –
focalnet_base_srf.ms_in1k 224x224 17.1 20.0 14.2 – – – 542 / 558 211 / 234 –
focalnet_small_lrf.ms_in1k 224x224 16.8 22.0 15.8 – – – 323 / 338 136 / 152 –
focalnet_small_srf.ms_in1k 224x224 15.9 20.1 14.0 – – – 320 / 336 136 / 152 –
focalnet_tiny_lrf.ms_in1k 224x224 8.92 11.4 8.07 – – – 199 / 208 95 / 118 –
focalnet_tiny_srf.ms_in1k 224x224 8.39 10.4 7.36 – – – 197 / 206 95 / 118 –

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
focalnet_base_lrf.ms_in1k 0.91 1.89 – 0.03 0.62 –
focalnet_base_srf.ms_in1k 0.89 1.92 – 0.05 0.49 –
focalnet_small_lrf.ms_in1k 0.54 1.86 – 0.03 0.69 –
focalnet_small_srf.ms_in1k 0.54 1.73 – 0.02 0.49 –
focalnet_tiny_lrf.ms_in1k 0.36 2.56 – 0.15 0.77 –
focalnet_tiny_srf.ms_in1k 0.40 1.49 – 0.45 0.49 –

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)
focalnet_base_lrf.ms_in1k 224x224 15.5 ms 7.47 ms 2.08 ms – 19.7 s 63.9 s 212 / 234 235 / 274 – 404 / 432
focalnet_base_srf.ms_in1k 224x224 14.2 ms 6.94 ms 1.74 ms – 17.8 s 63.8 s 211 / 234 234 / 272 – 401 / 430
focalnet_small_lrf.ms_in1k 224x224 15.8 ms 7.28 ms 1.70 ms – 19.5 s 62.5 s 136 / 152 162 / 192 – 265 / 286
focalnet_small_srf.ms_in1k 224x224 14.0 ms 6.90 ms 1.49 ms – 17.8 s 65.3 s 136 / 152 161 / 190 – 234 / 256
focalnet_tiny_lrf.ms_in1k 224x224 8.07 ms 3.94 ms 0.952 ms – 13.3 s 49.9 s 95 / 118 120 / 138 – 158 / 186
focalnet_tiny_srf.ms_in1k 224x224 7.36 ms 3.72 ms 0.864 ms – 12.9 s 52.0 s 95 / 118 120 / 136 – 159 / 188

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

On public datasets

fastforks: the model's own resize / crop on the GPU; original: the library's pretrained loader and its PIL eval transform. Original = the original code named under Details, same checkpoint and protocol.

variant dataset original fp32 original fp16 portable note
focalnet_tiny_srf.ms_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 82.86 / 96.28 82.88 / 96.30 82.90 / 96.28
focalnet_small_srf.ms_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 84.34 / 96.84 84.30 / 96.84 84.44 / 96.76
focalnet_base_srf.ms_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 84.20 / 96.86 84.18 / 96.86 84.16 / 96.84
focalnet_tiny_lrf.ms_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 82.88 / 96.12 82.86 / 96.10 82.86 / 96.20
focalnet_small_lrf.ms_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 84.36 / 96.80 84.34 / 96.80 84.32 / 96.80
focalnet_base_lrf.ms_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 84.64 / 96.64 84.64 / 96.64 84.56 / 96.56

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 7ba7307f, 2026-11-11.

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 2.9e-06 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0027 and p99 <= 0.0098 (absolute error of the output logits).

Usage

from fastforks import classification

model = classification.load("focalnet_base_lrf.ms_in1k")  # mode="auto": the preferred backend from recorded measurements
result = model("photo.jpg")
result.top(5)  # top-5 classes
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: focalnet_base_lrf.ms_in1k, focalnet_base_srf.ms_in1k, focalnet_small_lrf.ms_in1k, focalnet_small_srf.ms_in1k, focalnet_tiny_lrf.ms_in1k, focalnet_tiny_srf.ms_in1k.

Kernels backend

No kernels backend: not implemented: LayerNorm blocks with depthwise convolutions and gated sums: needs its own runner.

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. πŸ™‹ Ask for provider support

Collection including fastforks/focalnet