starnet / README.md
Flor's picture
card: update from fastforks 1cbf1671
b517a7e verified
|
Raw History Blame Contribute Delete
7.83 kB
metadata
library_name: fastforks
pipeline_tag: image-classification
license: apache-2.0
tags:
  - classification
  - image-classification
  - cuda
  - pytorch
  - fastforks
  - fastforks-card

StarNet (fastforks)

StarNet (Ma et al. 2024, Rewrite the Stars): a small convolutional classifier whose blocks multiply two branches element-wise (the star operation).

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/starnet_s1.in1k, https://huggingface.co/timm/starnet_s2.in1k, https://huggingface.co/timm/starnet_s3.in1k, https://huggingface.co/timm/starnet_s4.in1k.

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

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
starnet_s1.in1k@1 starnet_s1.in1k.safetensors 314e10df155c Apache-2.0 yes
starnet_s2.in1k@1 starnet_s2.in1k.safetensors dc73c27d1b71 Apache-2.0 yes
starnet_s3.in1k@1 starnet_s3.in1k.safetensors 7fbdee07ac6b Apache-2.0 yes
starnet_s4.in1k@1 starnet_s4.in1k.safetensors 6d75002d5e77 Apache-2.0 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/starnet_s1.in1k/blob/e3fbe6d1e24b6293fa6e7b897f380bd62f77e01f/README.md

Input size

fixed 224x224: the input is resized (short side 256) and centre-cropped to it Tested up to (portable, fp16, RTX 5060 Ti 16 GB): starnet_s1.in1k 224x224 (0.0 GiB peak allocated, 0.1 reserved); starnet_s2.in1k 224x224 (0.0 GiB peak allocated, 0.1 reserved); starnet_s3.in1k 224x224 (0.0 GiB peak allocated, 0.1 reserved); starnet_s4.in1k 224x224 (0.0 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
starnet_s1.in1k 224x224 4.57 7.29 3.88 0.683 6.7x 5.7x 51 / 60 41 / 52 41 / 56
starnet_s2.in1k 224x224 3.64 5.63 3.02 0.578 6.3x 5.2x 55 / 64 43 / 54 43 / 58
starnet_s3.in1k 224x224 4.96 7.67 4.19 0.715 6.9x 5.9x 67 / 76 47 / 60 47 / 64
starnet_s4.in1k 224x224 6.84 11.3 5.63 0.912 7.5x 6.2x 77 / 86 50 / 66 51 / 72

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
starnet_s1.in1k 0.18 1.60 2.20 0.26 0.24 0.21
starnet_s2.in1k 0.17 1.62 1.86 0.02 0.24 0.06
starnet_s3.in1k 0.30 1.60 1.82 0.01 0.24 0.05
starnet_s4.in1k 0.21 1.66 2.00 0.01 0.24 0.05

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)
starnet_s1.in1k 224x224 3.88 ms 2.41 ms 0.774 ms 0.683 ms 6.23 s 24.4 s 41 / 52 73 / 92 41 / 56 73 / 86
starnet_s2.in1k 224x224 3.02 ms 1.96 ms 0.622 ms 0.578 ms 5.71 s 24.7 s 43 / 54 75 / 94 43 / 58 74 / 88
starnet_s3.in1k 224x224 4.19 ms 2.60 ms 0.878 ms 0.715 ms 6.03 s 27.5 s 47 / 60 79 / 100 47 / 64 80 / 96
starnet_s4.in1k 224x224 5.63 ms 3.37 ms 1.18 ms 0.912 ms 6.84 s 28.0 s 50 / 66 82 / 106 51 / 72 85 / 104

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 kernels note
starnet_s1.in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 74.50 / 91.94 74.50 / 91.96 74.48 / 91.92 74.52 / 91.92
starnet_s2.in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 75.82 / 92.14 75.84 / 92.12 75.84 / 92.10 75.88 / 92.12
starnet_s3.in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 78.08 / 93.72 78.14 / 93.74 78.10 / 93.80 78.06 / 93.80
starnet_s4.in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 80.08 / 94.56 80.06 / 94.56 79.98 / 94.58 79.92 / 94.58

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 9e68b313, 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| <= 1.1e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0035 and p99 <= 0.012 (absolute error of the output logits).

Usage

from fastforks import classification

model = classification.load("starnet_s1.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: starnet_s1.in1k, starnet_s2.in1k, starnet_s3.in1k, starnet_s4.in1k.

Kernels backend

classification.load("starnet_s1.in1k", 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.