Models · Classification

StarNet

StarNet (Ma et al. 2024; S1 to S4, ImageNet-1k): a small convolutional classifier whose blocks multiply two branches element-wise. Apache-2.0.

experimental Fork of huggingface/pytorch-image-models@0df212b · backends: portable, kernels

  • Paper Rewrite the Stars CVPR 2024 · first released 29 Mar 2024
  • Authors’ code ma-xu/Rewrite-the-Stars (PyTorch), not the code forked here: timm carries the architecture and the converted authors' checkpoints in its own, simpler API.
Checkpoint

Paper 2024 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 0.683 ms 224x224, fp16
vs the faster original 6.7× original fp32: 4.57 ms
Peak memory 41 MiB original: 51 MiB
Accuracy vs original fp32 +0.02 kernels top-1 %, ImageNet-1k val (74.50 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install "fastforks-core[timm]"

from fastforks import classification

model = classification.load("starnet_s1.in1k")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.logits
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench starnet_s1.in1k

Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.

Speed and memory

Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.

Input: synthetic, 224x224 (the task’s standard size)

BackenddtypeTime per call (ms)vs original fp32Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32 4.57–51 MiB/58 MiB2.80.30
Original fp16torch.autocast(float16) around the callfp16 7.290.63×51 MiB/60 MiB0.20.26
Portable fp32fp32not measured ––––
Portablefp16 3.881.2×41 MiB/52 MiB1.60.24
Kernelsfp16 0.6836.7×41 MiB/56 MiB2.20.21
Portable + torch.compile 2.411.9×73 MiB/92 MiB6 (build)–
TensorRT fp16fp16 0.7745.9×34 MiBenginenot comparable24 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-11. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.

Other devices

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14Originalfp3213.0 436 MiB process–224x224
Portablefp1614.0 400 MiB processmatches224x224
Portablefp328.33 424 MiB processmatches224x224
Kernelsfp16not implemented–224x224

CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.

Licence and weights

Commercial use allowed

Code
Apache-2.0
Weights
Apache-2.0 source
Training data

Our reading of the licences, not legal advice.

Weights

starnet_s1.in1k.safetensors 11.5 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download starnet_s1.in1k).

Accuracy

The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.

ImageNet-1k val

5000 random images, top-1 / top-5 %

Rundtypetop-1 % ↑top-5 % ↑Δ top-1 % vs fp32 originalimagesms / image
Original fp32fp3274.5091.94 reference 5000 0.534
Original fp16fp1674.5091.96 ±0.00 5000 0.411
Portable fp32fp3274.4891.92 −0.02 5000 0.931
Portablefp1674.4891.92 −0.02 5000 0.646
Kernelsfp1674.5291.92 +0.02 5000 0.324

fastforks: the model's own resize / crop on the GPU; original: the library's pretrained loader and its PIL eval transform. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-11; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 9e68b313, 2026-11-11.

Closeness to the original

Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. 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).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.000011 not recorded
GPU default dtypes, mean |diff|≤ 0.0035 0.00142
GPU default dtypes, p99 |diff|≤ 0.012 0.00462

Recorded for starnet_s1.in1k on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-11, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

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

Input / output
Input: one RGB image. Output: ImageNet-1k class logits.
Input size
fixed 224x224: the input is resized (short side 256) and centre-cropped to it; tested up to 224x224 (41 MiB allocated)
Outputs
logits
Checkpoint
arch="starnet_s1.in1k"
CUDA kernels
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
  • fvm_convfused convolutions for image classifiers
  • fvm_extrafused transformer row operations (rotary split, SwiGLU with LayerNorm)
  • fvm_mbfused MBConv blocks (MobileNet and EfficientNet families)
  • fvm_r3further fused classifier blocks
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

StarNet (Ma et al. 2024; S1 to S4, ImageNet-1k): a small convolutional classifier whose blocks multiply two branches element-wise. Apache-2.0.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
starnet_s1.in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
4.57
7.29
3.88
0.683
2.41
0.774
51 / 58
51 / 60
41 / 52
41 / 56
73 / 92
73 / 86
starnet_s2.in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
3.64
5.63
3.02
0.578
1.96
0.622
56 / 80
55 / 64
43 / 54
43 / 58
75 / 94
74 / 88
starnet_s3.in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
4.96
7.67
4.19
0.715
2.60
0.878
64 / 88
67 / 76
47 / 60
47 / 64
79 / 100
80 / 96
starnet_s4.in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
6.84
11.3
5.63
0.912
3.37
1.18
71 / 94
77 / 86
50 / 66
51 / 72
82 / 106
85 / 104
Details
  • Original code: https://github.com/huggingface/pytorch-image-models at 0df212b369a5
  • Code licence: Apache-2.0 (permissive); files:
  • Weights licence: Apache-2.0
  • Licence evidence: https://huggingface.co/timm/starnet_s1.in1k/blob/e3fbe6d1e24b6293fa6e7b897f380bd62f77e01f/README.md
  • Training data: ImageNet-1k (ILSVRC 2012 classification)
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: starnet_s1.in1k (default of starnet_s1), starnet_s2.in1k (default of starnet_s2), starnet_s3.in1k (default of starnet_s3), starnet_s4.in1k (default of starnet_s4)
  • Outputs: logits; label_set: imagenet-1k
  • 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)
  • starnet_s1.in1k 224x224, first call: portable + torch.compile 6 s (cold Inductor cache), TensorRT fp16 24 s (ONNX + engine).
  • starnet_s2.in1k 224x224, first call: portable + torch.compile 6 s (cold Inductor cache), TensorRT fp16 25 s (ONNX + engine).
  • starnet_s3.in1k 224x224, first call: portable + torch.compile 6 s (cold Inductor cache), TensorRT fp16 28 s (ONNX + engine).
  • starnet_s4.in1k 224x224, first call: portable + torch.compile 7 s (cold Inductor cache), TensorRT fp16 28 s (ONNX + engine).
  • starnet_s1.in1k original 224x224: torch.autocast(float16) around the call
  • starnet_s2.in1k original 224x224: torch.autocast(float16) around the call
  • starnet_s3.in1k original 224x224: torch.autocast(float16) around the call
  • starnet_s4.in1k original 224x224: torch.autocast(float16) around the call
  • starnet_s1.in1k load / first call: original 0.18 s / 0.26 s, portable 1.6 s / 0.241 s, kernels 2.2 s / 0.214 s
  • starnet_s2.in1k load / first call: original 0.17 s / 0.021 s, portable 1.62 s / 0.24 s, kernels 1.86 s / 0.06 s
  • starnet_s3.in1k load / first call: original 0.3 s / 0.01 s, portable 1.6 s / 0.239 s, kernels 1.82 s / 0.052 s
  • starnet_s4.in1k load / first call: original 0.21 s / 0.013 s, portable 1.66 s / 0.245 s, kernels 2.0 s / 0.051 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

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.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
starnet_s1.in1kImageNet-1k val, 5000 random images, top-1 / top-5 %74.50 / 91.9474.50 / 91.9674.48 / 91.9274.48 / 91.9274.52 / 91.92
starnet_s2.in1kImageNet-1k val, 5000 random images, top-1 / top-5 %75.82 / 92.1475.84 / 92.1275.84 / 92.1075.82 / 92.1075.88 / 92.12
starnet_s3.in1kImageNet-1k val, 5000 random images, top-1 / top-5 %78.08 / 93.7278.14 / 93.7478.10 / 93.8078.06 / 93.8278.06 / 93.80
starnet_s4.in1kImageNet-1k val, 5000 random images, top-1 / top-5 %80.08 / 94.5680.06 / 94.5679.98 / 94.5879.98 / 94.5879.92 / 94.58

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 9e68b313, 2026-11-11; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 9e68b313, 2026-11-11.

Parity

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).

variantkernels error vs original, mean / p99GPU, torchkernels used
starnet_s1.in1k0.00142 / 0.00462RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
starnet_s2.in1k0.00143 / 0.00519RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
starnet_s3.in1k0.00187 / 0.00719RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
starnet_s4.in1k0.00229 / 0.00742RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Input

Top 5

  1. Egyptian cat 52%
  2. tiger cat 17%
  3. tabby 11%
  4. lynx 0.4%
  5. Band Aid 0.3%
Chelsea the cat, input
Input · 100%

Input: Chelsea the cat, Stéfan van der Walt (CC0).