Models · Classification

HGNetV2

HGNetV2 (PaddlePaddle's PP-HGNetV2; B0 to B6, SSLD-distilled, fine-tuned on ImageNet-1k): a convolutional classifier built for GPU speed. Apache-2.0.

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

  • Paper PP-HGNetV2 PaddleClas docs · first released 26 Sept 2023
  • Authors’ code PaddlePaddle/PaddleClas (PaddlePaddle), not the code forked here: timm carries the architecture and the converted PaddleClas checkpoints in its own PyTorch API; the original is PaddlePaddle.
Checkpoint

Paper 2023 Added in 0.2.0 Evaluated with 0.2.0

Portable, RTX 5060 Ti 16 GB 3.46 ms 224x224, fp16
vs the faster original 1.3× original fp32: 4.45 ms
Peak memory 48 MiB original: 65 MiB
Accuracy vs original fp32 ±0.00 portable top-1 %, ImageNet-1k val (78.50 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("hgnetv2_b0.ssld_stage2_ft_in1k")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.logits
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench hgnetv2_b0.ssld_stage2_ft_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 4.45–65 MiB/70 MiB3.20.33
Original fp16torch.autocast(float16) around the callfp16 5.890.76×68 MiB/72 MiB0.21.36
Portable fp32fp32not measured ––––
Portablefp16 3.461.3×48 MiB/70 MiB1.70.73
Kernelsnot measured ––––
Portable + torch.compile 1.782.5×78 MiB/110 MiB7 (build)–
TensorRT fp16fp16 0.33413×32 MiBenginenot comparable21 (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, 2026-11-14Originalfp3211.6 480 MiB process–224x224
Portablefp167.01 513 MiB processmatches224x224
Portablefp326.34 503 MiB processmatches224x224
Kernelsfp32not 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

hgnetv2_b0.ssld_stage2_ft_in1k.safetensors 24.1 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download hgnetv2_b0.ssld_stage2_ft_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 fp32fp3278.5093.98 reference 5000 0.474
Original fp16fp1678.4093.96 −0.10 5000 0.492
Portable fp32fp3278.5094.04 ±0.00 5000 1.15
Portablefp1678.5094.04 ±0.00 5000 0.535
Kernelsnot measured

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: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 7ba7307f, 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| <= 1.8e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.008 and p99 <= 0.033 (absolute error of the output logits).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.000018 not recorded
GPU default dtypes, mean |diff|≤ 0.008 0.0054
GPU default dtypes, p99 |diff|≤ 0.033 0.02154

Recorded for hgnetv2_b0.ssld_stage2_ft_in1k on RTX 5060 Ti, torch 2.10.0+cu130, 2026-11-17, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

HGNetV2 (PaddlePaddle, PP-HGNetV2): a convolutional classifier of aggregation blocks with light 5x5 convolutions, SSLD-distilled and fine-tuned on ImageNet-1k.

Input / output
Input: one RGB image. Output: ImageNet-1k class logits.
Input size
fixed 224x224: the input is resized (short side 232) and centre-cropped to it; tested up to 224x224 (48 MiB allocated)
Outputs
logits
Checkpoint
arch="hgnetv2_b0.ssld_stage2_ft_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

HGNetV2 (PaddlePaddle's PP-HGNetV2; B0 to B6, SSLD-distilled, fine-tuned on ImageNet-1k): a convolutional classifier built for GPU speed. Apache-2.0.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
hgnetv2_b0.ssld_stage2_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
4.45
5.89
3.46
–
1.78
0.334
65 / 70
68 / 72
48 / 70
–
78 / 110
77 / 100
hgnetv2_b1.ssld_stage2_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
4.57
6.17
3.47
–
1.84
0.307
66 / 94
70 / 92
50 / 68
–
80 / 108
81 / 104
hgnetv2_b2.ssld_stage2_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
6.25
7.98
4.55
–
2.33
0.383
89 / 114
98 / 118
59 / 94
–
89 / 134
91 / 128
hgnetv2_b3.ssld_stage2_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
7.16
9.30
5.22
–
2.55
0.436
112 / 134
127 / 140
76 / 114
–
99 / 152
100 / 150
hgnetv2_b4.ssld_stage2_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
7.06
8.41
4.53
–
2.66
0.463
131 / 140
147 / 166
82 / 116
–
107 / 134
139 / 182
hgnetv2_b5.ssld_stage2_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
11.4
13.8
7.04
–
4.13
0.758
213 / 222
260 / 280
114 / 150
–
146 / 190
209 / 250
hgnetv2_b6.ssld_stage2_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
16.5
17.9
9.44
–
5.44
1.42
354 / 376
465 / 490
198 / 224
–
217 / 264
347 / 394
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/hgnetv2_b0.ssld_stage2_ft_in1k/blob/9e0a904b210d65e4fcd6582f8ca6e0921fac2650/README.md
  • Training data: ImageNet-21k / ImageNet-22k (full ImageNet, fall 2011 release), ImageNet-1k (ILSVRC 2012 classification)
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: hgnetv2_b0.ssld_stage2_ft_in1k (default of hgnetv2_b0), hgnetv2_b1.ssld_stage2_ft_in1k (default of hgnetv2_b1), hgnetv2_b2.ssld_stage2_ft_in1k (default of hgnetv2_b2), hgnetv2_b3.ssld_stage2_ft_in1k (default of hgnetv2_b3), hgnetv2_b4.ssld_stage2_ft_in1k (default of hgnetv2_b4), hgnetv2_b5.ssld_stage2_ft_in1k (default of hgnetv2_b5), hgnetv2_b6.ssld_stage2_ft_in1k (default of hgnetv2_b6)
  • Outputs: logits; label_set: imagenet-1k
  • Input size: fixed 224x224: the input is resized (short side 232) and centre-cropped to it
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): hgnetv2_b0.ssld_stage2_ft_in1k 224x224 (0.0 GiB peak allocated, 0.1 reserved); hgnetv2_b1.ssld_stage2_ft_in1k 224x224 (0.0 GiB peak allocated, 0.1 reserved); hgnetv2_b2.ssld_stage2_ft_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved); hgnetv2_b3.ssld_stage2_ft_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved); hgnetv2_b4.ssld_stage2_ft_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved); hgnetv2_b5.ssld_stage2_ft_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved); hgnetv2_b6.ssld_stage2_ft_in1k 224x224 (0.2 GiB peak allocated, 0.2 reserved)
  • hgnetv2_b0.ssld_stage2_ft_in1k 224x224, first call: portable + torch.compile 7 s (cold Inductor cache), TensorRT fp16 21 s (ONNX + engine).
  • hgnetv2_b1.ssld_stage2_ft_in1k 224x224, first call: portable + torch.compile 7 s (cold Inductor cache), TensorRT fp16 20 s (ONNX + engine).
  • hgnetv2_b2.ssld_stage2_ft_in1k 224x224, first call: portable + torch.compile 8 s (cold Inductor cache), TensorRT fp16 23 s (ONNX + engine).
  • hgnetv2_b3.ssld_stage2_ft_in1k 224x224, first call: portable + torch.compile 8 s (cold Inductor cache), TensorRT fp16 25 s (ONNX + engine).
  • hgnetv2_b4.ssld_stage2_ft_in1k 224x224, first call: portable + torch.compile 8 s (cold Inductor cache), TensorRT fp16 20 s (ONNX + engine).
  • hgnetv2_b5.ssld_stage2_ft_in1k 224x224, first call: portable + torch.compile 10 s (cold Inductor cache), TensorRT fp16 24 s (ONNX + engine).
  • hgnetv2_b6.ssld_stage2_ft_in1k 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 27 s (ONNX + engine).
  • hgnetv2_b0.ssld_stage2_ft_in1k original 224x224: torch.autocast(float16) around the call
  • hgnetv2_b1.ssld_stage2_ft_in1k original 224x224: torch.autocast(float16) around the call
  • hgnetv2_b2.ssld_stage2_ft_in1k original 224x224: torch.autocast(float16) around the call
  • hgnetv2_b3.ssld_stage2_ft_in1k original 224x224: torch.autocast(float16) around the call
  • hgnetv2_b4.ssld_stage2_ft_in1k original 224x224: torch.autocast(float16) around the call
  • hgnetv2_b5.ssld_stage2_ft_in1k original 224x224: torch.autocast(float16) around the call
  • hgnetv2_b6.ssld_stage2_ft_in1k original 224x224: torch.autocast(float16) around the call
  • hgnetv2_b0.ssld_stage2_ft_in1k load / first call: original 0.19 s / 1.357 s, portable 1.67 s / 0.735 s
  • hgnetv2_b1.ssld_stage2_ft_in1k load / first call: original 0.31 s / 0.313 s, portable 1.87 s / 0.396 s
  • hgnetv2_b2.ssld_stage2_ft_in1k load / first call: original 0.23 s / 0.59 s, portable 1.79 s / 0.39 s
  • hgnetv2_b3.ssld_stage2_ft_in1k load / first call: original 0.27 s / 0.165 s, portable 1.81 s / 0.395 s
  • hgnetv2_b4.ssld_stage2_ft_in1k load / first call: original 0.32 s / 0.297 s, portable 1.74 s / 0.389 s
  • hgnetv2_b5.ssld_stage2_ft_in1k load / first call: original 0.49 s / 0.305 s, portable 1.95 s / 0.4 s
  • hgnetv2_b6.ssld_stage2_ft_in1k load / first call: original 0.86 s / 0.453 s, portable 2.21 s / 0.46 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 fp32note
hgnetv2_b0.ssld_stage2_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %78.50 / 93.9878.40 / 93.9678.50 / 94.0478.50 / 94.04
hgnetv2_b1.ssld_stage2_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %80.08 / 94.4080.00 / 94.4680.12 / 94.3880.16 / 94.42
hgnetv2_b2.ssld_stage2_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %82.50 / 95.9282.48 / 95.9282.50 / 95.9282.50 / 95.92
hgnetv2_b3.ssld_stage2_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %83.68 / 96.0883.68 / 96.0883.66 / 96.0283.64 / 96.06
hgnetv2_b4.ssld_stage2_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %83.96 / 96.6683.98 / 96.6684.00 / 96.6484.02 / 96.64
hgnetv2_b5.ssld_stage2_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %85.50 / 97.1485.52 / 97.1485.52 / 97.1885.52 / 97.16
hgnetv2_b6.ssld_stage2_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %86.60 / 97.9486.60 / 97.9486.74 / 98.0086.66 / 98.00

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 7ba7307f, 2026-11-11; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17.

Parity

CPU, float32: max |diff| <= 1.8e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.008 and p99 <= 0.033 (absolute error of the output logits).

variantkernels error vs original, mean / p99GPU, torchkernels used
hgnetv2_b0.ssld_stage2_ft_in1k0.0054 / 0.02154RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
hgnetv2_b1.ssld_stage2_ft_in1k0.00362 / 0.01274RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
hgnetv2_b2.ssld_stage2_ft_in1k0.00244 / 0.01084RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
hgnetv2_b3.ssld_stage2_ft_in1k0.00242 / 0.01019RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
hgnetv2_b4.ssld_stage2_ft_in1k0.00152 / 0.00782RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
hgnetv2_b5.ssld_stage2_ft_in1k0.00146 / 0.00686RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
hgnetv2_b6.ssld_stage2_ft_in1k0.00201 / 0.00906RTX 5060 Ti, torch 2.10.0+cu130cudnn_hacc@cudnn91501, 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 (portable · float16 · NVIDIA GeForce RTX 5060 Ti).

Input

Top 5

  1. Egyptian cat 40%
  2. tiger cat 31%
  3. tabby 28%
  4. lynx 0.4%
  5. shower curtain <0.1%
Chelsea the cat, input
Input · 100%

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