HGNetV2 (fastforks)
HGNetV2 (PaddlePaddle, PP-HGNetV2): a convolutional classifier of aggregation blocks with light 5x5 convolutions, SSLD-distilled and fine-tuned on ImageNet-1k.
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/hgnetv2_b0.ssld_stage2_ft_in1k, https://huggingface.co/timm/hgnetv2_b1.ssld_stage2_ft_in1k, https://huggingface.co/timm/hgnetv2_b2.ssld_stage2_ft_in1k, https://huggingface.co/timm/hgnetv2_b3.ssld_stage2_ft_in1k, https://huggingface.co/timm/hgnetv2_b4.ssld_stage2_ft_in1k, https://huggingface.co/timm/hgnetv2_b5.ssld_stage2_ft_in1k, https://huggingface.co/timm/hgnetv2_b6.ssld_stage2_ft_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 |
|---|---|---|---|---|
hgnetv2_b0.ssld_stage2_ft_in1k@1 |
hgnetv2_b0.ssld_stage2_ft_in1k.safetensors | a84094d2a9ed |
Apache-2.0 | yes |
hgnetv2_b1.ssld_stage2_ft_in1k@1 |
hgnetv2_b1.ssld_stage2_ft_in1k.safetensors | 4a9d49b24bca |
Apache-2.0 | yes |
hgnetv2_b2.ssld_stage2_ft_in1k@1 |
hgnetv2_b2.ssld_stage2_ft_in1k.safetensors | 184700043cfb |
Apache-2.0 | yes |
hgnetv2_b3.ssld_stage2_ft_in1k@1 |
hgnetv2_b3.ssld_stage2_ft_in1k.safetensors | f9769b32ab1f |
Apache-2.0 | yes |
hgnetv2_b4.ssld_stage2_ft_in1k@1 |
hgnetv2_b4.ssld_stage2_ft_in1k.safetensors | 8d05f750c530 |
Apache-2.0 | yes |
hgnetv2_b5.ssld_stage2_ft_in1k@1 |
hgnetv2_b5.ssld_stage2_ft_in1k.safetensors | 839782a6a43c |
Apache-2.0 | yes |
hgnetv2_b6.ssld_stage2_ft_in1k@1 |
hgnetv2_b6.ssld_stage2_ft_in1k.safetensors | 4febffb69be0 |
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-21k / ImageNet-22k (full ImageNet, fall 2011 release), ImageNet-1k (ILSVRC 2012 classification).
Licence evidence: https://huggingface.co/timm/hgnetv2_b0.ssld_stage2_ft_in1k/blob/9e0a904b210d65e4fcd6582f8ca6e0921fac2650/README.md
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)
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 |
|---|---|---|---|---|---|---|---|---|---|---|
hgnetv2_b0.ssld_stage2_ft_in1k |
224x224 | 4.45 | 5.89 | 3.46 | β | β | β | 68 / 72 | 48 / 70 | β |
hgnetv2_b1.ssld_stage2_ft_in1k |
224x224 | 4.57 | 6.17 | 3.47 | β | β | β | 70 / 92 | 50 / 68 | β |
hgnetv2_b2.ssld_stage2_ft_in1k |
224x224 | 6.25 | 7.98 | 4.55 | β | β | β | 98 / 118 | 59 / 94 | β |
hgnetv2_b3.ssld_stage2_ft_in1k |
224x224 | 7.16 | 9.30 | 5.22 | β | β | β | 127 / 140 | 76 / 114 | β |
hgnetv2_b4.ssld_stage2_ft_in1k |
224x224 | 7.06 | 8.41 | 4.53 | β | β | β | 147 / 166 | 82 / 116 | β |
hgnetv2_b5.ssld_stage2_ft_in1k |
224x224 | 11.4 | 13.8 | 7.04 | β | β | β | 260 / 280 | 114 / 150 | β |
hgnetv2_b6.ssld_stage2_ft_in1k |
224x224 | 16.5 | 17.9 | 9.44 | β | β | β | 465 / 490 | 198 / 224 | β |
Startup and first call
| model | load (s): original | portable | kernels | first call (s): original | portable | kernels |
|---|---|---|---|---|---|---|
hgnetv2_b0.ssld_stage2_ft_in1k |
0.19 | 1.67 | β | 1.36 | 0.73 | β |
hgnetv2_b1.ssld_stage2_ft_in1k |
0.31 | 1.87 | β | 0.31 | 0.40 | β |
hgnetv2_b2.ssld_stage2_ft_in1k |
0.23 | 1.79 | β | 0.59 | 0.39 | β |
hgnetv2_b3.ssld_stage2_ft_in1k |
0.27 | 1.81 | β | 0.17 | 0.40 | β |
hgnetv2_b4.ssld_stage2_ft_in1k |
0.32 | 1.74 | β | 0.30 | 0.39 | β |
hgnetv2_b5.ssld_stage2_ft_in1k |
0.49 | 1.95 | β | 0.30 | 0.40 | β |
hgnetv2_b6.ssld_stage2_ft_in1k |
0.86 | 2.21 | β | 0.45 | 0.46 | β |
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) |
|---|---|---|---|---|---|---|---|---|---|---|---|
hgnetv2_b0.ssld_stage2_ft_in1k |
224x224 | 3.46 ms | 1.78 ms | 0.334 ms | β | 6.74 s | 21.2 s | 48 / 70 | 78 / 110 | β | 77 / 100 |
hgnetv2_b1.ssld_stage2_ft_in1k |
224x224 | 3.47 ms | 1.84 ms | 0.307 ms | β | 7.07 s | 20.1 s | 50 / 68 | 80 / 108 | β | 81 / 104 |
hgnetv2_b2.ssld_stage2_ft_in1k |
224x224 | 4.55 ms | 2.33 ms | 0.383 ms | β | 7.65 s | 23.4 s | 59 / 94 | 89 / 134 | β | 91 / 128 |
hgnetv2_b3.ssld_stage2_ft_in1k |
224x224 | 5.22 ms | 2.55 ms | 0.436 ms | β | 7.99 s | 24.6 s | 76 / 114 | 99 / 152 | β | 100 / 150 |
hgnetv2_b4.ssld_stage2_ft_in1k |
224x224 | 4.53 ms | 2.66 ms | 0.463 ms | β | 7.84 s | 19.8 s | 82 / 116 | 107 / 134 | β | 139 / 182 |
hgnetv2_b5.ssld_stage2_ft_in1k |
224x224 | 7.04 ms | 4.13 ms | 0.758 ms | β | 9.97 s | 23.8 s | 114 / 150 | 146 / 190 | β | 209 / 250 |
hgnetv2_b6.ssld_stage2_ft_in1k |
224x224 | 9.44 ms | 5.44 ms | 1.42 ms | β | 11.2 s | 26.9 s | 198 / 224 | 217 / 264 | β | 347 / 394 |
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 |
|---|---|---|---|---|---|
hgnetv2_b0.ssld_stage2_ft_in1k |
ImageNet-1k val, 5000 random images, top-1 / top-5 % | 78.50 / 93.98 | 78.40 / 93.96 | 78.50 / 94.04 | |
hgnetv2_b1.ssld_stage2_ft_in1k |
ImageNet-1k val, 5000 random images, top-1 / top-5 % | 80.08 / 94.40 | 80.00 / 94.46 | 80.12 / 94.38 | |
hgnetv2_b2.ssld_stage2_ft_in1k |
ImageNet-1k val, 5000 random images, top-1 / top-5 % | 82.50 / 95.92 | 82.48 / 95.92 | 82.50 / 95.92 | |
hgnetv2_b3.ssld_stage2_ft_in1k |
ImageNet-1k val, 5000 random images, top-1 / top-5 % | 83.68 / 96.08 | 83.68 / 96.08 | 83.66 / 96.02 | |
hgnetv2_b4.ssld_stage2_ft_in1k |
ImageNet-1k val, 5000 random images, top-1 / top-5 % | 83.96 / 96.66 | 83.98 / 96.66 | 84.00 / 96.64 | |
hgnetv2_b5.ssld_stage2_ft_in1k |
ImageNet-1k val, 5000 random images, top-1 / top-5 % | 85.50 / 97.14 | 85.52 / 97.14 | 85.52 / 97.18 | |
hgnetv2_b6.ssld_stage2_ft_in1k |
ImageNet-1k val, 5000 random images, top-1 / top-5 % | 86.60 / 97.94 | 86.60 / 97.94 | 86.74 / 98.00 |
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| <= 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).
Usage
from fastforks import classification
model = classification.load("hgnetv2_b0.ssld_stage2_ft_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: hgnetv2_b0.ssld_stage2_ft_in1k, hgnetv2_b1.ssld_stage2_ft_in1k, hgnetv2_b2.ssld_stage2_ft_in1k, hgnetv2_b3.ssld_stage2_ft_in1k, hgnetv2_b4.ssld_stage2_ft_in1k, hgnetv2_b5.ssld_stage2_ft_in1k, hgnetv2_b6.ssld_stage2_ft_in1k.
Kernels backend
No kernels backend: not built yet: the portable CNN path is validated; HGNetV2 aggregation blocks need a dedicated kernels runner.