efficientformerv2 / README.md
Flor's picture
card: update from fastforks 1cbf1671
5b77504 verified
|
Raw History Blame Contribute Delete
8.23 kB
metadata
library_name: fastforks
pipeline_tag: image-classification
license: apache-2.0
tags:
  - classification
  - image-classification
  - cuda
  - pytorch
  - fastforks
  - fastforks-card

EfficientFormerV2 (fastforks)

EfficientFormerV2 (Li et al. 2022, Snap Research): a mobile hybrid classifier, convolutional stages with attention in the last two.

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/efficientformerv2_l.snap_dist_in1k, https://huggingface.co/timm/efficientformerv2_s0.snap_dist_in1k, https://huggingface.co/timm/efficientformerv2_s1.snap_dist_in1k, https://huggingface.co/timm/efficientformerv2_s2.snap_dist_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
efficientformerv2_l.snap_dist_in1k@1 efficientformerv2_l.snap_dist_in1k.safetensors a92de90055b1 Apache-2.0 yes
efficientformerv2_s0.snap_dist_in1k@1 efficientformerv2_s0.snap_dist_in1k.safetensors ad4ce93472a2 Apache-2.0 yes
efficientformerv2_s1.snap_dist_in1k@1 efficientformerv2_s1.snap_dist_in1k.safetensors 3ef67aa25321 Apache-2.0 yes
efficientformerv2_s2.snap_dist_in1k@1 efficientformerv2_s2.snap_dist_in1k.safetensors c46c28768173 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/efficientformerv2_s0.snap_dist_in1k/blob/aa226e91a076e760bc7957e5990f19f6ae322c5f/README.md

Input size

fixed 224x224: the input is resized (short side 235) and centre-cropped to it Tested up to (portable, fp16, RTX 5060 Ti 16 GB): efficientformerv2_l.snap_dist_in1k 224x224 (0.1 GiB peak allocated, 0.1 reserved); efficientformerv2_s0.snap_dist_in1k 224x224 (0.0 GiB peak allocated, 0.1 reserved); efficientformerv2_s1.snap_dist_in1k 224x224 (0.0 GiB peak allocated, 0.1 reserved); efficientformerv2_s2.snap_dist_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
efficientformerv2_l.snap_dist_in1k 224x224 19.9 29.0 17.1 – – – 186 / 198 93 / 144 –
efficientformerv2_s0.snap_dist_in1k 224x224 8.16 11.8 7.62 – – – 55 / 64 44 / 64 –
efficientformerv2_s1.snap_dist_in1k 224x224 10.1 14.9 8.91 – – – 70 / 80 50 / 88 –
efficientformerv2_s2.snap_dist_in1k 224x224 15.2 21.8 13.9 – – – 108 / 116 64 / 104 –

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
efficientformerv2_l.snap_dist_in1k 0.32 2.74 – 0.73 0.43 –
efficientformerv2_s0.snap_dist_in1k 0.17 1.88 – 1.76 0.50 –
efficientformerv2_s1.snap_dist_in1k 0.19 1.96 – 0.59 0.43 –
efficientformerv2_s2.snap_dist_in1k 0.24 2.68 – 0.86 0.44 –

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)
efficientformerv2_l.snap_dist_in1k 224x224 17.1 ms 7.32 ms 1.20 ms – 20.7 s 45.1 s 93 / 144 125 / 184 – 161 / 214
efficientformerv2_s0.snap_dist_in1k 224x224 7.62 ms 3.34 ms 0.550 ms – 11.3 s 36.3 s 44 / 64 76 / 104 – 78 / 100
efficientformerv2_s1.snap_dist_in1k 224x224 8.91 ms 3.91 ms 0.632 ms – 12.7 s 41.2 s 50 / 88 82 / 108 – 86 / 126
efficientformerv2_s2.snap_dist_in1k 224x224 13.9 ms 5.55 ms 0.919 ms – 16.4 s 38.9 s 64 / 104 96 / 124 – 100 / 142

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
efficientformerv2_s0.snap_dist_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 77.04 / 92.94 77.00 / 92.76 77.34 / 92.82
efficientformerv2_s1.snap_dist_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 80.80 / 95.00 81.02 / 94.88 80.88 / 95.00
efficientformerv2_s2.snap_dist_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 83.22 / 96.28 83.10 / 96.34 83.18 / 96.30
efficientformerv2_l.snap_dist_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 83.88 / 96.58 83.96 / 96.58 84.02 / 96.50

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| <= 4.7e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.12 and p99 <= 0.42 (absolute error of the output logits).

Usage

from fastforks import classification

model = classification.load("efficientformerv2_l.snap_dist_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: efficientformerv2_l.snap_dist_in1k, efficientformerv2_s0.snap_dist_in1k, efficientformerv2_s1.snap_dist_in1k, efficientformerv2_s2.snap_dist_in1k.

Kernels backend

No kernels backend: not built yet: Conv-BN blocks with depthwise 3x3, feature-map attention with learned bias tables and strided depthwise convolutions need a dedicated kernels runner.