CoreEpoch/tinyvit-5m-int8-imagenet

TinyViT-5M — INT8 ONNX, 80.5% ImageNet in 9.2 MB

TinyViT-5M (timm/tiny_vit_5m_224.dist_in22k_ft_in1k, Apache-2.0) quantized to INT8 with Kenosis — 128-image calibration, no retraining. 80.53% top-1 from a 9.2 MB single file, on ONNX Runtime or OpenVINO, CPU or GPU, no accelerator required.

Accuracy

ImageNet-1K validation, 49,872 images (disjoint from the 128 calibration images).

model	top-1	Δ vs FP32	size
FP32 baseline	80.87%	—	22.1 MB
Kenosis quantized	80.53%	−0.34	9.23 MB

Measured on a CPU with AVX-512 VNNI; on CPUs without VNNI this model's INT8 top-1 sits ~0.9 below FP32 rather than 0.34 (its per-channel-quantized layers are VNNI-sensitive).

Run

from huggingface_hub import hf_hub_download
import numpy as np, onnxruntime as ort
from PIL import Image
path = hf_hub_download("CoreEpoch/tinyvit-5m-int8-imagenet", "tinyvit_5m_224_int8_kenosis.onnx")
sess = ort.InferenceSession(path, providers=["CPUExecutionProvider"])
img = Image.open("your_image.jpg").convert("RGB")
scale = 236 / min(img.size)  # shorter side to 236, center crop to 224 — the measured transform
img = img.resize((round(img.width*scale), round(img.height*scale)), Image.BICUBIC)
l, t = (img.width-224)//2, (img.height-224)//2
img = img.crop((l, t, l+224, t+224))
x = (np.asarray(img, np.float32)/255.0 - [0.485,0.456,0.406]) / [0.229,0.224,0.225]
logits = sess.run(None, {"input": x.transpose(2,0,1)[None].astype(np.float32)})[0]
print(int(np.argmax(logits)))

Input 1x3x224x224, RGB, /255, ImageNet mean/std. Output logits [1,1000], sorted-synset order. run_classify.py / eval_imagenet.py reproduce the demo and table.

Integrity & license

tinyvit_5m_224_int8_kenosis.onnx (9,228,567 B) — SHA-256 EDA9007C0449F118F02B1E8C4BA6B43BEC51A7D7AB677566D42929EB3DA02874. Apache-2.0 (base TinyViT, Wu et al. 2022, retained). Quantized with Kenosis (patent pending) · coreepoch.dev