rockerritesh/EUPE-ONNX

EUPE — All 6 Models in ONNX (FP32 + INT8)

ONNX exports of Meta AI's Efficient Universal Perception Encoder (EUPE) — a single lightweight vision backbone that matches or exceeds domain-specialist models on diverse tasks, designed for on-device / edge deployment.

• 📄 Paper: arxiv 2603.22387
• 🔥 Original weights: facebook/EUPE collection
• 💻 Code: facebookresearch/EUPE

---

Files

File	Architecture	Params	Size	Type
eupe_vitt16.onnx	ViT-T/16	6M	22.1 MB	FP32
eupe_vitt16_int8.onnx	ViT-T/16	6M	5.9 MB	INT8 ✅ smallest
eupe_vits16.onnx	ViT-S/16	21M	86.6 MB	FP32
eupe_vits16_int8.onnx	ViT-S/16	21M	22.2 MB	INT8
eupe_vitb16.onnx	ViT-B/16	86M	342.8 MB	FP32
eupe_vitb16_int8.onnx	ViT-B/16	86M	86.4 MB	INT8
eupe_convnext-tiny.onnx	ConvNeXt-T	29M	111.4 MB	FP32
eupe_convnext-tiny_int8.onnx	ConvNeXt-T	29M	28.2 MB	INT8
eupe_convnext-small.onnx	ConvNeXt-S	50M	198.0 MB	FP32
eupe_convnext-small_int8.onnx	ConvNeXt-S	50M	50.2 MB	INT8
eupe_convnext-base.onnx	ConvNeXt-B	89M	350.4 MB	FP32
eupe_convnext-base_int8.onnx	ConvNeXt-B	89M	88.5 MB	INT8

INT8 models are ~75% smaller than FP32 with negligible accuracy loss.

---

Inputs & Outputs

ViT models (vitt16, vits16, vitb16)

	Name	Shape	dtype
Input	input	[batch, 3, 224, 224]	float32
Output 0	cls_token	[batch, D]	float32
Output 1	patch_tokens	[batch, 196, D]	float32

Where D = 192 (ViT-T) / 384 (ViT-S) / 768 (ViT-B)

ConvNeXt models (convnext-tiny, convnext-small, convnext-base)

	Name	Shape	dtype
Input	input	[batch, 3, 224, 224]	float32
Output	features	[batch, D]	float32

Where D = 768 (Tiny/Small) / 1024 (Base)

Preprocessing: ImageNet normalisation — mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225]

---

Quick Start

import onnxruntime as ort
import numpy as np
from PIL import Image
from torchvision.transforms import v2
import torch

# Load model
sess = ort.InferenceSession(
    "eupe_vitt16_int8.onnx",
    providers=["CPUExecutionProvider"]
)

# Preprocess image
transform = v2.Compose([
    v2.ToImage(),
    v2.Resize((224, 224), antialias=True),
    v2.ToDtype(torch.float32, scale=True),
    v2.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
])
img = Image.open("image.jpg").convert("RGB")
inp = transform(img)[None].numpy()   # (1, 3, 224, 224)

# Inference
cls_token, patch_tokens = sess.run(None, {"input": inp})
print(cls_token.shape)     # (1, 192)
print(patch_tokens.shape)  # (1, 196, 192)

---

What can you do with the outputs?

Task	How
Image classification	k-NN on cls_token
Image similarity / retrieval	Cosine similarity of cls_token
Depth estimation	Linear layer on patch_tokens
Semantic segmentation	Linear layer on patch_tokens
Visual QA	Feed patch_tokens into a language model

---

Edge Deployment Targets

Platform	Recommended
Android / iOS	INT8 + ONNX Runtime Mobile
Raspberry Pi / Jetson Nano	INT8
Browser	FP32 via onnxruntime-web
Server / Cloud	FP32

---

Conversion Details

• Exported with torch.onnx.export (legacy TorchScript path, opset 16)
• Quantized with onnxruntime.quantization.quantize_dynamic (QInt8 weights)
• Validated: max absolute diff FP32 vs ONNX < 0.0001 on all models

---

License

FAIR Noncommercial Research License. See original repo for details.