ipsilondev/calorie-clip-onnx

CalorieCLIP ONNX

ONNX export of CalorieCLIP for food calorie estimation.

Models

File	Size	Precision	Use Case
calorie_clip_fp16.onnx	170 MB	FP16	Mobile/Edge (recommended)
calorie_clip.onnx	354 MB	FP32	Desktop/Server

Quick Start

import onnxruntime as ort
import numpy as np
from PIL import Image

CLIP_MEAN = np.array([0.48145466, 0.4578275, 0.40821073], dtype=np.float32)
CLIP_STD = np.array([0.26862954, 0.26130258, 0.27577711], dtype=np.float32)

def preprocess(image_path):
    img = Image.open(image_path).convert('RGB').resize((224, 224))
    arr = np.array(img, dtype=np.float32) / 255.0
    arr = (arr - CLIP_MEAN) / CLIP_STD
    arr = np.transpose(arr, (2, 0, 1))
    return np.expand_dims(arr, axis=0)

session = ort.InferenceSession("calorie_clip_fp16.onnx")
image = preprocess("food.jpg")
calories = session.run(None, {"image": image})[0][0]
print(f"Calories: {calories:.0f}")

Scripts

• inference_onnx.py - Run inference on images
• export_onnx.py - Export PyTorch to ONNX
• quantize_onnx.py - Quantize to FP16 or INT8
• final_comparison.py - Compare model versions

Performance

Model	MAE (cal)	Size	Use Case
FP16	~13.6	170 MB	Mobile (recommended)
FP32	~13.7	354 MB	Desktop/Server
INT8	~20	86 MB	Maximum compression

• Inference time: ~35ms on CPU
• Input: 224x224 RGB image
• Output: single float (estimated calories)

References

• Base model: jc-builds/CalorieCLIP
• Trained on Nutrition5k dataset
• MAE: 51.4 cal on full test set (1951 samples)