MobileViT-XXS — 3D Print Failure Detector

Binary image classifier: normal vs failure (spaghetti / bed detachment). Designed for on-device inference on Raspberry Pi 4 via ONNX Runtime (~200–400 ms/frame).

Labels

ONNX Inference (Raspberry Pi)

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

processor = MobileViTImageProcessor.from_pretrained("Masamsa/3d-print-failure-mobilevit-xxs")
sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])

img = Image.open("frame.jpg").convert("RGB")
inputs = processor(images=img, return_tensors="np")
logits = sess.run(["logits"], dict(inputs))[0]
pred   = int(np.argmax(logits, axis=-1)[0])
label  = {0: "normal", 1: "failure"}[pred]
print(label)

Training Details

• Base: apple/mobilevit-xx-small (1.3M params)
• Image size: 256×256
• Best val macro-F1: 0.9011
• Near-balanced dataset (normal≈3.7K / failure≈3.3K) — standard cross-entropy loss