Reproducible recipe for Swin-Tiny ONNX

General Discussion
, , ,
1

Dear Hailo Community / Support Team,

I’m trying to compile the Swin-based classifier from Hugging Face as referenced in the Hailo Model Zoo config.

To reproduce the error see the attached zip https://artifacts.iris-sensing.com/web/client/pubshares/Ktht5AhekvsTBxcUAwsyUn/download:
$ python3 run_hf.py
$ hailo parser onnx swin-tiny-patch4-window7-224_single.onnx --hw-arch hailo8

I downloaded the source model you refer to, but the ONNX I export differs from the ONNX provided by Hailo (see attached onnx_diff.png). Because of this mismatch, the compilation fails (see attached error.log).

Could you please share the exact recipe you used to generate the “compilable” ONNX for this network—specifically:

  • the Hugging Face model identifier / exact revision (commit/tag)

  • preprocessing / input assumptions (size, normalization, layout)

  • export steps (PyTorch → ONNX), including opset, dynamic axes settings, and any graph simplifications

  • any patches or custom modifications applied prior to export

  • the exact command(s) used to compile the ONNX into a .hef

If there’s a script or a reference export notebook used for the Model Zoo ONNX, sharing that would also solve the issue.

Thanks in advance,
Ruben

Solider ReID (Swin - based)
2

We replicated the issue and asked our R&D colleagues to have a look into it.

3

Hi @Ruben_Khachaturyan,

Please try the below, works for me on DFC 3.33.0. The trick was using opset=15 and simplifying the ONNX:

from PIL import Image
import requests
import torch
import onnx
from onnx import external_data_helper
from onnxsim import simplify
from transformers import AutoImageProcessor, AutoModelForImageClassification

opset_version = 15
onnx_path = f"swin-tiny-patch4-window7-224_{opset_version}.onnx"

# ====== LOAD MODEL
model = AutoModelForImageClassification.from_pretrained("microsoft/swin-tiny-patch4-window7-224")
model.eval()
# ====== EXPORT TO ONNX
dummy = torch.randn(1, 3, 224, 224)
torch.onnx.export(
    model, dummy, onnx_path,
    export_params=True,
    training=torch.onnx.TrainingMode.PRESERVE,
    input_names=["input"],
    output_names=["output"],
    opset_version=opset_version,
    do_constant_folding=False,
    dynamo=False
)

# ====== SIMPLIFY ONNX
onnx_model = onnx.load(onnx_path)

# Provide input shape hints so onnxsim can reason well about shapes
simplified_model, check = simplify(
    onnx_model,
    input_shapes={"input": [1, 3, 224, 224]}
)

if not check:
    raise RuntimeError("ONNX simplification check failed")

onnx.save(simplified_model, onnx_path)

# ====== (OPTIONAL) Convert external data -> embedded
# Only needed if your export produced external tensor files and you want everything in one .onnx
m = onnx.load(onnx_path, load_external_data=True)
external_data_helper.convert_model_from_external_data(m)
onnx.save(m, onnx_path)


print(f"✅ Exported + simplified ONNX saved to: {onnx_path}")
1 Like
4

Dear Nadav, it totally worked! Thank you very much!

1 Like