estock-dinov2-vitb14-onnx

DINOv2 ViT-B/14 Slim FP16 ONNX

This repository contains a deployment-oriented ONNX export of the standard dinov2_vitb14 backbone from facebookresearch/dinov2.

Artifact

The graph was optimized with onnxslim and is intended for single-image inference with ONNX Runtime DirectML, with CPU execution as a fallback.

Inputs

The exported graph has two inputs:

• pixel_values: FP16 tensor with shape [1, 3, 224, 224].
• masks: scalar boolean set to false.

Benchmark preprocessing:

1. Convert the image to RGB.
2. Resize with bicubic interpolation so the shortest side is 224.
3. Take a centered 224 x 224 crop.
4. Convert to CHW and scale pixels to 0 through 1.
5. Normalize with ImageNet mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225].
6. Cast pixel_values to FP16.

The EStock catalog uses a compatibility profile that first resizes the shortest side to 256 and then takes the same 224 x 224 center crop. The ONNX graph itself is identical for both profiles.

For similarity search, cast the output to FP32 and L2-normalize it before storage or comparison.

Runtime

Tested on Windows with:

• Python 3.14.5, 64-bit
• onnxruntime-directml==1.24.4
• numpy==2.5.2
• pillow==12.3.0

Provider preference:

1. DmlExecutionProvider
2. CPUExecutionProvider

Verification

Verify the downloaded artifact before loading it:

(Get-FileHash .\dinov2_vitb14_opset18_slim_fp16.onnx -Algorithm SHA256).Hash.ToLowerInvariant()

Expected value:

d0737e9262093604e0b847e4388dfaca6e21b6c35663a81bc6f7f1085a56ffba

License and attribution

The source DINOv2 code and standard S/B/L/g model weights are distributed by Meta under the Apache License 2.0. This optimized export retains that license; see the upstream repository and model card for attribution and usage details.