JElvis/upernet-convnext-tiny-ade-onnx

UperNet + ConvNeXt-Tiny (ADE20K) — ONNX

An ONNX export of openmmlab/upernet-convnext-tiny, unmodified weights, fp32. Released under the same MIT licence as the source model.

Input

Name	Shape	Type
pixel_values	[1, 3, 512, 512]	float32

RGB, resized to 512×512 (aspect ignored), scaled to 0..1, then normalised with ImageNet mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225], CHW. This matches the source model's preprocessor_config.json.

Output

Name	Shape	Type
logits	[1, 150, 128, 128]	float32

The decode head's logits at 1/4 of the input resolution. The source model's final bilinear upsample to the input size is not in the graph: upsample the logits (bilinear, align_corners=False) to the size you need, then argmax.

Classes: the 150 ADE20K classes, 0-based, no background class (0 = wall, 1 = building, 2 = sky, …), identical to the source model's id2label.

Export

export_upernet.py in this repo. torch 2.14.1, transformers 5.18.0, onnx 1.23.1, torch.onnx.export(..., opset_version=18, dynamo=True). Checked against PyTorch on random input: max abs logit difference 2.7e-5, 100% argmax agreement.

Citation

• Xiao et al., Unified Perceptual Parsing for Scene Understanding, ECCV 2018.
• Liu et al., A ConvNet for the 2020s, CVPR 2022.