ZumaInnovations/C-RADIOv4-H-ONNX

C-RADIOv4-H ONNX

ONNX conversion of nvidia/C-RADIOv4-H for spatial patch features. This repository is not published by NVIDIA.

The weights are licensed by NVIDIA Corporation under the NVIDIA Open Model License (June 14, 2024). A copy of that agreement is in nvidia-open-model-license-agreement-june-2024.pdf. See Notice.

Graph

Files	model.onnx + model.onnx_data
Input	pixel_values, float32, [B, 3, H, W], RGB in [0, 1]
Output	last_hidden_state, float32, [B, N, 1280]
Patch size	16
Tokens	N = (H / 16) * (W / 16)
Traced at	512x512, with dynamic height and width
Opset	17

H and W must be multiples of 16, up to the source model's 2048 limit. C-RADIO normalizes internally, so do not apply ImageNet mean and standard deviation. The export returns spatial tokens only. Summary tokens are not in the output. ViTDet windowed attention is off.

Converted from c-radio_v4-h_half.pth.tar with NVlabs/RADIO, then cast to float32 before export.