A tool like onnx-modifier for openvino to change model constants?

[cansik]

I am currently working with a model that contains a NonMaxSuppression operator with fixed score and iou threshold. Does anyone know a tool like onnx-modifier to change openvino models and weights and adjust the constants?

[Ganesh-403]

@cansik From what I can tell, there isn't an OpenVINO equivalent of onnx-modifier that lets you interactively edit an IR model's graph and constants.

However, you do have a couple of options:

1. Modify the model before conversion (recommended)

   If you still have the ONNX model, using onnx-modifier (or another ONNX graph editing tool) to change the NonMaxSuppression thresholds before converting to OpenVINO is probably the simplest workflow.

2. Modify the OpenVINO graph programmatically

   OpenVINO exposes the model as a computational graph, and operations such as NonMaxSuppression take iou_threshold and score_threshold as input nodes rather than fixed attributes. You can construct new Constant nodes and replace the existing inputs or replace the entire NonMaxSuppression node using the graph transformation APIs (e.g. ov::replace_node). :contentReference[oaicite:0]{index=0}

For example, the graph rewrite would conceptually look like:

auto new_score = ov::op::v0::Constant::create(...);
auto new_iou   = ov::op::v0::Constant::create(...);

// Recreate NMS with the desired constants
// or replace the corresponding constant inputs.

I couldn't find a standalone GUI or CLI tool that performs this kind of IR editing automatically, so if your only artifact is the OpenVINO IR, the supported approach appears to be modifying the graph through the OpenVINO Runtime/Transformation API.

One question:

• Do you still have the original ONNX model, or only the generated .xml + .bin OpenVINO IR?

That determines whether it's easier to modify the graph before conversion or perform a graph rewrite on the IR itself.

If this solves your problem, feel free to mark it as the accepted answer so others can find it easily.

[dev-jinwoohong]

You can edit the IR directly by replacing the two NMS input edges. For a model with one NonMaxSuppression node and existing fixed thresholds:

import openvino as ov
from openvino import opset13 as ops

model = ov.Core().read_model("model.xml")
nodes = [n for n in model.get_ops()
         if n.get_type_name() == "NonMaxSuppression"]
if len(nodes) != 1:
    raise ValueError("Select the intended NMS node by its friendly name")
nms = nodes[0]
if nms.get_input_size() < 5:
    raise ValueError("This example requires existing threshold inputs")

for index, value in ((3, 0.5), (4, 0.25)):  # IoU, score
    port = nms.input(index)
    replacement = ops.constant(value, dtype=port.get_element_type())
    port.replace_source_output(replacement.output(0))

model.validate_nodes_and_infer_types()
ov.save_model(model, "edited.xml", compress_to_fp16=False)

The NMS input order makes these zero-based ports 3 and 4. Replacing the input edge leaves other consumers of a shared constant alone. Load/compile the edited model afterward.

Checked on a synthetic shared-constant NMS graph with OpenVINO 2026.4.1, including save/reload and CPU inference. Your model was not tested.