Get shape for each output from the ONNX and infer layout + YOLO26 changes + output field re-names

[dtronmans]

Purpose

Shape and layout:

• Re-load the saved ONNX and use it to retrieve the output shapes
• For the output layout I found no other way than to infer it from the dim length, at some points it might be incorrect for example if NCD is supposed to be NDC instead but I don't see anywhere where the layout field is used, for example in depthai-nodes, to decide how to post-process the output

Renaming output names:

• YoloV8 family: previous keypoint_outputs was previously incorrect, should be ["kpt_output1", "kpt_output2", "kpt_output3"],
• YOLO26 outputs: output -> output_yolo26, mask_output -> output_masks
• Add tests for correct yolo_output and output names for YOLO26 and all exporters that use the YOLOV8 exporter
• Corresponding change in depthai-nodes (use updated field names): Remove need for cls_scores = scores.max(axis=-1) as ReduceMax resides on the model side for YOLO26 depthai-nodes#284

Yolo26 conversion with one2one head:

• tools yolo26n.pt --imgsz "640" --version yolov26_nms
• Since this is the first time the same .pt file can have two different export branches, I had to add cli_version flag to the tests

ReduceMax in graph:

• Not an option/flag it's just always there
• Corresponding change in depthai-nodes (remove objectness calculation there: Remove need for cls_scores = scores.max(axis=-1) as ReduceMax resides on the model side for YOLO26 depthai-nodes#284)

Specification

Dependencies & Potential Impact

Deployment Plan

Testing & Validation

yolov8n:

"outputs": [
    {
        "name": "output1_yolov6r2",
        "dtype": "float32",
        "shape": [
            1,
            85,
            80,
            80
        ],
        "layout": "NCHW"
    },
    {
        "name": "output2_yolov6r2",
        "dtype": "float32",
        "shape": [
            1,
            85,
            40,
            40
        ],
        "layout": "NCHW"
    },
    {
        "name": "output3_yolov6r2",
        "dtype": "float32",
        "shape": [
            1,
            85,
            20,
            20
        ],
        "layout": "NCHW"
    }
],

yolo26n:

"outputs": [
    {
        "name": "output",
        "dtype": "float32",
        "shape": [
            1,
            8400,
            84
        ],
        "layout": "NCD"
    }
],

yolo26n-seg:

"outputs": [
    {
        "name": "output",
        "dtype": "float32",
        "shape": [
            1,
            8400,
            84
        ],
        "layout": "NCD"
    },
    {
        "name": "mask_output",
        "dtype": "float32",
        "shape": [
            1,
            8400,
            32
        ],
        "layout": "NCD"
    },
    {
        "name": "protos_output",
        "dtype": "float32",
        "shape": [
            1,
            32,
            160,
            160
        ],
        "layout": "NCHW"
    }
],

yolo26n-pose:

"outputs": [
    {
        "name": "output",
        "dtype": "float32",
        "shape": [
            1,
            8400,
            5
        ],
        "layout": "NCD"
    },
    {
        "name": "kpt_output",
        "dtype": "float32",
        "shape": [
            1,
            8400,
            51
        ],
        "layout": "NCD"
    }
],

[github-actions]

☂️ Python Coverage

current status: ✅

Overall Coverage

Lines	Covered	Coverage	Threshold	Status
30404	6515	21%	0%	🟢

New Files

No new covered files...

Modified Files

File	Coverage	Status
tools/main.py	86%	🟢
tools/modules/exporter.py	84%	🟢
tools/modules/heads.py	51%	🟢
tools/version_detection/version_detection.py	47%	🟢
tools/yolo/yolo26_exporter.py	0%	🟢
tools/yolo/yolov8_exporter.py	63%	🟢
TOTAL	55%	🟢

updated for commit: 1ef0413 by action🐍

[github-actions]

Test Results

  6 files    6 suites   24m 4s ⏱️
 26 tests  26 ✅ 0 💤 0 ❌
156 runs  156 ✅ 0 💤 0 ❌

Results for commit 1ef0413.

♻️ This comment has been updated with latest results.

[aljazkonec1]

I've been testing pose model and the keypoint layer output shape differs compared to say Yolo v8 pose:

                "layout": "NCD",
                "name": "kpt_output1",
                "shape": [
                    1,
                    51,
                    3520
                ]
            },

Here you see C dimension is 51, having the order be:

        "name": "kpt_output",
        "dtype": "float32",
        "shape": [
            1,
            8400,
            51
        ],
        "layout": "NCD"
    }

breaks the current implementation of DetectionParser as it assumes data is flattened as [1, 51, 3520]. Which is the "correct" shape? Are both valid shapes?

[aljazkonec1]

I have added depthai-core support for YOLO26 here. I tested the PR against the nnarchives @dtronmans generated with this PR and all looks good.

The only thing I would change in the NNarchives is the parser is listed as "YOLOExtendedParser" and I would change it to "YOLO" so ParsingNeuralNetwork automatically switches to depthai-core DetectionParser?

[dtronmans]

I have added depthai-core support for YOLO26 here. I tested the PR against the nnarchives @dtronmans generated with this PR and all looks good.

The only thing I would change in the NNarchives is the parser is listed as "YOLOExtendedParser" and I would change it to "YOLO" so ParsingNeuralNetwork automatically switches to depthai-core DetectionParser?

@aljazkonec1 good point, done