Which system do you use? Android, Ubuntu, OOWOW or others?
Ubuntu 22.04 on VIM3
Which version of system do you use? Please provide the version of the system here:
VIM3 (Khadas VIM3, not VIM3L)
Ubuntu 22.04.3 LTS
KSNN Convert tools v1.4
aml_npu_sdk-ddk-6.4.8.7-250427
Please describe your issue below:
I’m trying to convert a custom YOLOv8 crowd detection model to .nb format for VIM3 NPU inference.
Steps taken:
- Exported PyTorch model (best.pt) to ONNX format successfully (best.onnx - 12.1 MB)
- Prepared calibration dataset (100 crowd detection images, 468.75 MB)
- Attempted conversion using KSNN converter with command:
./convert --model-name human --platform onnx --model ./human.onnx --input-size-list “3,640,640” --inputs images --outputs output0 --mean-values “127.5 127.5 127.5 0.007843137” --quantized-dtype asymmetric_affine --source-files ./data/dataset/dataset0.txt --kboard VIM3 --print-level 1
The ONNX model exports successfully, but the converter fails during shape inference with an invalid dimension (1,0,160,160).
Can you help with:
- Why this error occurs?
- Is there a workaround or newer converter version?
- Alternative conversion method for custom YOLOv8 models?
Post a console log of your issue below:
--+ KSNN Convert tools v1.4 +--
Start import model ...
I Namespace(import='onnx', input_dtype_list=None, input_size_list='3,640,640', inputs='images', model='./human.onnx', output_data='Model.data', output_model='Model.json', outputs='output0', size_with_batch=None, which='import')
I Start importing onnx...
I Current ONNX Model use ir_version 9 opset_version 20
[24] Failed to execute script pegasus
Traceback (most recent call last):
File "pegasus.py", line 131, in <module>
File "acuitylib/converter/onnx/convert_onnx.py", line 761, in _shape_inference
File "acuitylib/onnx_ir/onnx_numpy_backend/smart_toolkit.py", line 636, in multi_direction_broadcast_shape
ValueError: operands could not be broadcast together with shapes (1,0,160,160) (1,16,160,160)