VIM4 ONNX to KSNN model convert fail with multiple outputs

Which system do you use? Android, Ubuntu, OOWOW or others?

Ubuntu

Which version of system do you use? Please provide the version of the system here:

VIM4 with NPU kernel 5.15 241129

Please describe your issue below:

Running model convert in PC fails. System with WSL and Docker Desktop running with “numbqq/npu-vim4” container.

When using vim4_npu_sdk and running command “./convert-in-docker.sh ksnn” for python model cannot convert onnx model with multiple outputs.

Model converts fine if I specify only output0. I can share onnx file here also.

Post a console log of your issue below:

usage: convert [-h] [--model-name MODEL_NAME] [--print-level PRINT_LEVEL] [--model-type MODEL_TYPE] [--model MODEL] [--weights WEIGHTS] [--inputs INPUTS]
               [--input-shapes INPUT_SHAPES] [--shape-with-batch SHAPE_WITH_BATCH] [--dtypes DTYPES] [--outputs OUTPUTS] [--batch-size BATCH_SIZE]
               [--iterations ITERATIONS] [--outdir OUTDIR] [--quantize-dtype QUANTIZE_DTYPE] [--source-file SOURCE_FILE]
               [--channel-mean-value CHANNEL_MEAN_VALUE] [--disable-per-channel DISABLE_PER_CHANNEL] [--inference-input-type INFERENCE_INPUT_TYPE]
               [--inference-output-type INFERENCE_OUTPUT_TYPE] [--kboard KBOARD] [--quantize-algo QUANTIZE_ALGO] [--optimize-algo OPTIMIZE_ALGO]
               [--kl-threshold-size KL_THRESHOLD_SIZE] [--divergence-nbins DIVERGENCE_NBINS] [--percentile PERCENTILE]
               [--asymm-percentile ASYMM_PERCENTILE] [--moving-alpha MOVING_ALPHA] [--neg-scale NEG_SCALE] [--weights-quantize-algo WEIGHTS_QUANTIZE_ALGO]
               [--weights-optimize-algo WEIGHTS_OPTIMIZE_ALGO] [--weights-threshold-size WEIGHTS_THRESHOLD_SIZE] [--weights-neg-scale WEIGHTS_NEG_SCALE]
               [--export-template-code EXPORT_TEMPLATE_CODE] [--system-env SYSTEM_ENV] [--debugger-option DEBUGGER_OPTION]
convert: error: unrecognized arguments: output1 output2

Following is ksnn_args.txt file used
–model-name modified_yolov8n-base-version
–model-type onnx
–model modified_yolov8n-base-version.onnx
–inputs “images”
–input-shapes “3,640,640”
–dtypes “float32”
–outputs “output0 output1 output2”
–quantize-dtype int8
–outdir onnx_output
–channel-mean-value “0,0,0,255”
–source-file dataset.txt
–iterations 1
–batch-size 1
–kboard VIM4
–inference-input-type “float32”
–inference-output-type “float32”