Ksnn convert parameters documentation

The khadas doc mentions the meaning of certain parameters but the significance of the below mentioned arguments are still not clear

  --batch-size BATCH_SIZE
  --iterations ITERATIONS
  --device DEVICE
  --hybrid HYBRID
  1. What do batch_size and iterations denote? And how can they be initialised?
  2. What does device argument mean? Does it mention the device (CPU/GPU) where the conversion will run or the device in the khadas board where the inference will run
    I’m getting the following error on using either of the two options

  1. What does hybrid quantisation mean? Are we allowed to select the qtypes to be considered during conversion or does the script select the best performing ones (based on the board’s architecture)?

@johndoe These parameters are used when mixing, but now KSNN does not support mixing, so you can ignore these parameters. No matter which platform you use, you don’t need to consider these parameters for the time being.

1 Like

@Frank On a similar note, would you recommend using the C/C++ scripts (NPU SDK) to convert rather than relying on ksnn?

@johndoe If you do not plan to use python but use C/C++, then you should directly use the NPU SDK

@Frank I’m benchmarking variety of networks on different quantisation techniques. Will NPU SDK allow more choices?

@johndoe No, the quantitative options for the two are the same. Unless you use mixed quantization, but I do not recommend using mixed quantization, the data will appear abnormal jitter

1 Like