I am confused about model quantization and inputs. Suppose I have an ONNX network that expects 256x256 inputs as 32bit floats. Then I convert this network to a quantized dlc format. What is the expected input format? Would the quantized model expect 256x256 8bit values, or does SNPE perform input quantization from 32bit floats to 8bit to run on AIP/DSP?
Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Qualcomm Incorporated or its subsidiaries (“Qualcomm”). The content is provided for informational purposes only and is not meant to be an endorsement or representation by Qualcomm or any other party. This site may also provide links or references to non-Qualcomm sites and resources. Qualcomm makes no representations, warranties, or other commitments whatsoever about any non-Qualcomm sites or third-party resources that may be referenced, accessible from, or linked to this site.
Hi,
Quantization means the network parameters of the model are made to 8-bit floating-point representation. So, this does not have any effect on the input provided to the model. But as seen from the documentation the snpe-onnx-to-dlc just converts the onnx model to a dlc model and needs to use snpe-dlc-quantize for quantizing the model. The input provided to the model needs to have a batch size as 1 (1,256,256,3 in your case if it is a color image).
For more details on NPE quantization keenly refer this link