PTQ For Per-Tensor Quantization

[kossyrev-bg]

I'm trying to upload a custom model onto an ARTPEC-8 Camera (M4328-P). This model is a tensorflow model that we're trying to convert to tflite. Since the DLPU is optimized for per-tensor, we added the following line to the converter:

converter._experimental_disable_per_channel = True

However, after converting, the model outputs nonsense data for every input image. The output is always

  ([0.9928777813911438, 0.007122197654098272],

no matter what input image we supply. This conversion works correctly when we do per-channel quantization; the model outputs correct predictions.

I'm not really sure where to go from here, if there are other flags or methods to try.