Hello,
We want to port our weights/biases to TIDL model with our own quantization parameters, and ensure same result and value as PC since our algorithm is very sensitive to value.
So questions are about how I can use our quantization parameters, with Caffe data.
1. What quantization parameters do TIDL use? (e.g fractional length of weight)
1. How can we encode our parameters into TIDL model?
2. Can you let us know how TIDL operations(e.g convolutions, batchnorm) use quantization parameters?
Thank you.
