Ternary weights not reflected in caffe model about stochastic-quantization HOT 3 CLOSED

Yes, the weights in caffemodel files are still full-precision. We actually save the full-precision weights instead of low-bits weights.
The accuracy results in the first stage are sometimes useless. Only the models after the forth(final) stage can obtain good results.
If you look at our codes, you will find that we first quantize the full-precision weights to low-bit weights and then perform forward propagation, even at the testing stage. So if you want to get the quantized weights for further usage, you probably need to quantize the weights and save them manually.

from stochastic-quantization.

Thanks for the clarification, I was able to obtain the quantized weights.

from stochastic-quantization.

Hi, @sunejas
How can you obtain the quantized weights, can you provide the script?
Thank you.

from stochastic-quantization.