AojunZhou

@KangolHsu 0.007812 ？ Maybe you truncate the 0.0078125 (2^-7).

yes, for 4 bits, you can use 15 no-zero values + 0, unbalanced quantization doesn't influence INQ result, specifically, the INQ is very flexible，you can adjust bit-width according your task,...

@mynameischaos yes, network quantization can accelerate inference speed, but your hardware must support low-precision bit shift， you can pay close attention to intel-altera and intel-movidious product.

@lamperouge11 for ImageNet, most convNet models (ResNet, GoogleNet, AlexNet, VGG) weight range are within -1,1.

@TwistedfateKing line 235 at /src/caffe/solvers/sgd_solver.cpp. if only one weight is float, I fixed "std::abs(data_vec[i])>data_copy[partition]" (line 536) at src/caffe/blob.cpp

@KangolHsu Sorry, I don't know, INQ hardware implementation was completed by another team.

@TwistedfateKing yes, the number 7 (default) is corresponding to 5 bits in paper, you can modify it, 3 for 4 bits, 1 for 3 bits, 0 for 2 bits.

n2 = n1 + 1 −2^(b−1)/2. For instance, if b = 3 and n1 = −1, it is easy to get n2 = −2, if b=5, n2=-1+1-(2^(5-1))/2=-8

@KangolHsu

you can read the code from ./src/caffe/blob.cpp in line 480, and I am writing the readme and tutorial.