[tf2]How to generate the binary lookup table of approximate multiplier? about tf-approximate HOT 7 CLOSED

Hi,
it is very easy in C. But you can also try to do it in different tool (the accurate 8-bit multiplier output should be exactly same as tf2/examples/axmul_8x8/mul8u_1JFF.bin

FILE * f = fopen("output.bin", "wb");

for(unsigned int a = 0; a < 256; a++)
    for(unsigned int b = 0; b < 256; b++) {
      uint16_t val = approximate_mult(a, b); // replace by your own function call
      fwrite(&val, sizeof(uint16_t), 1, f);
    }

fclose(f);

from tf-approximate.

I have included this example to the README file.

from tf-approximate.

Greetings,

Can we employ approximate multipliers with different bit-widths?

from tf-approximate.

The inputs of convolutional layers are quantized for 8 bits. If you want to simulate lower bit widths, you can simulate it as:

int approximate_mult(int a, int b) {
   return (a & 0xFE) * (b & 0xFE); // for 7-bits
}

from tf-approximate.

Is it possible to quantize weights using higher bit-widths. In the code, I saw that you can adjust the the quantization parameter.

from tf-approximate.

We have tried to keep the interface somewhat open to that possibility. However at the moment this is not supported as multiplication lookup tables would get too large very quickly and wouldn't fit into fast caches of the GPU.

from tf-approximate.

Thanks for response and thank you for publishing the code.

from tf-approximate.