Fixed-point model numbers

[laski007]

Dear Prof. Wang,
For the quantized parameter, you used (n,m) pair to denote the precision. For example, in VGG-16 1st layer, you used (8,7) (8, 0) (8,-2) to denote frac_w, frac_input and frac_output, for last FC layer, you used (8, 2) (8,2) (4,7). Is there any rules or constrains how to decide the numbers? Or just use any numbers you like? Can I change to other numbers? If I changed the fraction numbers and convert a new model, does it works or not? Thank you so much.

[laski007]

Okay, I find
rule #1, previous layer's output fraction number must equal to next layer's input fraction number.
rule #2, 1st layer's input fraction number is 0.
Anything else? Why in AlexNet the 1st layer's output fraction is (8, -4) but in VGG16 the output is (8, -2)? For the weight fraction, sometime it is (8,7), sometime it is (8,8) or (4,7) or other numbers such as (8,9) (8,10) (8,11), why?
Because I have trained my own model with different number of Convolution layers and FC layers but I don't how to set the weight fraction numbers to convert my own model.
Many thanks.

[laski007]

Dear Prof. Wang,
Could you please help me to briefly describe how to decide the frac_w, frac_input and frac_output numbers when you have spare time? I'm still waiting ... Thank you so much. 跪谢Orz

[aazz44ss]

frac_w, frac_input and frac_output depends on the distribution of your data.
If your weight is range from -32.0 to 32.0, like -19.83, 2.15, -5.14, 6.81, 21.02, 32.11.....
you should choose frac_w of 2 (1 signed bit, 5 integer bits, 2 fractional bit)
you have to use 5 integer bits to hold your range of weight. ( 2^5 = 32 )
every number larger than 32 will become 32. (you can see conv.cl code)
and if your weight is range only from -1 to 1, you don't need so many integer bits,
so you can increase your fractional bit to increase precision.

fran_input and frac_output done by the same way.
you have to try fractional bit with different layers, and see the accurate with your test data set.

the description of how to choose fractional bit is in the latest PipeCNN paper just in the document folder.

[doonny]

Hi, all, please refer to matlab's fixed-point number's defination in the Fixed-point Toolbox. We are using a very similar data format.