Accuracy Problem

[wonderzy]

In your paper, you say '.In our architecture, inputs and feature maps are in the dynamic fixed-point format'.
Do you use this in your ImageNet result or just use weights quantization and float input? From your python script, it seems that you just use approximate float weights to produce the shift loss. @gudovskiy

[gudovskiy]

@wonderzy good question. Yes, in the script activations are, obviously, floating-point. To get results in the Tables, activations were quantized using Ristretto

[wonderzy]

Hi, I tried the example Squeeze Net. Sadly I found that the accuracy of only quantizing activations of Squeeze Net v1.0 drops 0.006. (0.57058 vs 0.5768) and Squeeze Net v1.1 drops 0.004 (0.5839 vs 0.5793). But in your paper, you got the accuracy of only 0.01% drop. It looks like you just put the approximate float weights into caffe and get the accuracy or do something else? This troubles me a lot, any suggestions? @gudovskiy @gudovskiy

[wonderzy]

Another question, in your script you divide each value by a max value, how did you deal with this at inference? @gudovskiy

[gudovskiy]

@wonderzy Did you finetune after quantizing activations? The 2nd question is answered here.

[lawrencewxj]

Hi, I want to know when you reimplement it on the computer. Is the Squeeze Net significantly accelerated? I am curious about the speed-up. Thank you very much. @wonderzy . And speed-up results on PC are not reflected in the ImageNet result table. Can you tell me a rough result ? Thank you very much. @gudovskiy

[gudovskiy]

@lawrencewxj it is not accelerated since it is emulation. For acceleration you need specialized hardware.

[lawrencewxj]

OK, I get it, thank you anyway.