cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 
trived76
Observer
Observer
269 Views
Registered: ‎10-03-2019

Which quantization method - 0 for non-overflow and 1 for min-diffs - to use for which network?

Hi Xilinx Community,

I saw from vai_q_caffe command usage that there are two quantization methods: 0 for non-overflow and 1 for min-diffs. To the best of my knowledge, the old Edge-AI tutorial for SSD and Yolo both used method 0, whereas the new Vitis-AI tutorial uses method 1 for YoloV4.

Is there any general rule for using method 0? I meant, for which types of network we need to use method 1 and method 0?

Thanks very much.

0 Kudos
2 Replies
jheaton
Xilinx Employee
Xilinx Employee
194 Views
Registered: ‎03-21-2008

I have seen that method 1 generally gives better accuracy results for yolo.

I usually try both options, and the use the one that gives the best results.

0 Kudos
mirresnick
Newbie
Newbie
149 Views
Registered: ‎04-09-2021

Many businesses, even the largest and most successful ones, invest huge amounts of money on ineffective or under-effective telemarketing efforts, especially when there is little or no direct competition https://tubreveespacio.com/foros/viewtopic.php?t=13463&sid=93329620dd5feb3ad888c7001b84b2d3 . This strategy tends to drain a company's finances, as it consumes large amounts of time and money to follow up with each lead, many of which are not likely to convert to customers. IVR technologies such as Salespro lead capture software allows businesses to bypass this problem, as they will immediately begin making contact with leads on their own, as soon as the technology is implemented. Businesses can create and maintain effective and productive follow-up lines, which greatly reduces the amount of money wasted on ineffective telemarketing techniques.

0 Kudos