Continue to Site

Welcome to EDAboard.com

Welcome to our site! EDAboard.com is an international Electronics Discussion Forum focused on EDA software, circuits, schematics, books, theory, papers, asic, pld, 8051, DSP, Network, RF, Analog Design, PCB, Service Manuals... and a whole lot more! To participate you need to register. Registration is free. Click here to register now.

Register Log in

FPGA neural network training estimation time

Status
Not open for further replies.
A

adwnis123

Full Member level 4
Joined
Jun 19, 2014
Messages
214
Helped
0
Reputation
0
Reaction score
1
Trophy points
18
Activity points
1,737
Hello,

Is there a formula to estimate how much time does a neural network need for training? I am trying to build a CNN for image processing, 640x480 images on Virtex 5 (without success however...). I have read articles about that, but I haven't seen anywhere talking about training time.

Thank you...
 

there is no formula.

training is much harder than inference, and typically is not performed online. to give you a rough reference, it took me 2-3 days of CPU time to train a 400 neuron spiking neural network with STPD. the input set consisted of the MNIST database of 27x27 pixel images that represent handwritten digits. The time to train will change based on input size and characteristics, target accuracy, network architecture, and so on. too many parameters to generalise in one equation.
 
ThisIsNotSam said:
there is no formula.

training is much harder than inference, and typically is not performed online. to give you a rough reference, it took me 2-3 days of CPU time to train a 400 neuron spiking neural network with STPD. the input set consisted of the MNIST database of 27x27 pixel images that represent handwritten digits. The time to train will change based on input size and characteristics, target accuracy, network architecture, and so on. too many parameters to generalise in one equation.
Click to expand...

So many neurons for 27x27 pixels?
 

Yes, but then again my network is far from traditional.
 

I am trying to build a CNN for image processing, 640x480 images
Click to expand...

Seems like you are trying to use the full image captured for that NN, but in general the processing is made with a smaler picture, the ROC (region of interest) and/or with a reduced scale of the original image so that the training time can be substantially reduced.
 
andre_teprom said:
Seems like you are trying to use the full image captured for that NN, but in general the processing is made with a smaler picture, the ROC (region of interest) and/or with a reduced scale of the original image so that the training time can be substantially reduced.
Click to expand...

This sounds better, rather than using full 640x480 image, but how I "center" to the ROC?
 

but how I "center" to the ROC
Click to expand...

It is applicable when the pattern to be detected lies within a known region; or depending on the application, if the recognition can be based on a statistical rather than a deterministic approach, you could alternatively consider the most probable region where the object is found in most samples, and reduce the ROC within that bounds.
 
Status
Not open for further replies.

Similar threads

Part and Inventory Search

Welcome to EDABoard.com

Sponsor

Back
Top