Re: [help-GIFT] patch-fu

[David Squire]

address@hidden wrote:
> Jonas:
> Yes, not doing a calloc() every time through improves performance by arround 5%.
>
> David,
>
> it is my intention to expand the image size handling of the feature extractor 
> as soon as i'm done with performance hacks, however, the curent code wont 
> properly process an image larger than 256x256. therefore, i think allocating
> the array automatically, to the MAX_WIDTH and MAX_HEIGHT is appropriate.
>   
OK.... but I am concerned about a direction of modification that seeks a 
few percentage points of speed improvement here and there at the expense 
of good design (i.e. low coupling and extensibility) - not that I am 
suggesting that the feature extraction code as it was is an example of 
good design. We should be aiming always for good design.

The above approach tightly couples parts of the code to others via an 
essentially arbitrary choice of image size, that *happens* to be 
constant in the current version, but need not be. I would be *much* 
happier seeing things dynamic and parameterized, even if that makes 
things a little slower.

Remember the "rules" of optimization:

1. Don't optimize by hand.
2. If you think you need to optimize by hand, think again.
3. If you still really think you need to optimize by hand, optimize late.

This is particularly true of research software (such as the GIFT 
essentially is), where the requirements are moving targets. For example, 
I have my own versions of the feature extraction code that also look for 
and index text files associated with the images. The feature extraction 
code is intended to be as separate as possible from the indexing and 
query engine, so that others can write and use their own feature 
extraction code, using whatever features they like.

Also, the typical IR scenario is that you extract features and index 
once, and query many times. Consequently, optimizing query performance 
is much more important than optimizing feature extraction. Users are 
often quite happy for indexing to take hours or days.

I am not saying that optimization of the feature code is not 
appreciated, just that the trade-offs need to be kept closely in mind. 
IMHO, extending to arbitrary image sizes and shapes first would make 
more sense.

Regards,

David


--
Dr David McG. Squire, Senior Lecturer, on sabbatical in 2006
Caulfield School of Information Technology, Monash University, Australia
CRICOS Provider No. 00008C        http://www.csse.monash.edu.au/~davids/