Compare pruner for various methods and precision

[lducas]

Include option to choose pruning method and precision (in fpylll ?).
Then, run a simple strategizer (for example with just lll preproc) with
various methods (gradient, ND, hybrid) and various precision.
Plot them and identify when things diverge.

@malb maybe you can handle the fpylll side, and I can (try) to deal with the strategizer side ?

[malb]

Sounds like a plan!

[lducas]

Poke me when ready ;)

[malb]

it seems we’re okay on the probability front. I added svp_probability to fpylll and ran:

from fpylll import *
from fpylll.fplll.pruner import svp_probability
strategies = load_strategies_json(BKZ.DEFAULT_STRATEGY)

set_precision(512)

for strategy in strategies:
    if strategy.block_size < 3:
        continue
    print strategy,    
    for pruning in strategy.pruning_parameters:
        probability = svp_probability(pruning, float_type="mpfr")
        if abs(probability - pruning.probability) > 0.00001:
            print "FAIL"
    else:
        print "PASS"

I got PASS for all strategies.

[malb]

The fpylll is now implemented, i.e. the prune function takes a precision and a descent_method parameter now.

[lducas]

Sorry for the stupid question:
it seems that the submodules are not linked to the latest version of fpylll. How should I update them ?

[malb]

Not stupid at all, here's how I do it.

cd fplll  && git checkout master && git pull && cd ..
cd fpylll && git checkout master && git pull && cd ..
git add fplll
git add fpylll
git commit -m "update to newest upstream releases"

[lducas]

I've tested hybrid vs gradient and no difference appears up to blocksize 75.
@malb would you mind running the comparison further using your X many cores ?

 python ./strategize.py --max-block-size 90 --threads X --samples 128 --method hybrid --prec 53 --filename hybrid53.json
 python ./strategize.py --max-block-size 90 --threads X --samples 128 --method gradient --prec 53 --filename gradient53.json

[malb]

Will do, it is currently under heavy use by others so it will be a few weeks until I can run it. Maybe @shi-bai or @dstehle have some spare idle cores?

[lducas]

I'm toying with progressive preprocessing strategies. Seems to start saving after dim 70. I'll launch bigger comparison when its more stable.

[shi-bai]

Hi Martin,

I could do that. Just tried, however, I could not find the option for "--method" in strategize.py?

strategize.py: error: unrecognized arguments: --method hybrid --prec 53

Cheers,
Shi

[malb]

Sorry @shi-bai , @lducas pulled his pull request: #4

[lducas]

I'm adding progressive strategies. If all goes well, I will push tonight.

Best
-- Leo

2016-08-29 14:42 GMT+02:00 Martin Albrecht notifications@github.com:

Sorry @shi-bai https://github.com/shi-bai , @lducas
https://github.com/lducas pulled his pull request: #4
#4

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#1 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AGBp4EQnwrwekQbDRLMIsByqj-CTxkI7ks5qktO7gaJpZM4JoXdT
.

[shi-bai]

Thanks. It works now. I started it yesterday and it is reaching blocksize 85 now.

Cheers,
Shi

[shi-bai]

Just finished the two runs with --max-block-size 90. Please see attached files for the json and logs. Github restricts the file types so I zipped them.

exp.zip