Simulation Frequency

[aprajita]

The simulation frequency seems to go down v rapidly. can you remind me of the the frequency steps.

[drphilmarshall]

It was half as much as it should be, now fixed. Here's the code, in classifier.coffee:

nClassified = @counter.classified
baseLevel = Math.floor(nClassified / 20) + 1
@level = Math.min(baseLevel, 3)
denominator = (5 * Math.pow(Math.sqrt(2), @level - 1))
@simRatio = 2 / denominator
Subject.group = if @simRatio > Math.random() then @simulationGroup else

@subjectGroup

So 2 in 5 for the first 20, then 2 in 7, 2 in 10, 2 in 14 etc in successive groups of 20. Should we make that 20 something longer? Like 40, perhaps? To keep the bacon.training up?

On Mon, Jan 6, 2014 at 5:33 PM, aprajita notifications@github.com wrote:

The simulation frequency seems to go down v rapidly. can you remind me of the the frequency steps.

— Reply to this email directly or view it on GitHubhttps://github.com/drphilmarshall/Lens-Zoo/issues/59 .

[drphilmarshall]

Set to 30 for now.

[aprajita]

30 sounds ok thanks

On 6 Jan 2014, at 23:12, Phil Marshall wrote:

It was half as much as it should be, now fixed. Here's the code, in classifier.coffee:

nClassified = @counter.classified baseLevel = Math.floor(nClassified / 20) + 1 @level = Math.min(baseLevel, 3) denominator = (5 * Math.pow(Math.sqrt(2), @level - 1)) @simRatio = 2 / denominator Subject.group = if @simRatio > Math.random() then @simulationGroup else @subjectGroup

So 2 in 5 for the first 20, then 2 in 7, 2 in 10, 2 in 14 etc in successive groups of 20. Should we make that 20 something longer? Like 40, perhaps? To keep the bacon.training up?

On Mon, Jan 6, 2014 at 5:33 PM, aprajita notifications@github.com wrote:

The simulation frequency seems to go down v rapidly. can you remind me of the the frequency steps.

— Reply to this email directly or view it on GitHubhttps://github.com/drphilmarshall/Lens-Zoo/issues/59 .

— Reply to this email directly or view it on GitHub.

[aprajita]

@drphilmarshall there seems to be something odd with dud frequency as I discussed in another issue I just did a test in 100 images I get 5 sims - this is a 1/20 sim frq so makes sense as I've looked at 2300 images. but in the same 100 images I got 15 Duds. I think this is too many and they are often bunched up together. Can u please check the dud frequency? as i mentioned in #61, of the 5 sims in this test even though I marked the sim I didnt get any feedback.

[drphilmarshall]

This is odd - the code is the same in the classifier. Michael @parrish, would you mind checking that the number of duds is correct please? There should be only 196, the same as the sims. Can you think of a reason why they should be over-represented in the training group?

The only thing I can think of is that Aprajita and I have spent more time looking at sims than duds, and so that Aprajita's personal pool of sims is quite small.

On Tue, Jan 7, 2014 at 7:13 AM, aprajita notifications@github.com wrote:

@drphilmarshall https://github.com/drphilmarshall there seems to be something odd with dud frequency as I discussed in another issue I just did a test in 100 images I get 5 sims - this is a 1/20 sim frq so makes sense as I've looked at 2300 images. but in the same 100 images I got 15 Duds. I think this is too many and they are often bunched up together. Can u please check the dud frequency? as i mentioned in #61https://github.com/drphilmarshall/Lens-Zoo/issues/61, of the 5 sims in this test even though I marked the sim I didnt get any feedback.

— Reply to this email directly or view it on GitHubhttps://github.com/drphilmarshall/Lens-Zoo/issues/59#issuecomment-31732629 .

[aprajita]

"The only thing I can think of is that Aprajita and I have spent more time looking at sims than duds, and so that Aprajita's personal pool of sims is quite small."

no that's not it, as i noticed lots of duds even when i had only classified a small number of galaxies