Rugby Analytics commit #764

springcoil · 2015-06-17T21:12:50Z

Second rugby analytics commit

jsalvatier · 2015-06-17T21:40:05Z

Cool! Thanks :)

A few comments.

You don't need the first model = pm3.Model() since you already use the with statement.
I would fit a slightly different model. I don't think you need to de-mean atts and defs. Or maybe I'm misunderstanding why you've done that.

Also, you don't seem to be using away_theta?

with pm3.Model() as model:
    # global model parameters
    home        = pm3.Normal('home',      0, .0001)
    tau_att     = pm3.Gamma('tau_att',   .1, .1)
    tau_def     = pm3.Gamma('tau_def',   .1, .1)
    intercept   = pm3.Normal('intercept', 0, .0001)

    # team-specific model parameters

    atts   = pm3.Normal("atts", 
                           mu   =0,
                           tau  =tau_att, 
                           shape=num_teams)

    home = pm3.Normal("mean_defs", 0, .0001)
    defs   = pm3.Normal("defs", 
                           mu   =home,
                           tau  =tau_def,  
                           shape=num_teams) 

    home_theta  = tt.exp(intercept + atts[away_team] + defs[home_team])

    # likelihood of observed data
    home_points = pm3.Poisson('home_points', mu=home_theta, observed=observed_home_goals)

springcoil · 2015-06-17T21:46:57Z

The demeaning was necessary when reproducing the paper. It might be just a
hack though.

On Wednesday, 17 June 2015, John Salvatier notifications@github.com wrote:

Cool! Thanks :)

A few comments.

You don't need the first model = pm3.Model() since you already use the
with statement.
I would fit a slightly different model. I don't think you need to de-mean
atts and defs. Or maybe I'm misunderstanding why you've done that.

Also, you don't seem to be using away_theta?

with pm3.Model() as model:
# global model parameters
home = pm3.Normal('home', 0, .0001)
tau_att = pm3.Gamma('tau_att', .1, .1)
tau_def = pm3.Gamma('tau_def', .1, .1)
intercept = pm3.Normal('intercept', 0, .0001)
# team-specific model parameters

atts   = pm3.Normal("atts",
                       mu   =0,
                       tau  =tau_att,
                       shape=num_teams)

home = pm3.Normal("mean_defs", 0, .0001)
defs   = pm3.Normal("defs",
                       mu   =home,
                       tau  =tau_def,
                       shape=num_teams)

home_theta  = tt.exp(intercept + atts[away_team] + defs[home_team])

# likelihood of observed data
home_points = pm3.Poisson('home_points', mu=home_theta, observed=observed_home_goals)
—
Reply to this email directly or view it on GitHub
#764 (comment).

Peadar Coyle
Skype: springcoilarch
www.twitter.com/springcoil
peadarcoyle.wordpress.com

springcoil · 2015-06-17T21:48:03Z

The away theta should be used but maybe I deleted it. I will look tomorrow

On Wednesday, 17 June 2015, Peadar Coyle peadarcoyle@googlemail.com wrote:

The demeaning was necessary when reproducing the paper. It might be just a
hack though.

On Wednesday, 17 June 2015, John Salvatier <notifications@github.com
javascript:_e(%7B%7D,'cvml','notifications@github.com');> wrote:
Cool! Thanks :)

A few comments.

You don't need the first model = pm3.Model() since you already use the
with statement.
I would fit a slightly different model. I don't think you need to de-mean
atts and defs. Or maybe I'm misunderstanding why you've done that.

Also, you don't seem to be using away_theta?

with pm3.Model() as model:
# global model parameters
home = pm3.Normal('home', 0, .0001)
tau_att = pm3.Gamma('tau_att', .1, .1)
tau_def = pm3.Gamma('tau_def', .1, .1)
intercept = pm3.Normal('intercept', 0, .0001)
# team-specific model parameters

atts   = pm3.Normal("atts",
                       mu   =0,
                       tau  =tau_att,
                       shape=num_teams)

home = pm3.Normal("mean_defs", 0, .0001)
defs   = pm3.Normal("defs",
                       mu   =home,
                       tau  =tau_def,
                       shape=num_teams)

home_theta  = tt.exp(intercept + atts[away_team] + defs[home_team])

# likelihood of observed data
home_points = pm3.Poisson('home_points', mu=home_theta, observed=observed_home_goals)
—
Reply to this email directly or view it on GitHub
#764 (comment).
Peadar Coyle
Skype: springcoilarch
www.twitter.com/springcoil
peadarcoyle.wordpress.com

Peadar Coyle
Skype: springcoilarch
www.twitter.com/springcoil
peadarcoyle.wordpress.com

jsalvatier · 2015-06-17T21:48:12Z

Ah, gotcha, its reproducing a paper. That makes a lot of sense. Maybe add a
comment that you might not want to do that.

On Wed, Jun 17, 2015 at 2:46 PM, springcoil notifications@github.com
wrote:

The demeaning was necessary when reproducing the paper. It might be just a
hack though.

On Wednesday, 17 June 2015, John Salvatier notifications@github.com
wrote:

Cool! Thanks :)

A few comments.

You don't need the first model = pm3.Model() since you already use the
with statement.
I would fit a slightly different model. I don't think you need to de-mean
atts and defs. Or maybe I'm misunderstanding why you've done that.

Also, you don't seem to be using away_theta?

with pm3.Model() as model:

global model parameters

home = pm3.Normal('home', 0, .0001)
tau_att = pm3.Gamma('tau_att', .1, .1)
tau_def = pm3.Gamma('tau_def', .1, .1)
intercept = pm3.Normal('intercept', 0, .0001)

team-specific model parameters

atts = pm3.Normal("atts",
mu =0,
tau =tau_att,
shape=num_teams)

home = pm3.Normal("mean_defs", 0, .0001)
defs = pm3.Normal("defs",
mu =home,
tau =tau_def,
shape=num_teams)

home_theta = tt.exp(intercept + atts[away_team] + defs[home_team])

likelihood of observed data

home_points = pm3.Poisson('home_points', mu=home_theta,
observed=observed_home_goals)

—
Reply to this email directly or view it on GitHub
#764 (comment).

Peadar Coyle
Skype: springcoilarch
www.twitter.com/springcoil
peadarcoyle.wordpress.com

—
Reply to this email directly or view it on GitHub
#764 (comment).

springcoil · 2015-06-17T21:50:37Z

I will add a comment.
Something about the mean of teams affected the results. I forget the
details.

I will flesh it out :)

On Wednesday, 17 June 2015, John Salvatier notifications@github.com wrote:

Ah, gotcha, its reproducing a paper. That makes a lot of sense. Maybe add a
comment that you might not want to do that.

On Wed, Jun 17, 2015 at 2:46 PM, springcoil <notifications@github.com
javascript:_e(%7B%7D,'cvml','notifications@github.com');>
wrote:

The demeaning was necessary when reproducing the paper. It might be just
a
hack though.

On Wednesday, 17 June 2015, John Salvatier <notifications@github.com
javascript:_e(%7B%7D,'cvml','notifications@github.com');>
wrote:

Cool! Thanks :)

A few comments.

You don't need the first model = pm3.Model() since you already use the
with statement.
I would fit a slightly different model. I don't think you need to
de-mean
atts and defs. Or maybe I'm misunderstanding why you've done that.

Also, you don't seem to be using away_theta?

with pm3.Model() as model:

global model parameters

home = pm3.Normal('home', 0, .0001)
tau_att = pm3.Gamma('tau_att', .1, .1)
tau_def = pm3.Gamma('tau_def', .1, .1)
intercept = pm3.Normal('intercept', 0, .0001)

team-specific model parameters

atts = pm3.Normal("atts",
mu =0,
tau =tau_att,
shape=num_teams)

home = pm3.Normal("mean_defs", 0, .0001)
defs = pm3.Normal("defs",
mu =home,
tau =tau_def,
shape=num_teams)

home_theta = tt.exp(intercept + atts[away_team] + defs[home_team])

likelihood of observed data

home_points = pm3.Poisson('home_points', mu=home_theta,
observed=observed_home_goals)

—
Reply to this email directly or view it on GitHub
#764 (comment).

Peadar Coyle
Skype: springcoilarch
www.twitter.com/springcoil
peadarcoyle.wordpress.com

—
Reply to this email directly or view it on GitHub
#764 (comment).

—
Reply to this email directly or view it on GitHub
#764 (comment).

Peadar Coyle
Skype: springcoilarch
www.twitter.com/springcoil
peadarcoyle.wordpress.com

springcoil · 2015-08-02T12:19:26Z

I know this is still on me, but I've been busy lately. I'll get back to it soon

springcoil · 2015-08-13T16:50:09Z

I made some updates based on what you said John. I think it should be ready to be merged soon. Let me know if you need any other documentation.

springcoil · 2015-08-13T19:56:06Z

It passed the checks and the documentation is ready. Can someone like @jsalvatier or @twiecki merge this?

twiecki · 2015-08-14T06:29:58Z

Thanks! Isn't there more description in PP for hackers or a blog post we could include here? Would make the example much stronger.

springcoil · 2015-08-14T07:05:38Z

Hey Thomas,
Yeah sure - I'll dive into my blog posts and turn the .py file into
something with more substance.

I agree the example could be made much stronger

ᐧ

On Fri, Aug 14, 2015 at 8:30 AM, Thomas Wiecki notifications@github.com
wrote:

Thanks! Isn't there more description in PP for hackers or a blog post we
could include here? Would make the example much stronger.

—
Reply to this email directly or view it on GitHub
#764 (comment).

Peadar Coyle
Skype: springcoilarch
www.twitter.com/springcoil
peadarcoyle.wordpress.com

twiecki · 2015-08-14T07:32:54Z

We probably don't even need the .py file, just the ipython NB which we could then convert to a .py.

Fix critical bug for elementwise transforms where summing was done incorrectly.

slight update 2

springcoil · 2015-08-22T16:55:56Z

I updated the IPython notebook, I thought I correctly used rebase,
however it seems I haven't I'll need to squash these commits.
I used git fetch upstream and git rebase -i origin master

Fix typos in backend documentation re parallelize tests move deps back where they belong fix parallel tests typo upgrade to pymc3 Beta! add link for NUTS commas set transparency for histograms make transforms classes typo add basic transform tests actually basic test transforms categorical foolings improvements for simplex, but still broken test and fix simplex jacobian fix logtransform test all jacobians check transforms are going the right way fix categorical rename transforms test and fix sum_to_1 fix unit continuous try docker builds remove sudo Revert "remove sudo" This reverts commit f2ce835. Revert "try docker builds" This reverts commit b4d43ad. added some docstrings don't import nonexistant things Combine chains by default Make all methods for accessing variable values concatenate chains by default, because this is likely to be the desired output most of the time. See #758 and #759. Simplify single-trace __getitem__ method The get_values fallback in the single-trace __getitem__ method was present mostly as leftover code from a previous design. MultiTrace is the only user-facing trace object, so there is not much benefit of letting __getitem__ fall back to get_values for convenience. Allow trace variable values to be sliced Extend indexing of trace objects to support an additional slice object that specifies the burn and thin arguments of get_values: >>> trace[x, start::step] Above differs from >>> trace[x][start::step] because the second form operates on an array that is the combination of all chains. BaseTrace: Remove chain keyword argument to point This is stale code. The current BaseTrace only deals with a for a chain argument in the point method. The point method signatures in the children classes are already correct. MultiTrace: Rework docstring * Add more information about getting variable values. * Remove parameter information, which isn't relevant because users get back an initialized instance from sampling. backends/base.py: Fix docstring typo Differentiate names for single-chain traces Previously, the variable name "trace" was confusingly used to refer to both BaseTrace instances and MultiTrace instances, making readers infer which type it was from context. Change code to use the following conventions: * Use "strace" for variable names of BaseTrace instances to indicate a single-chain trace (where "single-chain" trace refers to an object derived from BaseTrace, but not a MultiTrace object with only one chain). * MultiTrace instances are called either "mtrace" or "trace". All changes in this commit are purely variable renames and are not user-facing. Added alpha as an argument to traceplot NDArray: Rename variable Rename variable to make it clear that it refers to values for a single model variable rather than a trace object. #771 (comment) Remove gridspec import check Commit 3c3273d added a check on the gridspec import to support older versions of matplotlib. However, the minimum version that is now specified (1.2.1) has gridspec, so this check is no longer needed. combine leapfrog and energy computations combine metropolis density comparison cleanup things revert some unintentional changes avoid infinite recursion actually fix arraystep work with tensors too shapes have to be ints fix lkjcorr distribution fastarraystep->arraystepshared DOC Remove comment about TransformedVar in stoch vol example. Removed TransformedVar from profiling example Added standard normal cdf Added ExGaussian distribution and (rubbish) tests Added 2-parameter Inverse Gaussian distribution Added 3-parameter shifted inverse gaussian. Fix for phi parametrization in the inverse gaussian Added newline at end of file. Improved handling of alternative parametrization for inverse gaussian. Added mean attribute to inverse gaussian. Added switch to ExGaussian logp and missing self to method random. Moved inverse gaussian stuff (and tests) to Wald. Added condition to bound function in Wald logp Fixed Wald logp Fixed typos and removed whitespace Remove unnecssary size checks from Wald/exGaussian random Added disc_vars property to complement cont_vars. Might be useful for assigning discretes to a Metropolis sampler, for example. MultiTrace: Check output before returning slice Out-of-memory backends give a warning when the user tries to slice them, so the result may be a list of Nones. re: #790 Support slicing in SQlite and Text backends Return a NDArray slice instead of warning when an out-of-memory backend is sliced. Fixes #790 ENH Add example of a Gaussian Mixture Model. Add author fix inappropriate summing fix test for elemwise transform jacobian dets remove debug statement

springcoil · 2015-08-22T19:00:58Z

Anyone have any idea why this build is failing?

twiecki · 2015-08-22T19:02:02Z

#803

twiecki · 2015-08-22T20:20:15Z

But also, something seems to have gone wrong with the rebase.

git checkout mybranch
git rebase master
git push -f origin mybranch

springcoil · 2015-08-22T20:34:54Z

Opening a new branch :)

'

Second commit removed get_ipython Update of the ipython notebook with more information I added a slight update to include more tutorial material slight update 2 Use Text as the example backend in docstring Fix typos in backend documentation re parallelize tests move deps back where they belong fix parallel tests typo upgrade to pymc3 Beta! add link for NUTS commas set transparency for histograms make transforms classes typo add basic transform tests actually basic test transforms categorical foolings improvements for simplex, but still broken test and fix simplex jacobian fix logtransform test all jacobians check transforms are going the right way fix categorical rename transforms test and fix sum_to_1 fix unit continuous try docker builds remove sudo Revert "remove sudo" This reverts commit f2ce835. Revert "try docker builds" This reverts commit b4d43ad. added some docstrings don't import nonexistant things Combine chains by default Make all methods for accessing variable values concatenate chains by default, because this is likely to be the desired output most of the time. See #758 and #759. Simplify single-trace __getitem__ method The get_values fallback in the single-trace __getitem__ method was present mostly as leftover code from a previous design. MultiTrace is the only user-facing trace object, so there is not much benefit of letting __getitem__ fall back to get_values for convenience. Allow trace variable values to be sliced Extend indexing of trace objects to support an additional slice object that specifies the burn and thin arguments of get_values: >>> trace[x, start::step] Above differs from >>> trace[x][start::step] because the second form operates on an array that is the combination of all chains. BaseTrace: Remove chain keyword argument to point This is stale code. The current BaseTrace only deals with a for a chain argument in the point method. The point method signatures in the children classes are already correct. MultiTrace: Rework docstring * Add more information about getting variable values. * Remove parameter information, which isn't relevant because users get back an initialized instance from sampling. backends/base.py: Fix docstring typo Differentiate names for single-chain traces Previously, the variable name "trace" was confusingly used to refer to both BaseTrace instances and MultiTrace instances, making readers infer which type it was from context. Change code to use the following conventions: * Use "strace" for variable names of BaseTrace instances to indicate a single-chain trace (where "single-chain" trace refers to an object derived from BaseTrace, but not a MultiTrace object with only one chain). * MultiTrace instances are called either "mtrace" or "trace". All changes in this commit are purely variable renames and are not user-facing. Added alpha as an argument to traceplot NDArray: Rename variable Rename variable to make it clear that it refers to values for a single model variable rather than a trace object. #771 (comment) Remove gridspec import check Commit 3c3273d added a check on the gridspec import to support older versions of matplotlib. However, the minimum version that is now specified (1.2.1) has gridspec, so this check is no longer needed. combine leapfrog and energy computations combine metropolis density comparison cleanup things revert some unintentional changes avoid infinite recursion actually fix arraystep work with tensors too shapes have to be ints fix lkjcorr distribution fastarraystep->arraystepshared DOC Remove comment about TransformedVar in stoch vol example. Removed TransformedVar from profiling example Added standard normal cdf Added ExGaussian distribution and (rubbish) tests Added 2-parameter Inverse Gaussian distribution Added 3-parameter shifted inverse gaussian. Fix for phi parametrization in the inverse gaussian Added newline at end of file. Improved handling of alternative parametrization for inverse gaussian. Added mean attribute to inverse gaussian. Added switch to ExGaussian logp and missing self to method random. Moved inverse gaussian stuff (and tests) to Wald. Added condition to bound function in Wald logp Fixed Wald logp Fixed typos and removed whitespace Remove unnecssary size checks from Wald/exGaussian random Added disc_vars property to complement cont_vars. Might be useful for assigning discretes to a Metropolis sampler, for example. MultiTrace: Check output before returning slice Out-of-memory backends give a warning when the user tries to slice them, so the result may be a list of Nones. re: #790 Support slicing in SQlite and Text backends Return a NDArray slice instead of warning when an out-of-memory backend is sliced. Fixes #790 ENH Add example of a Gaussian Mixture Model. Add author fix inappropriate summing fix test for elemwise transform jacobian dets remove debug statement readme update' ' update readme.md

slight update

springcoil · 2015-08-23T16:41:57Z

This is a lot of commits - can anyone advise me on how to properly rebase this?

springcoil · 2015-08-24T17:26:07Z

Hmm I tried rebasing etc and it didn't work...

kyleam · 2015-08-24T18:15:52Z

Hmm I tried rebasing etc and it didn't work...

It seems like things got into a pretty messy state. I've pushed what (I
think) was the intended state to km/rubgy-fix.

So I'd suggest you

Back up your local rugby_analytics branch from Rugby analytics: Another commit #807 to another branch.
Fetch from the pymc3 repo.
With rugby_analytics as your current branch, reset it to the new
branch I pushed.

git reset --hard /km/rubgy-fix
Re-run the notebook and update the commit with 'git commit --amend'.
Force push to the pymc3 repo. This will update PR Rugby analytics: Another commit #807.

kyleam · 2015-08-24T18:17:31Z

Doh, please look under the folded text for my last comment. GitHub thinks the rest of my email is my signature.

Rugby Analytics commit

caea43d

Second commit

eaad24b

removed get_ipython

3b77bba

twiecki and others added 5 commits August 17, 2015 15:01

ENH Add example of a Gaussian Mixture Model.

cd52ba4

Add author

2e9a8a6

Merge pull request #800 from pymc-devs/transform_fix

9476ba5

Fix critical bug for elementwise transforms where summing was done incorrectly.

Update of the ipython notebook with more information

a23ba6f

I added a slight update to include more tutorial material

c82b868

slight update 2

springcoil and others added 5 commits August 23, 2015 16:45

readme update'

cd7cc89

'

update readme.md

d12de8e

Merge branch 'master' of https://github.com/springcoil/pymc3

8436452

Updating imports based on feedback - included more of a tutorial

eceac1c

slight update

twiecki mentioned this pull request Aug 25, 2015

DOC Add Rugby analysis example #808

Merged

twiecki closed this Aug 25, 2015

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Rugby Analytics commit #764

Rugby Analytics commit #764

springcoil commented Jun 17, 2015

jsalvatier commented Jun 17, 2015

springcoil commented Jun 17, 2015

springcoil commented Jun 17, 2015

jsalvatier commented Jun 17, 2015

global model parameters

team-specific model parameters

likelihood of observed data

springcoil commented Jun 17, 2015

global model parameters

team-specific model parameters

likelihood of observed data

springcoil commented Aug 2, 2015

springcoil commented Aug 13, 2015

springcoil commented Aug 13, 2015

twiecki commented Aug 14, 2015

springcoil commented Aug 14, 2015

twiecki commented Aug 14, 2015

springcoil commented Aug 22, 2015

springcoil commented Aug 22, 2015

twiecki commented Aug 22, 2015

twiecki commented Aug 22, 2015

springcoil commented Aug 22, 2015

springcoil commented Aug 23, 2015

springcoil commented Aug 24, 2015

kyleam commented Aug 24, 2015

kyleam commented Aug 24, 2015

Rugby Analytics commit #764

Rugby Analytics commit #764

Conversation

springcoil commented Jun 17, 2015

jsalvatier commented Jun 17, 2015

springcoil commented Jun 17, 2015

springcoil commented Jun 17, 2015

jsalvatier commented Jun 17, 2015

global model parameters

team-specific model parameters

likelihood of observed data

springcoil commented Jun 17, 2015

global model parameters

team-specific model parameters

likelihood of observed data

springcoil commented Aug 2, 2015

springcoil commented Aug 13, 2015

springcoil commented Aug 13, 2015

twiecki commented Aug 14, 2015

springcoil commented Aug 14, 2015

twiecki commented Aug 14, 2015

springcoil commented Aug 22, 2015

springcoil commented Aug 22, 2015

twiecki commented Aug 22, 2015

twiecki commented Aug 22, 2015

springcoil commented Aug 22, 2015

springcoil commented Aug 23, 2015

springcoil commented Aug 24, 2015

kyleam commented Aug 24, 2015

kyleam commented Aug 24, 2015