What about transferring this repository into an organization. Maybe inside pymc-devs? Together with other books/docs? Thoughts @junpenglao, @AustinRochford @twiecki

It might be helpful for contributors to have a conda environment file so they can quickly create a new environment and get things rolling. I have one I'll paste below, but I haven't checked it against all of the notebooks to make sure everything necessary is included and anything extraneous is not.

Here's mine:

name: pymc
channels:
- conda-forge/label/rc
- conda-forge
- defaults
dependencies:
- cycler=0.10.0=py36_0
- freetype=2.7=1
- joblib=0.11=py36_0
- libgpuarray=0.6.2=np112py36_0
- libpng=1.6.28=0
- mako=1.0.6=py36_0
- matplotlib=2.0.0=np112py36_3
- nose=1.3.7=py36_2
- pandas=0.19.2=np112py36_1
- patsy=0.4.1=py36_0
- pyparsing=2.2.0=py36_0
- pytz=2017.2=py36_0
- theano=0.9.0=py36_0
- tqdm=4.11.2=py36_0
- pymc3=3.1rc3=py36_0
- appnope=0.1.0=py36_0
- bleach=1.5.0=py36_0
- decorator=4.0.11=py36_0
- entrypoints=0.2.2=py36_1
- h5py=2.7.0=np112py36_0
- hdf5=1.8.17=1
- html5lib=0.999=py36_0
- ipykernel=4.6.1=py36_0
- ipython=5.3.0=py36_0
- ipython_genutils=0.2.0=py36_0
- jinja2=2.9.6=py36_0
- jsonschema=2.5.1=py36_0
- jupyter_client=5.0.1=py36_0
- jupyter_core=4.3.0=py36_0
- markupsafe=0.23=py36_2
- mistune=0.7.4=py36_0
- mkl=2017.0.1=0
- mkl-service=1.1.2=py36_3
- nbconvert=5.1.1=py36_0
- nbformat=4.3.0=py36_0
- notebook=5.0.0=py36_0
- numpy=1.12.1=py36_0
- openssl=1.0.2k=1
- pandocfilters=1.4.1=py36_0
- path.py=10.1=py36_0
- pexpect=4.2.1=py36_0
- pickleshare=0.7.4=py36_0
- pip=9.0.1=py36_1
- prompt_toolkit=1.0.14=py36_0
- ptyprocess=0.5.1=py36_0
- pygments=2.2.0=py36_0
- python=3.6.1=0
- python-dateutil=2.6.0=py36_0
- pyzmq=16.0.2=py36_0
- readline=6.2=2
- scipy=0.19.0=np112py36_0
- seaborn=0.7.1=py36_0
- setuptools=27.2.0=py36_0
- simplegeneric=0.8.1=py36_1
- six=1.10.0=py36_0
- sqlite=3.13.0=0
- terminado=0.6=py36_0
- testpath=0.3=py36_0
- tk=8.5.18=0
- tornado=4.4.2=py36_0
- traitlets=4.3.2=py36_0
- wcwidth=0.1.7=py36_0
- wheel=0.29.0=py36_0
- xz=5.2.2=1
- zlib=1.2.8=3
- pip:
  - ipython-genutils==0.2.0
  - jupyter-client==5.0.1
  - jupyter-core==4.3.0
  - prompt-toolkit==1.0.14
  - pygpu==0.6.2
prefix: /Users/Peter/anaconda/envs/pymc

Parts of Chapter 11. Two notes:

• This PyMC3 pull request will make ordinal regression simpler, so I'll wait on that.
• The book uses a nonstandard parametrization of the beta(-binomial) distribution, which requires some care (and reading the R rethinking source) to replicate, which is slowing me down a bit.

Cant understand this piece of code in 4.54 and 4.58:

weigth_seq = np.arange(25, 71)
# Given that we have a lot of samples we can use less of them for plotting (or we can use all!)
chain_N_thinned = chain_N[::10]
mu_pred = np.zeros((len(weigth_seq), len(chain_N_thinned)*chain_N.nchains))
for i, w in enumerate(weigth_seq):
    mu_pred[i] = chain_N_thinned['alpha'] + chain_N_thinned['beta'] * w

What is chain_N[::10]? and chain_N.nchains doing? What is mu_pred supposed to be?

I've been working through the not yet released 2nd version of McElreath's book and have been using this repository to help guide me along the way. Its been extremely helpful, thanks! As of about ch.6 there is enough difference between the versions that it has become quite difficult to use. Any interest in getting a head start on the 2nd edition code translations so when he releases it, the code with be available in python as well? The 2nd edition can be found on his website w/ a password on one of the first 2019 youtube videos.

If I run this cell:

rt = chimp_ensemble['pulled_left']
pred_mean = np.zeros((1000, 4))
cond = d.condition.unique()
prosoc_l = d.prosoc_left.unique()

for i in range(len(rt)):
    
    tmp = []
    for cp in cond:
        for pl in prosoc_l:
            tmp.append(np.mean(rt[i][(d.prosoc_left==pl) & (d.chose_prosoc==cp)]))
    pred_mean[i, :] = tmp
        
ticks = range(4)
mp = pred_mean.mean(0)
hpd = pm.hpd(pred_mean)

plt.figure(figsize=(13,6))
plt.fill_between(ticks, hpd[:,1], hpd[:,0], alpha=0.25, color='k')
plt.plot(mp, color='k')
plt.xticks(ticks, ("0/0","1/0","0/1","1/1"))
chimps = d.groupby(['actor', 'prosoc_left', 'condition']).agg('mean')['pulled_left'].values.reshape(7, -1)
for i in range(7):
    plt.plot(chimps[i], 'C0')

plt.ylim(0, 1.1);

I get this error:

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-18-509a6637b6be> in <module>()
     11     for cp in cond:
     12         for pl in prosoc_l:
---> 13             tmp.append(np.mean(rt[i][(d.prosoc_left==pl) & (d.chose_prosoc==cp)]))
     14     pred_mean[i, :] = tmp
     15 

IndexError: too many indices for array

which it is very strange. Fortunately, after an hour, I found a fix. Change the loop to this one:

for i in range(len(rt)):
    
    tmp = []
    if rt[i].shape == ():
        continue
    for cp in cond:
        for pl in prosoc_l:
            tmp.append(np.mean(rt[i][(d.prosoc_left==pl) & (d.chose_prosoc==cp)]))
    pred_mean[i, :] = tmp

Since almost all the chapter is ported (I will also work on Chapter 14), maybe it is good to

• [ ] add legends to the figures in the early chapter.
• [ ] reproduce the rest of the figures shown in the book. Some of the figures are related to the analysis without code presented.

In 4.66, we set beta to generate two numbers for our quadratic model so both are generated from the same distribution. But the book generates it from two different distributions. Why would we do this? Doesn't this decrease the flexibility of the model?

with pm.Model() as m_4_5:
    alpha = pm.Normal('alpha', mu=178, sd=100)
    
    beta = pm.Normal('beta', mu=0, sd=10, shape=2) # shape of 2 generates two betas
    sigma = pm.Uniform('sigma', lower=0, upper=50)
    mu = pm.Deterministic('mu', alpha + beta[0] * d.weight_std + beta[1] * d.weight_std2)
    height = pm.Normal('height', mu=mu, sd=sigma, observed=d.height)
    trace_4_5 = pm.sample(1000, tune=1000)

It seems to me that the Code 5.55, as it presently stands, does not reflect the intended "index variable" approach introduced in the book? Hence, I propose the following code:

with pm.Model() as m5_16_alt:
    a = pm.Normal('a',mu = 0.6, sd=10, shape=len(d['clade_id'].unique()))
    mu = pm.Deterministic('mu', a[d['clade_id'].values])
    sigma = pm.Uniform('sigma', lower= 0 , upper= 10)
    kcal_per_g = pm.Normal('kcal_per_g', mu = mu, sd=sigma, observed = d['kcal.per.g'])
    trace_5_16_alt = pm.sample(1000, tune=1000) 

Proposed code includes shape parameter for the variable a and uses index variable the way it was intended by the author of the book. Is my reasoning on this correct? The summary produces following output (excerpt):

a:

  Mean             SD               MC Error         89% HPD interval
  -------------------------------------------------------------------
  
  0.544            0.044            0.001            [0.482, 0.620]
  0.713            0.044            0.001            [0.641, 0.781]
  0.788            0.054            0.002            [0.709, 0.882]
  0.506            0.059            0.002            [0.407, 0.593]

sigma:

  Mean             SD               MC Error         89% HPD interval
  -------------------------------------------------------------------
  
  0.131            0.019            0.001            [0.101, 0.159]

Compare this output with the book (page 159).

I am working to finish Chp 10 and meanwhile also did some clean up of the the finished notebooks:

1, add column names to pm.compare output, so that the dataframe and compare plot has better label (should add this feature to pymc3 main retro as well): 2, using new init (jitter+adapt_diag) and increase tuning samples (partly get rid of the warning of acceptance probability mismatch the target) 3, retina display across notebook 4, remove pm.effective_n and pm.gelman_rubin as it is now included in pm.summary. 5, changing pm.df_summary to pm.summary.