Average Loss = inf

Although it’s not new topic, but i couldn’t understand much on this subject from the previous posts also because I am a beginner. I have tried changing SD to 10, or even 100.

Code: https://github.com/Harpreetsingh31/MachineLearning/blob/master/Classification/Breastcancer/Logistic-PyMC3.ipynb
Data: https://github.com/Harpreetsingh31/MachineLearning/tree/master/Classification/Breastcancer

Non-probabilistic models (sci-kit) are in Jbook: “aio.ipynb”

Code implementation, I followed was from Nicole Carlson.

I want to know what is causing this error and possible solutions.

Many thanks in advance.

Hello Harpreetsingh31,

In your code you use

df[df == '?'] = np.nan

which I can imagine will mess things up if those nan values are present while training. Can you make sure all those values are dropped?

Yeah, i thought the same, so I basically dropped that entire column (‘bare_nuclei’)

##IMP NOTE: I basically dropped ‘bare_nuclei’ bcuz of 16 “?” values
X = scale(np.array(df.drop([‘class’,‘bare_nuclei’],1)))

When you do

for RV in logistic_model.basic_RVs:
    print(RV.name, RV.logp(linear_model.test_point))

The output shows that the logp of observed y is inf
This means that the input p is out of support.
Try doing p.tag.test_value and see where p is returning value outside of [0, 1] (the support of p)

I added the line “p.tag.test_value” in for loop, it prints

alpha -0.9189385332046727
betas -7.351508265637381
y -inf

So, Bernoulli returns inf for p=0.5.

I would say there are a couple of problem:
1, the observed y contains value 2 and 4 instead of 0 and 1, which cause Bernoulli goes inf
2, you get one value for p, but you should get a vector the same size as y

Try the code below:

#Split Data
X_tr, X_te, y_tr, y_te = train_test_split(X, y/2-1, test_size=0.2, random_state=42)

model_input = shared(X_tr)
model_output = shared(y_tr)
with pm.Model() as logistic_model:
    # Priors for unknown model parameters    
    alpha = pm.Normal("alpha", mu=0,sd=1)
    betas = pm.Normal("betas", mu=0, sd=1, shape=(X.shape[1], 1))
    # Expected value of outcome
    p = pm.invlogit(alpha + T.dot(X_tr, betas))
    # Likelihood (sampling distribution of observations)
    y = pm.Bernoulli('y', p, observed=model_output)
1 Like

Thank you so much Junpenglao, I am sorry for that silly mistake, I had actually made changes to the dataframe but for some reason, they didn’t get reflected in ‘y’. Anyways, I think things are making sense now.

However, when i am calculating accuracy, I see that pred vector has got the dimensions as that of y_tr instead of y_te.

Unrelated questions:

  1. Do we have probabilistic logistic regression model for multi-class problems?

  2. And is there easier or general way to plot “Uncertainty in predicted value”?.
    In the linked article, Thomas Wiecki does by defining a grid and then plotting contour on top but i replicate for my application of NeuralNetwork (git) because of reshape function .

You need to do model_input.set_value(X_te)

Yes, you can model the observed classes as a Categorical random variable. Usually ppl do a softmax on a matrix and use the matrix as p for the Categorical.

There are many ways to do it, one way is to plot each ppc_sample and then plot the observed on top. For example, see the visualization in http://junpenglao.xyz/Blogs/posts/2017-10-23-OOS_missing.html

You can also get some inspiration from http://docs.pymc.io/notebooks/posterior_predictive.html

1 Like