Sampling error - Initial evaluation of model at starting point failed!

Questions v5
1

I’m just trying to carry out a bayesian regression and I keep getting a sampling error. This is my model:

product_array = data_aggr_period_prod_BR[‘product’].unique()
for product in product_array:
# Get the prior Normal parameters
regression_results_loop = regression_results_BR[regression_results_BR[‘Product’]==product]
data_aggr_period_prod_loop = data_aggr_period_prod_BR[data_aggr_period_prod_BR[‘product’]==product]
slope_mean = regression_results_loop[[‘Slope’]].mean().values[0]
slope_std = regression_results_loop[[‘Slope’]].std().values[0]
intercept_mean = regression_results_loop[[‘Intercept’]].mean().values[0]
intercept_std = regression_results_loop[[‘Intercept’]].std().values[0]

# Create bayesian model
with pm.Model() as bayesian_model_prod6:
    slope = pm.Normal('slope', mu = slope_mean, sd = slope_std)
    intercept = pm.Normal('intercept', mu = intercept_mean, sd = intercept_mean)

    # Linear regression line
    mean = intercept + slope*data_aggr_period_prod_loop['x']

    # Describe the distribution of our conditional output
    y = pm.Normal('y', mu = mean, observed = data_aggr_period_prod_loop['y']) # Run the sampling using pymc3 for 100 samples
    trace_600 = pm.sample(100, return_inferencedata=True)

The sampling error is being produced from the line where I define trace_600. How could I fix this?

2

Try model.debug() and see if it helps.

Otherwise this seems like a typo?

    intercept = pm.Normal('intercept', mu = intercept_mean, sd = intercept_mean)

I imagine you want sd = intercept_std

3

Yes, it was a typo but now that I have changes that I got this:

SamplingError: Initial evaluation of model at starting point failed!
Starting values:
{‘slope’: array(0.00049181), ‘intercept’: array(0.80299521)}

Initial evaluation results:
slope -inf
intercept -inf
y -23.72
Name: Log-probability of test_point, dtype: float64

I will try the debug that you suggested

4

What is your slope_mean and slope_std values. Maybe they are too extreme and you are getting numerical issues.

5

my slope mean value is 0.0004918058105516407
and slope std is nan which is probably the problem

6

That is definitely bad :slight_smile: