Pymc3: Silent integer overflow in test_values

[mschmidt87]

Description of your problem

I am trying to fit a Negative-Binomial regression model. However, I am running into SamplingError: Initial evaluation of model at starting point failed!, even if I am trying to fit samples from the model prior.

I found out that the test_value of my observation value (which should be positive because of the Negative-Binomial distribution) are sometimes negative, which leads to the error. I then checked model.check_test_point() and get:

alpha_log__   -1.14
mu            -1.61
obs            -inf
Name: Log-probability of test_point, dtype: float64

This happens because some of the test values of the observed variable are negative:

model_instance.obs.tag.test_value.min()

If I understand correctly, the test values of the observed variable are basically the same as the values I am trying to fit:

>>> model_instance.obs.tag.test_value[0][:10]
array([136519719, 139778075, 162329337, 147874471, 176346820, 131886145,
       131664941, 145409569, 141731980, 152439463])
>>> prior_trace['obs'][0][:10]
array([136519719, 139778075, 162329337, 147874471, 176346820, 131886145,
       131664941, 145409569, 141731980, 152439463])

However, for very large sampled values, the test value is negative:

>>> model_instance.obs.tag.test_value.min()
-2146971425
>>> prior_trace['obs'].reshape(-1)[model_instance.obs.tag.test_value.reshape(-1).argmin()]
2147995871
>>> model_instance.obs.tag.test_value.reshape(-1)[np.where(prior_trace['obs'].reshape(-1) != model_instance.obs.tag.test_value.reshape(-1))]
array([-2041161858, -1615277538, -2008654162, -1830866193, -2066486365,
       -1939932683, -1927434696, -1920384235, -1997835169, -1856732049,
       -1849204327, -1857766493, -2045704721, -1855359298, -2123503404,
       -1829876267, -2071833989, -1976140492, -2110619502, -1816045663, ...

Thus, it seems that the test value is, for some reason, overflowing and turning negative for very large values.

Below is a minimal reproducer:

Please provide a minimal, self-contained, and reproducible example.

def model_factory(y):
    with pm.Model() as model:
        alpha = pm.HalfCauchy("alpha", 100)
        mu = pm.Normal('mu', 17., 2.)
        lam = pm.Deterministic('lambda', tt.exp(mu))
        pm.NegativeBinomial("obs", mu=lam, alpha=alpha, observed=y)
    return model

# Just some dummy data to let us sample from the prior
y = np.random.binomial(n=10, p=0.5, size=100)

# Sample from the model's prior
with model_factory(y):
    prior_trace = pm.sample_prior_predictive(1000)

# Fit the model to the prior samples
model_instance = model_factory(prior_trace['obs'])
with model_instance:
    trace = pm.sample()

Please provide the full traceback.

---------------------------------------------------------------------------
SamplingError                             Traceback (most recent call last)
<ipython-input-21-1bfc50620c9f> in <module>
      1 model_instance = model_factory(prior_trace['obs'])
      2 with model_instance:
----> 3     trace = pm.sample()

~/path/to/lib/python3.7/site-packages/pymc3/sampling.py in sample(draws, step, init, n_init, start, trace, chain_idx, chains, cores, tune, progressbar, model, random_seed, discard_tuned_samples, compute_convergence_checks, callback, return_inferencedata, idata_kwargs, mp_ctx, pickle_backend, **kwargs)
    423                 update_start_vals(chain_start_vals, model.test_point, model)
    424 
--> 425     check_start_vals(start, model)
    426     if cores is None:
    427         cores = min(4, _cpu_count())

~/path/to/lib/python3.7/site-packages/pymc3/util.py in check_start_vals(start, model)
    228                 "Initial evaluation of model at starting point failed!\n"
    229                 "Starting values:\n{}\n\n"
--> 230                 "Initial evaluation results:\n{}".format(elem, str(initial_eval))
    231             )
    232 

SamplingError: Initial evaluation of model at starting point failed!
Starting values:
{'alpha_log__': array(4.60517019), 'mu': array(17.)}

Initial evaluation results:
alpha_log__   -1.14
mu            -1.61
obs            -inf
Name: Log-probability of test_point, dtype: float64

Please provide any additional information below.

Versions and main components

• PyMC3 Version: current master (9deb5eedd331c20a63c50d05c8cb169abf8fa47d) (Note that this also happens with the current release)
• Theano Version: 1.0.11
• Python Version: 3.7.7
• Operating system: Mac OS X
• How did you install PyMC3: (conda/pip)

[Spaak]

Thanks @mschmidt87 for the detailed report! Pinging @StephenHogg here, who recently did some work (#4211) on initial model checks before sampling; could be related.

[StephenHogg]

check_test_eval is really just calling model.check_test_point under the hood, which in turn just calls each of the RV objects contained in the model. So I would try just doing that on your initial values in order to pick apart what's going on a bit more.

[mschmidt87]

Yes, I looked a bit in the code, but I think the problem is in the construction of the test value for my observation variable (which becomes negative for very high values for some reason). But I don't really know where those test values are determined/initialized.

[mschmidt87]

Okay, I have digged deeper into this and found the source of the error:

In my example, I am specifying the data as a numpy array with dtype np.int64. PyMC3 then takes that data and calls pandas_to_array on it:

https://github.com/pymc-devs/pymc3/blob/162fd92da607e133a83d360e97fbde2b364cb703/pymc3/model.py#L1785-L1786

Now, in the pandas_to_array, we end in the else branch of a long list of type checks:

https://github.com/pymc-devs/pymc3/blob/162fd92da607e133a83d360e97fbde2b364cb703/pymc3/model.py#L1686-L1703

which executes np.asarray(data). Since my data is already an array, nothing really happens here. But then the code checks if the dtype is an integer and, if yes, calls pm.intX(ret):

https://github.com/pymc-devs/pymc3/blob/162fd92da607e133a83d360e97fbde2b364cb703/pymc3/model.py#L1705-L1708

Now, pm.intX converts the dtype into int32 which overflows at a bit above 2.e9 and this causes the negative values which then lead to the error in the check_test_eval.

Now, I am not familiar with all the intricacies of the code, but the comments suggest this casting might happen because PyMC3 thinks that I am passing an array of indices since the dtype is integer. I don't quite understand why that necessitates casting it to int32 but I presume that is related to theano?
Another thing that is a bit odd is that we are calling pandas_to_array in the first place, because we already passed an array to the model. Thus, either we don't actually want to call pandas_to_array for numpy arrays or the function should be renamed to a more informative name, I would say.

[mschmidt87]

So, to me, there are essentially two questions:

• Why are integers converted to int32 using the conversion map from float64 to int32? (The line in pandas_to_array() I quoted above) Pinging @AlexAndorra who authored the PR (#3925 ) that introduced that line.
• Why are they then converted to int32 and not int64? Pinging @lucianopaz here because he authored the pull-request (#3300 ) that introduced the conversion map.

[lucianopaz]

The reason that we cast to int32 and not to int64 is that the latter can overflow float64 calculations for large integer values, whereas int32 won't. You could try to see if this happens by casting your observations to float64. Maybe some inf values will start popping up.
How were your observations generated? Is there a chance to use int32 throughout the entire process? Is there a way to rescale your observations so that they don't end up being huge?

[mschmidt87]

Thank you for your reply. I could scale down my observations, I guess, at the (acceptable) cost of losing some precision in the model.

However, could you explain what you mean with "we cast to int32 and not to int64 is that the latter can overflow float64 calculations for large integer values"? It's not clear to me what you mean with this.

[michaelosthege]

However, could you explain what you mean with "we cast to int32 and not to int64 is that the latter can overflow float64 calculations for large integer values"? It's not clear to me what you mean with this.

I think he refers to operations like int64 * float64.

Now let's appreciate for a moment, that the recently introduced checks from #4211 surfaced a bug that would have caused very hard to debug problems in the MCMC.

What if we just add a check in intX to raise an error when the conversion results in an under/overflow?

[mschmidt87]

However, could you explain what you mean with "we cast to int32 and not to int64 is that the latter can overflow float64 calculations for large integer values"? It's not clear to me what you mean with this.

I think he refers to operations like int64 * float64.

Okay, I was not aware that this cannot happen with int32 * float64. If there is no safe way around casting those integers to int32, I would suggest to add some warnings to the code to make the user aware of this behavior. In addition, I would suggest to rename the pandas_to_array function to a more informative name.