theano max works but not min

[sammosummo]

Its seems that my version of PyMC3/Theano won't compile models containing tt.min() but will compile models with tt.max(). I've tried replacing tt.min([a, b], axis=0) with the workaround -tt.max([-a, -b], axis=0) and while this works for the toy example below, in real models, it sometimes yields the same compilation error as using min. This must be related but I haven't figured out the critical issue.

import pymc3 as pm
import theano.tensor as tt


def main():
    with pm.Model() as m1:

        a = pm.Normal(name="a", shape=5)
        b = pm.Normal(name="b", shape=5)
        c = tt.max([a, b], axis=0)

    pm.sample(model=m1)  # works fine

    with pm.Model() as m2:

        a = pm.Normal(name="a", shape=5)
        b = pm.Normal(name="b", shape=5)
        c = tt.min([a, b], axis=0)

    pm.sample(model=m2)  # raises this exception:
    # ... error: non-constant-expression cannot be narrowed from type 'npy_intp' (aka
    # 'long') to 'int' in initializer list [-Wc++11-narrowing]


if __name__ == '__main__':
    main()

• PyMC3 Version: 3.8
• Theano Version: 1.0.4
• Python Version: 3.8.1
• Operating system: MacOS
• How did you install PyMC3: conda

[aloctavodia]

I am not able to reproduce this error.

• PyMC3 Version: master
• Theano Version: 1.0.4
• Python Version: 3.7.6
• Operating system: Linux

[Ahanmr]

I tried with the exact same specs, I too was not able to reproduce the error, it ran successfully. Not sure, what's causing the error.

Auto-assigning NUTS sampler... Initializing NUTS using jitter+adapt_diag... Sequential sampling (2 chains in 1 job) NUTS: [b, a] Sampling chain 0, 0 divergences: 100%|██████████| 1000/1000 [00:01<00:00, 774.57it/s] Sampling chain 1, 0 divergences: 100%|██████████| 1000/1000 [00:00<00:00, 1363.44it/s] Auto-assigning NUTS sampler... Initializing NUTS using jitter+adapt_diag... Sequential sampling (2 chains in 1 job) NUTS: [b, a] Sampling chain 0, 0 divergences: 100%|██████████| 1000/1000 [00:00<00:00, 1280.23it/s] Sampling chain 1, 0 divergences: 100%|██████████| 1000/1000 [00:00<00:00, 1363.76it/s]

• PyMC3 Version: 3.8

• Theano Version: 1.0.4

• Python Version: 3.6.9 and 3.8.1

• Operating system: Linux

[Ahanmr]

I'm not too sure if this is a work-around, please feel free to correct me,
import theano
theano.config.gcc.cxxflags = "-Wno-c++11-narrowing"

[fonnesbeck]

This flag is set automatically for macOS to solve a different issue, but is not set for Linux. If this solves the problem, we could add it for that platform as well.

[ricardoV94]

@sammosummo Did you keep encountering this issue in newer releases?

[sammosummo]

Works fine with this setup:

• PyMC3 Version: 3.9.3
• Theano Version: 1.0.5
• Python Version: 3.8.5
• Operating system: macOS

[ricardoV94]

Thanks for the feedback. Will close this issue as it seems we did not hear more reports like this