Bambi error: "cannot pickle 'fortran' object" (original) (raw)

November 19, 2022, 6:04pm 1

a = np.random.normal(6, 2.5, 160)
b = np.random.normal(8, 2, 120)
df = pd.DataFrame({“Group”: [“a”] * 160 + [“b”] * 120, “Val”: np.hstack([a, b])})
df

model_1 = bmb.Model(“Val ~ Group”, df)
results_1 = model_1.fit()

#I am a novice:
problem: even the simplest example from bambi gives the error: “cannot pickle ‘fortran’ object”. I reinstalled anaconda and pip installed bambi but did not correct error. I am new to pymc3 and bambi. Can anyone suggest a managable solution?
thanking you in advance. DD

tcapretto November 22, 2022, 1:30am 2

Hello!

Could you share which versions of PyMC and Bambi you’re using?

You can do

import bambi as bmb
import pymc as pm

print("Bambi:", bmb.__version__)
print("PyMC": pm.__version__)

If import pymc as pm does not work, it may indicate you’re using pymc3. We encourage using the new version of PyMC, which is v4.

Nn_Nnn November 22, 2022, 12:30pm 3

[quote=“tcapretto, post:2, topic:10858”]

`Thank you for your help.

I installed pymc with: ‘pip install pymc’`
then I imported:
import arviz as az
import bambi as bmb
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
import pymc as pm

print(“Bambi:”, bmb.version)
print(“PyMC:”, pm.version)
gives:
Bambi: 0.9.1
PyMC: 4.3.0

then following the Bambi first example from the Bambi documentation:

Comparison of two means (T-test):

bambinos.github Comparison of two means (T-test) — Bambi 0.9.1 documentation)

a = np.random.normal(6, 2.5, 160)
b = np.random.normal(8, 2, 120)
df = pd.DataFrame({“Group”: [“a”] * 160 + [“b”] * 120, “Val”: np.hstack([a, b])})
df.sample(10)

model_1 = bmb.Model(“Val ~ Group”, df)
results_1 = model_1.fit()
this still gives:
‘cannot pickle ‘fortran’ object’ error.

If instead I
import pymc3 as pm

print(“Bambi:”, bmb.version)
print(“PyMC3:”, pm.version)
I get:
Bambi: 0.9.1
PyMC: 3.11.5
and then run the model, I still get the same error.
mayby I am not installing pymc correctly and the two versions are interferring with each other?

I have since gone on to run other models on Bambi successfully, but when I return to this example above from the Bambi website it fails with the aforementioned error.
If I modify the model by setting the intercept to zero it then works!:
model_2 = bmb.Model(“Val ~ 0+Group”, df)
results_2 = model_2.fit()
but why the first model, designated model1(model_1 = bmb.Model(“Val ~ Group”, df)) on the Bambi website fails on my computer I do not know. Comparing two means is a useful model.
Putting this all together I am not sure if its just this example model that is at fault, or that my install is mixed up between pymc3 and pymc4. I was using pymc3 since the learning examples seem to be mostly pymc3, however I would be happy to migrate to pymc4 if the code transition is not too difficult.
Any ideas as to why the Bambi documentation example model 1 is failing? would give confidence to using other Bambi models. Thank you for expertise and guidance on the above. DD

tcapretto November 24, 2022, 12:36pm 4

Could you create a new environment and install the development version?

pip install git+https://github.com/bambinos/bambi.git

Nn_Nnn November 24, 2022, 3:07pm 5

I tried pip install git+https://github.com/bambinos/bambi.git into c promp and error message returns "Cannot find command ‘git’. Do you have ‘git’ installed and in your path.
I since tried the following (added cores=1 to results_1 = model_1.fit) and it appears to now work and does not give the error message:“cannot pickle fortran object.”
Example:
np.random.seed(1234)

a = np.random.normal(6, 2.5, 160)
b = np.random.normal(8, 2, 120)
df = pd.DataFrame({“Group”: [“a”] * 160 + [“b”] * 120, “Val”: np.hstack([a, b])})
df.sample(10)

model_1 = bmb.Model(“Val ~ Group”, df)

results_1 = model_1.fit(cores=1)
az.summary(results_1)

also tried adding the argument “cores =1” to other simple models and it prevented the error occurring and gives good results.
eg:
#Bayesian Linear Regression with Bambi | by Khuyen Tran | Towards Data Science
import pandas as pd

df = pd.read_csv(
“https://media.githubusercontent.com/media/khuyentran1401/Data-science”
“/master/statistics/bayes_linear_regression/student-score.csv”)

df.head(10)
plt.scatter(df[‘mat’], df.por, label=“sampled data”)
import bambi as bmb

gauss_model = bmb.Model(‘por ~mat’, data=df)

Fit the model using 1000 on each of 4 chains

gauss_fitted = gauss_model.fit(draws=1000, chains=4, cores=1)
az.summary(gauss_fitted)
Again putting ‘cores=1’ as a model.fit argument allowed the model to work without error. I am using a new HP laptop just purchased running windows11 to learn pymc, Bayesian and Bambi (and python). It could be useful to other novices to know that setting cores=1 will allow the simple starting examples on Bambi to run. Not sure why this fix works?
I am happy now to continue my exploration into Bambi and I am encouraged by the helpful support that is afforded to novices.
Comment:
Bambi is exactly what I think is useful to the novice statistician to explore their data with Bayesian models. A cursory look at the examples (on the Bambi site and elsewhere), I would love to see Bambi models for paired means and paired proportions added to the examples to bridge the gap when moving from classical to Bayesian models. A short introductory book or manual on Bambi for the Novice programmer would be a great starting point for new programmers to cover a broader range of models and would be very welcome.
Meanwhile I will continue to explore the available models.
Thanks again for your time, expertise, interest and help.
DD

tcapretto November 24, 2022, 5:34pm 6

So it looks like it’s a problem with multiprocessing.

For there record, I saw this issue in Github, perhaps it’s related NUTS Sampler Stuck as 0.00% [0/8000 00:00<? Sampling 4 chains, 0 divergences] · Issue #6315 · pymc-devs/pymc · GitHub