abc.run stuck at t=1 for integer parameters

Hi, so if for models with integer parameters the abc.run gets stuck at t=1 without running the model. Weirdly, there is no error message or crush. Maybe there is an issue with the transition kernel for integer parameters. The problem can be circumvented by using a float parameter that gets rounded to the next integer, however, it still consider the current behavior a bug. Here’s a MWE:

import pyabc
from pyabc.sampler import SingleCoreSampler
import numpy as np
import tempfile
import os
import matplotlib.pyplot as plt

def model(par):
    # print('new run!')
    return {"data": np.random.binomial(par['n'], 0.5)}

def distance(x, y):
    return abs(x["data"] - y["data"])

prior = pyabc.Distribution(n=pyabc.RV("randint", 1, 21))  # it works if I switch to 'uniform'
abc = pyabc.ABCSMC(model, prior, distance, sampler=SingleCoreSampler())  # SingleCoreSampler because I am on Windows

db_path = ("sqlite:///" +
           os.path.join(tempfile.gettempdir(), "test.db"))
observation = 5
abc.new(db_path, {"data": observation})
history = abc.run(minimum_epsilon=0., max_nr_populations=10)


fig, ax = plt.subplots()
for t in range(1, history.max_t+1, 2):
    df, w = history.get_distribution(m=0, t=t)
    pyabc.visualization.plot_kde_1d(
        df, w,
        xmin=0, xmax=20,
        x="n", ax=ax,
        label="PDF t={}".format(t))
plt.legend()
plt.show()

I am on Windows 10x64 with the latest Anaconda environment, where I installed pyabc via pypi. My pyabc version is 0.10.4.

[sisyga]

Thank you @yannikschaelte , that sounds very promising, I will contact you via email about it!

[yannikschaelte]

warning added in #331