Parallel Data Collection problem

[dxyzx0]

I want to reproduce the result in the paper(10000 for training, 20000 for validation) by running the example branching-imitaion.ipynb, but I find that the generation process is too slow to finish.
For example, this follows the setting in the paper for Capacitated Facility Location problem with n_customer=100 and I shrink the training/validation set to 12000, but still too slow to collect all the instances.

My cpu is Intel(R) Xeon(R) Gold 6226R CPU @ 2.90GHz, which I think is enough.

Users that are interested in reproducing competitive performance should use a larger sample size, such as the n=100,000 samples used for training in the paper. In this case, we strongly recommend to parallelize data collection, as in the original Gasse et al. (2019) code.

I know I need to parallelize data collection, but how can I accomplish this process?
I tried with multiprocessing.Pool but I get the following error TypeError: cannot pickle 'ecole.core.reward.IsDone' object.

[AntoinePrv]

Hi @DingXiangYuanZhiXing

Indeed multiprocessing requires pickle support, which is not fully supported by Ecole.
You can use the threading, which should not be (too) affected by the GIL, given that it is released during computationally intensive C/C++ functions calls.

[dxyzx0]

I solved this problem by putting Ecole.environment.Environment() object into function body instead of function arguments.

It turns out this method works and I haven't found any problem about it.