Unable to run parallel backtest with neural network models

[tinkoff-ai]

Unable to run parallel backtest with neural network models

[julia-shenshina]

how can we fix it?

[martins0n]

We add joblib parameters customization via additional parameter joblib_params in #253 for pipeline.backtest

tscv_deepar.backtest(ts, n_jobs=2, metrics=[SMAPE()], n_folds=2, joblib_params=dict(verbose=11, backend="loky", mmap_mode="c"))

Now by default we use multiprocessing backend for joblib because of global nature of our tslogger.
You can use loky or multiprocessing with multiprocessing.get_start_method() == 'spawn' but results of folds runs are not be logged -- only total results of backtest will be logged (It may be important for experiments tracked by wandb)

To sum up:

So in case of working with cuda, you can use parallelization of backtest with loky specification or multiprocessing with multiprocessing.set_start_method('spawn') specification.

For now we can close the issue.

You can open it if additional recommendation are needed or you have new proposal.