RuntimeError at training due to multiprocessing about reservoirpy

[nTrouvain]

Hello,

Could you try to import reservoirpy from within the 'if name == "main":' statement ?

PS: Thank you for using reservoirpy, we are glad you like it :)

[anaston]

Thank you for the quick reply:)

Yes, I tried adding it into an 'if name == "main":' statement but still getting this RuntimeError. However, what is very odd that actually the code does not stop at the error. Now I added some additional code after this error is printed in the terminal and that runs. Do you have an idea what's happenning? Sorry, I am new to Python, so I might miss something basic.

Pleas the script below that I use:

def split_timeserie_for_task1(forecast, train_length=20000):

X_train, y_train = X[:train_length], X[forecast: train_length+forecast]
X_test, y_test = X[train_length: -forecast], X[train_length+forecast:]

return (X_train, y_train), (X_test, y_test)

def reset_esn():
Win = mat_gen.generate_input_weights(units, 1, input_scaling=input_scaling,
proba=input_connectivity, input_bias=True,
seed=seed)

W = mat_gen.generate_internal_weights(units, sr=spectral_radius,
                              proba=density, seed=seed)

reservoir = ESN(leak_rate, W, Win, ridge=regularization)

return reservoir

def r2_score(y_true, y_pred):
return 1 - (np.sum((y_true - y_pred)**2) / np.sum((y_true - y_true.mean())**2))

def nrmse(y_true, y_pred):
return np.sqrt((np.sum(y_true - y_pred)**2) / len(y_true)) / (y_true.max() - y_true.min())

if name == 'main':
import numpy as np
import matplotlib.pyplot as plt
from reservoirpy import mat_gen, ESN
from reservoirpy.datasets import mackey_glass

# Generate data

timesteps = 25000
tau = 17
X = mackey_glass(timesteps, tau=tau)

# rescale between -1 and 1
X = 2 * (X - X.min()) / (X.max() - X.min()) - 1

forecast = 10
(X_train, y_train), (X_test, y_test) = split_timeserie_for_task1(forecast)

sample = 500
figsizex = 15
figsizey = 3

# %% ESN training and prediction

units = 100
leak_rate = 0.3
spectral_radius = 1.25
input_scaling = 1.0
density = 0.1
input_connectivity = 0.2
regularization = 1e-8
seed = 1234

Win = mat_gen.generate_input_weights(units, 1, input_scaling=input_scaling,
                                    proba=input_connectivity, input_bias=True,
                                    seed=seed)

W = mat_gen.generate_internal_weights(units, sr=spectral_radius,
                            proba=density, seed=seed)

reservoir = ESN(leak_rate, W, Win, ridge=regularization)

states = reservoir.train([X_train.reshape(-1, 1)], [y_train.reshape(-1, 1)],
                return_states=True, verbose=True) # workers=1

y_pred, states1 = reservoir.run([X_test.reshape(-1, 1)], return_states=True,
                verbose=True) # init_state=states[0][-1] workers=1

y_pred = y_pred[0].reshape(-1, 1)
states1 = states1[0]

fig = plt.figure(figsize=(figsizex, figsizey)) # figsize=(15, 7)
plt.plot(np.arange(sample), y_pred[:sample], lw=3, label="ESN prediction")
plt.plot(np.arange(sample), y_test[:sample], linestyle="--", lw=2, label="True value")
plt.plot(np.abs(y_test[:sample] - y_pred[:sample]), label="Absolute deviation")

plt.legend()
plt.show()

[nTrouvain]

Hello,
We have partially fixed the issue, but I am affraid it is caused by the behaviour of Mac OS regarding how to instanciate processes.
Using the if __name__ == "__main__": statement is mandatory in your case when launching your code (it is actually a good practice in Python in any case). Thank you for drawing this to our attention, we will try to document this error the best we can. In the meantime, next releases v0.2.4-post1 and v0.3.0b1 will avoid raising this error all the time, even when parallelization is deactivated. You can deactivate it by using "sequential" as a backend in the trainnig functions, or by setting it globaly using the set_joblib_backend("sequential") function.

[anaston]

Thanks a lot for trying to fix this and the help with the workaround. Feel free to let me know if I can be helpful in any testing.