# esn = ESN(2, 1, 500, echo_param=0.85, regulariser=1e-6)
esn = EsnModel(2, 1, 100)
# esn = EsnModel(2, 1, 1000,
# spectral_scale=1.0, echo_param=0.85,
# input_weight_scale=1.0, regulariser=1e-5)
esn.train(X_train, y_train)
hm = np.zeros((100, 100))
n, m = hm.shape
for i in range(n):
print("iteration %d" % (i+1))
for j in range(m):
save = esn.layers[0].W_res[i, j]
esn.layers[0].W_res[i, j] = 0
esn_outputs = esn.generate(X_valid, len(y_valid))
error = nrmse(y_valid, esn_outputs, data_mean)
hm[i, j] = error
esn.layers[0].W_res[i, j] = save
del esn_outputs
# esn_outputs = []
# generate test-data
# esn_outputs = esn.generate(X_valid, len(y_valid))
# n_esn_outputs.append(esn_outputs)
# error = nrmse(y_valid, esn_outputs, data_mean)
# print('ESN NRMSE: %f' % error)
# for i in range(len(y_valid)):
# y_valid_slice = y_valid[:i]
esn_outputs = []
# esn.layers[0].drop_probability = 0.0
# generate test-data
#esn_outputs = esn.generate(X_valid, len(y_valid))
#error_first = nrmse(y_valid, esn_outputs, data_mean)
#print('ESN NRMSE: %f' % error_first)
gen_data = np.vstack((X_train[-warm_up_length:], X_valid))
esn_outputs = esn.generate(
gen_data, len(y_valid),
warmup_timesteps=warm_up_length) #, reset_increment=3)
#esn_outputs = esn.generate(gen_data, len(y_valid), warmup_timesteps=warm_up_length)
n_esn_outputs.append(esn_outputs)
error = nrmse(y_valid, esn_outputs, data_mean)
print('ESN NRMSE: %f' % error)
#assert np.round(error, 3) == np.round(error_first, 3), "errors after warm-up and generation must be the same"
for i in range(len(y_valid)):
y_valid_slice = y_valid[:i]
esn_outputs_slice = esn_outputs[:i]
err = nrmse(y_valid_slice, esn_outputs_slice, data_mean)
t_nrmse[n, i] = err
_nrmse[n, r] = error
print("--ERR. MEAN: {}+-{}".format(np.mean(_nrmse, 1),
np.std(_nrmse, 1)))
X_valid = np.hstack(
(np.array(data[split - 1:-1]), np.ones_like(data[split - 1:-1, :1])))
y_valid = np.array(data[split:])
'''
Train the ESN.
'''
esn = models.EsnModel(input_size=2, output_size=1, reservoir_size=1000)
esn.train(X_train, y_train, warmup_timesteps=300)
'''
Evaluate the ESN.
'''
# generate test-data starting from the next data point.
esn_outputs = esn.generate(X_valid, len(y_valid))
error_gen = nrmse(y_valid, esn_outputs, DATA_MEAN)
print('ESN NRMSE VALID 1: %f' % error_gen)
# generate test data from earlier data requires using a warmup of 300
# so that our reservoir state is in synch with the predicted data point.
warmup = 300
gen_data = np.vstack((X_train[-warmup:], X_valid))
esn_outputs = esn.generate(gen_data, len(y_valid), warmup_timesteps=warmup)
error_pre_gen = nrmse(y_valid, esn_outputs, DATA_MEAN)
print('ESN NRMSE VALID 2: %f' % error_pre_gen)
# notice how both methods of training produce the same predictions. Try changing
# the 'warmup' to a low value like zero and see if they are still the same.
assert np.round(error_gen, 3) == np.round(
error_pre_gen,