def run_episode(self, params, num_runs=3):
esn = self.params_to_model(params)
errors = np.zeros(num_runs)
for run_num in range(num_runs):
esn.train(self.data_train[0], self.data_train[1])
# Run generative test, calculate NRMSE
y_pred = []
u_n = self.data_val[0][0]
for _ in range(len(self.data_val[1])):
u_n = esn.forward(u_n)
y_pred.append(u_n)
y_pred = np.array(y_pred).squeeze()
y_vals = self.data_val[1].squeeze()
# print(np.hstack((y_pred[:, None], y_vals[:, None])))
nr = nrmse(y_vals, y_pred, self.MEAN_OF_DATA)
if nr > 1000000000: nr = 1000000000
errors[run_num] = nr
# for r in esn.reservoirs:
# print("E:{}, S:{}, W:{}, REG:{}, SIZE:{}, SPARSE:{}".format(r.echo_param, r.spectral_scale, r.input_weights_scale, esn.regulariser, r.N, r.sparsity))
# plt.plot(range(len(y_pred)), y_pred, label="pred")
# plt.plot(range(len(y_vals)), y_vals, label="true")
# plt.legend()
# plt.show()
return -np.mean(np.log(errors)), np.std(np.log(errors))
def run_episode2(self, params):
esn = self.params_to_model(params)
esn.train(self.data_train[0], self.data_train[1])
# run generative and check
y_pred = []
# GENERATIVE =================================================
u_n_ESN = self.data_val[0][0]
for _ in range(len(self.data_val[1])):
u_n_ESN = esn.forward(u_n_ESN)
y_pred.append(u_n_ESN)
y_pred = np.array(y_pred).squeeze()
y_vals = self.data_val[1].squeeze()
nrmse_err = nrmse(y_vals, y_pred, self.MEAN_OF_DATA)
# avoid explosions
if nrmse_err > 10:
nrmse_err = MAX_REWARD
return -nrmse_err
#esn_outputs = []
eesn_outputs = []
# GENERATIVE =================================================
#u_n_ESN = data[split]
u_n_EESN = np.array(X_valid[0])
for _ in range(len(data[split:])):
#u_n_ESN = esn.forward(u_n_ESN)
#esn_outputs.append(u_n_ESN)
u_n_EESN = eesn.forward(u_n_EESN)
eesn_outputs.append(u_n_EESN)
#esn_outputs = np.array(esn_outputs).squeeze()
eesn_outputs = np.array(eesn_outputs).squeeze()
error = nrmse(y_valid, eesn_outputs, data_mean)
#print(' ESN MSE: %f' % mse(y_valid, esn_outputs))
print('EESN NRMSE: %f' % error)
key = eesn.info()
if key not in to_save.keys():
to_save[key] = []
to_save[key].append(error)
if 0:
f, ax = plt.subplots(figsize=(12, 12))
#ax.plot(range(len(esn_outputs)), esn_outputs, label='ESN')
ax.plot(range(len(eesn_outputs)), eesn_outputs, label='EESN')
ax.plot(range(len(y_valid)), y_valid, label='True')
plt.legend()
plt.show()
# SUPERVISED ====
# for u_n in X_valid:
# esn_outputs.append(esn.forward(u_n))
# lcesn_outputs.append(lcesn.forward(u_n))
# ============================================================
# esn_outputs = np.array(esn_outputs).squeeze()
eesn_outputs = np.array(eesn_outputs).squeeze()
y_vals = y_valid.squeeze()
# print(np.shape(eesn_outputs))
# print(np.shape(y_vals))
# print(np.vstack((eesn_outputs, y_vals)))
# print(' ESN MSE: %f' % mse(y_valid, esn_outputs))
# print('EESN MSE: %f' % mse(y_valid, eesn_outputs))
nrmse_err = nrmse(y_vals, eesn_outputs, MEAN_OF_DATA)
if nrmse_err > 100000:
nrmse_err = 100000
print('DHESN NRMSE: %f' % nrmse_err)
#print('-log(NRMSE): %f' % -np.log(nrmse_err))
_errors.append(nrmse_err)
# plt.plot(range(len(eesn_outputs)), eesn_outputs, label="predicted")
# plt.plot(range(len(y_vals)), y_vals, label="true")
# plt.legend()
# plt.show()
# we = eesn.W_out.squeeze()
# plt.bar(range(len(we)), we)
# plt.show()
data_csv[idx, 0] = n
# for u_n in X_valid:
# esn_outputs.append(esn.forward(u_n))
# lcesn_outputs.append(lcesn.forward(u_n))
# ============================================================
# esn_outputs = np.array(esn_outputs).squeeze()
eesn_outputs = np.array(
eesn_outputs).squeeze()
y_vals = y_valid.squeeze()
# print(np.shape(eesn_outputs))
# print(np.shape(y_vals))
# print(np.vstack((eesn_outputs, y_vals)))
# print(' ESN MSE: %f' % mse(y_valid, esn_outputs))
# print('EESN MSE: %f' % mse(y_valid, eesn_outputs))
nrmse_err = nrmse(
y_vals, eesn_outputs, MEAN_OF_DATA)
if nrmse_err > 100000:
nrmse_err = 100000
print('DHESN NRMSE: %f' % nrmse_err)
#print('-log(NRMSE): %f' % -np.log(nrmse_err))
_errors.append(nrmse_err)
# plt.plot(range(len(eesn_outputs)), eesn_outputs, label="predicted")
# plt.plot(range(len(y_vals)), y_vals, label="true")
# plt.legend()
# plt.show()
# we = eesn.W_out.squeeze()
# plt.bar(range(len(we)), we)
# plt.show()
data_csv[idx, 0] = n