aboutsummaryrefslogtreecommitdiffstats
path: root/dpd/src/Model.py
diff options
context:
space:
mode:
Diffstat (limited to 'dpd/src/Model.py')
-rw-r--r--dpd/src/Model.py481
1 files changed, 248 insertions, 233 deletions
diff --git a/dpd/src/Model.py b/dpd/src/Model.py
index 2aa9feb..827027a 100644
--- a/dpd/src/Model.py
+++ b/dpd/src/Model.py
@@ -8,262 +8,221 @@
import datetime
import os
import logging
+
logging_path = os.path.dirname(logging.getLoggerClass().root.handlers[0].baseFilename)
import numpy as np
-import matplotlib
-matplotlib.use('agg')
import matplotlib.pyplot as plt
-from sklearn.linear_model import Ridge
+from sklearn import linear_model
class Model:
"""Calculates new coefficients using the measurement and the old
coefficients"""
- def __init__(self, coefs_am, coefs_pm):
+ def __init__(self,
+ c,
+ SA,
+ MER,
+ coefs_am,
+ coefs_pm,
+ learning_rate_am=1.,
+ learning_rate_pm=1.,
+ plot=False):
+ self.c = c
+ self.SA = SA
+ self.MER = MER
+
+ self.learning_rate_am = learning_rate_am
+ self.learning_rate_pm = learning_rate_pm
+
self.coefs_am = coefs_am
- self.coefs_history = [coefs_am, ]
- self.mses = [0, ]
- self.errs = [0, ]
+ self.coefs_am_history = [coefs_am, ]
+ self.mses_am = []
+ self.errs_am = []
+
+ self.tx_mers = []
+ self.rx_mers = []
self.coefs_pm = coefs_pm
self.coefs_pm_history = [coefs_pm, ]
- self.errs_phase = [0, ]
+ self.errs_pm = []
+
+ self.plot = plot
- def sample_uniformly(self, txframe_aligned, rxframe_aligned, n_bins=4):
+ def sample_uniformly(self, tx_dpd, rx_received, n_bins=5):
"""This function returns tx and rx samples in a way
that the tx amplitudes have an approximate uniform
- distribution with respect to the txframe_aligned amplitudes"""
- txframe_aligned_abs = np.abs(txframe_aligned)
+ distribution with respect to the tx_dpd amplitudes"""
+ mask = np.logical_and((np.abs(tx_dpd) > 0.01), (np.abs(rx_received) > 0.01))
+ tx_dpd = tx_dpd[mask]
+ rx_received = rx_received[mask]
+
+ txframe_aligned_abs = np.abs(tx_dpd)
ccdf_min = 0
ccdf_max = np.max(txframe_aligned_abs)
tx_hist, ccdf_edges = np.histogram(txframe_aligned_abs,
bins=n_bins,
range=(ccdf_min, ccdf_max))
- n_choise = np.min(tx_hist)
+ n_choise = np.min(tx_hist)
tx_choice = np.zeros(n_choise * n_bins, dtype=np.complex64)
rx_choice = np.zeros(n_choise * n_bins, dtype=np.complex64)
for idx, bin in enumerate(tx_hist):
indices = np.where((txframe_aligned_abs >= ccdf_edges[idx]) &
- (txframe_aligned_abs <= ccdf_edges[idx+1]))[0]
- indices_choise = np.random.choice(indices, n_choise, replace=False)
- rx_choice[idx*n_choise:(idx+1)*n_choise] = rxframe_aligned[indices_choise]
- tx_choice[idx*n_choise:(idx+1)*n_choise] = txframe_aligned[indices_choise]
+ (txframe_aligned_abs <= ccdf_edges[idx + 1]))[0]
+ indices_choise = np.random.choice(indices,
+ n_choise,
+ replace=False)
+ rx_choice[idx * n_choise:(idx + 1) * n_choise] = \
+ rx_received[indices_choise]
+ tx_choice[idx * n_choise:(idx + 1) * n_choise] = \
+ tx_dpd[indices_choise]
+
+ assert isinstance(rx_choice[0], np.complex64), \
+ "rx_choice is not complex64 but {}".format(rx_choice[0].dtype)
+ assert isinstance(tx_choice[0], np.complex64), \
+ "tx_choice is not complex64 but {}".format(tx_choice[0].dtype)
+
return tx_choice, rx_choice
- def get_next_coefs(self, txframe_aligned, rxframe_aligned):
- tx_choice, rx_choice = self.sample_uniformly(txframe_aligned, rxframe_aligned)
-
- # Calculate new coefficients for AM/AM correction
- rx_abs = np.abs(rx_choice)
- rx_A = np.vstack([rx_abs,
- rx_abs ** 3,
- rx_abs ** 5,
- rx_abs ** 7,
- rx_abs ** 9,
- ]).T
- rx_dpd = np.sum(rx_A * self.coefs_am, axis=1)
- rx_dpd = rx_dpd * (
- np.median(np.abs(tx_choice)) / np.median(np.abs(rx_dpd)))
-
- err = rx_dpd - np.abs(tx_choice)
- self.errs.append(np.mean(np.abs(err ** 2)))
-
- a_delta = np.linalg.lstsq(rx_A, err)[0]
- new_coefs = self.coefs_am - 0.1 * a_delta
- new_coefs = new_coefs * (self.coefs_am[0] / new_coefs[0])
- logging.debug("a_delta {}".format(a_delta))
- logging.debug("new coefs_am {}".format(new_coefs))
-
- # Calculate new coefficients for AM/PM correction
- phase_diff_rad = ((
- (np.angle(tx_choice) -
- np.angle(rx_choice) +
- np.pi) % (2 * np.pi)) -
- np.pi
- )
-
- tx_abs = np.abs(tx_choice)
- tx_abs_A = np.vstack([tx_abs,
- tx_abs ** 2,
- tx_abs ** 3,
- tx_abs ** 4,
- tx_abs ** 5,
+ def dpd_amplitude(self, sig, coefs=None):
+ if coefs is None:
+ coefs = self.coefs_am
+ assert isinstance(sig[0], np.complex64), "Sig is not complex64 but {}".format(sig[0].dtype)
+ sig_abs = np.abs(sig)
+ A_sig = np.vstack([np.ones(sig_abs.shape),
+ sig_abs ** 1,
+ sig_abs ** 2,
+ sig_abs ** 3,
+ sig_abs ** 4,
+ ]).T
+ sig_dpd = sig * np.sum(A_sig * coefs, axis=1)
+ return sig_dpd, A_sig
+
+ def dpd_phase(self, sig, coefs=None):
+ if coefs is None:
+ coefs = self.coefs_pm
+ assert isinstance(sig[0], np.complex64), "Sig is not complex64 but {}".format(sig[0].dtype)
+ sig_abs = np.abs(sig)
+ A_phase = np.vstack([np.ones(sig_abs.shape),
+ sig_abs ** 1,
+ sig_abs ** 2,
+ sig_abs ** 3,
+ sig_abs ** 4,
]).T
- phase_dpd = np.sum(tx_abs_A * self.coefs_pm, axis=1)
-
- err_phase = phase_dpd - phase_diff_rad
- self.errs_phase.append(np.mean(np.abs(err_phase ** 2)))
- a_delta = np.linalg.lstsq(tx_abs_A, err_phase)[0]
- new_coefs_pm = self.coefs_pm - 0.1 * a_delta
- logging.debug("a_delta {}".format(a_delta))
- logging.debug("new new_coefs_pm {}".format(new_coefs_pm))
-
- def dpd_phase(tx):
- tx_abs = np.abs(tx)
- tx_A_complex = np.vstack([tx,
- tx * tx_abs ** 1,
- tx * tx_abs ** 2,
- tx * tx_abs ** 3,
- tx * tx_abs ** 4,
- ]).T
- tx_dpd = np.sum(tx_A_complex * self.coefs_pm, axis=1)
- return tx_dpd
-
- tx_range = np.linspace(0, 2)
- phase_range_dpd = dpd_phase(tx_range)
-
- rx_A_complex = np.vstack([rx_choice,
- rx_choice * rx_abs ** 2,
- rx_choice * rx_abs ** 4,
- rx_choice * rx_abs ** 6,
- rx_choice * rx_abs ** 8,
- ]).T
- rx_post_distored = np.sum(rx_A_complex * self.coefs_am, axis=1)
- rx_post_distored = rx_post_distored * (
- np.median(np.abs(tx_choice)) /
- np.median(np.abs(rx_post_distored)))
- mse = np.mean(np.abs((tx_choice - rx_post_distored) ** 2))
- logging.debug("MSE: {}".format(mse))
- self.mses.append(mse)
-
- def dpd(tx):
- tx_abs = np.abs(tx)
- tx_A_complex = np.vstack([tx,
- tx * tx_abs ** 2,
- tx * tx_abs ** 4,
- tx * tx_abs ** 6,
- tx * tx_abs ** 8,
- ]).T
- tx_dpd = np.sum(tx_A_complex * self.coefs_am, axis=1)
- return tx_dpd
-
- rx_range = np.linspace(0, 1, num=100)
- rx_range_dpd = dpd(rx_range)
- rx_range = rx_range[(rx_range_dpd > 0) & (rx_range_dpd < 2)]
- rx_range_dpd = rx_range_dpd[(rx_range_dpd > 0) & (rx_range_dpd < 2)]
-
- if logging.getLogger().getEffectiveLevel() == logging.DEBUG:
- logging.debug("txframe: min %f, max %f, median %f" %
- (np.min(np.abs(txframe_aligned)),
- np.max(np.abs(txframe_aligned)),
- np.median(np.abs(txframe_aligned))
- ))
-
- logging.debug("rxframe: min %f, max %f, median %f" %
- (np.min(np.abs(rx_choice)),
- np.max(np.abs(rx_choice)),
- np.median(np.abs(rx_choice))
- ))
+ phase_diff_est = np.sum(A_phase * coefs, axis=1)
+ return phase_diff_est, A_phase
+ def _next_am_coefficent(self, tx_choice, rx_choice):
+ """Calculate new coefficients for AM/AM correction"""
+ rx_dpd, rx_A = self.dpd_amplitude(rx_choice)
+ rx_dpd = rx_dpd * (
+ np.median(np.abs(tx_choice)) /
+ np.median(np.abs(rx_dpd)))
+ err = np.abs(rx_dpd) - np.abs(tx_choice)
+ mse = np.mean(np.abs((rx_dpd - tx_choice) ** 2))
+ self.mses_am.append(mse)
+ self.errs_am.append(np.mean(err**2))
+
+ reg = linear_model.Ridge(alpha=0.00001)
+ reg.fit(rx_A, err)
+ a_delta = reg.coef_
+ new_coefs_am = self.coefs_am - self.learning_rate_am * a_delta
+ new_coefs_am = new_coefs_am * (self.coefs_am[0] / new_coefs_am[0])
+ return new_coefs_am
+
+ def _next_pm_coefficent(self, tx_choice, rx_choice):
+ """Calculate new coefficients for AM/PM correction
+ Assuming deviations smaller than pi/2"""
+ phase_diff_choice = np.angle(
+ (rx_choice * tx_choice.conjugate()) /
+ (np.abs(rx_choice) * np.abs(tx_choice))
+ )
+ plt.hist(phase_diff_choice)
+ plt.savefig('/tmp/hist_' + str(np.random.randint(0,1000)) + '.svg')
+ plt.clf()
+ phase_diff_est, phase_A = self.dpd_phase(rx_choice)
+ err_phase = phase_diff_est - phase_diff_choice
+ self.errs_pm.append(np.mean(np.abs(err_phase ** 2)))
+
+ reg = linear_model.Ridge(alpha=0.00001)
+ reg.fit(phase_A, err_phase)
+ p_delta = reg.coef_
+ new_coefs_pm = self.coefs_pm - self.learning_rate_pm * p_delta
+
+ return new_coefs_pm, phase_diff_choice
+
+ def get_next_coefs(self, tx_dpd, rx_received):
+ # Check data type
+ assert tx_dpd[0].dtype == np.complex64, \
+ "tx_dpd is not complex64 but {}".format(tx_dpd[0].dtype)
+ assert rx_received[0].dtype == np.complex64, \
+ "rx_received is not complex64 but {}".format(rx_received[0].dtype)
+ # Check if signals have same normalization
+ normalization_error = np.abs(np.median(np.abs(tx_dpd)) -
+ np.median(np.abs(rx_received))) / (
+ np.median(np.abs(tx_dpd)) + np.median(np.abs(rx_received)))
+ assert normalization_error < 0.01, "Non normalized signals"
+
+ tx_choice, rx_choice = self.sample_uniformly(tx_dpd, rx_received)
+ new_coefs_am = self._next_am_coefficent(tx_choice, rx_choice)
+ new_coefs_pm, phase_diff_choice = self._next_pm_coefficent(tx_choice, rx_choice)
+
+ logging.debug('txframe: min {:.2f}, max {:.2f}, ' \
+ 'median {:.2f}; rxframe: min {:.2f}, max {:.2f}, ' \
+ 'median {:.2f}; new coefs_am {};' \
+ 'new_coefs_pm {}'.format(
+ np.min(np.abs(tx_dpd)),
+ np.max(np.abs(tx_dpd)),
+ np.median(np.abs(tx_dpd)),
+ np.min(np.abs(rx_choice)),
+ np.max(np.abs(rx_choice)),
+ np.median(np.abs(rx_choice)),
+ new_coefs_am,
+ new_coefs_pm))
+
+ if logging.getLogger().getEffectiveLevel() == logging.DEBUG and self.plot:
+ off = self.SA.calc_offset(tx_dpd)
+ tx_mer = self.MER.calc_mer(tx_dpd[off:off + self.c.T_U])
+ rx_mer = self.MER.calc_mer(rx_received[off:off + self.c.T_U], debug=True)
+ self.tx_mers.append(tx_mer)
+ self.rx_mers.append(rx_mer)
+
+ if logging.getLogger().getEffectiveLevel() == logging.DEBUG and self.plot:
dt = datetime.datetime.now().isoformat()
- fig_path = logging_path + "/" + dt + "_Model.pdf"
+ fig_path = logging_path + "/" + dt + "_Model.svg"
- fig = plt.figure(figsize=(3*6, 1.5 * 6))
+ fig = plt.figure(figsize=(2 * 6, 2 * 6))
- ax = plt.subplot(3,3,1)
- ax.plot(np.abs(txframe_aligned[:128]),
- label="TX sent",
- linestyle=":")
- ax.plot(np.abs(rxframe_aligned[:128]),
- label="RX received",
- color="red")
- ax.set_title("Synchronized Signals of Iteration {}".format(len(self.coefs_history)))
- ax.set_xlabel("Samples")
- ax.set_ylabel("Amplitude")
- ax.text(0, 0, "TX (max {:01.3f}, mean {:01.3f}, median {:01.3f})".format(
- np.max(np.abs(txframe_aligned)),
- np.mean(np.abs(txframe_aligned)),
- np.median(np.abs(txframe_aligned))
- ), size = 8)
- ax.legend(loc=4)
+ i_sub = 1
- ax = plt.subplot(3,3,2)
- ax.plot(np.real(txframe_aligned[:128]),
+ ax = plt.subplot(4, 2, i_sub)
+ i_sub += 1
+ ax.plot(np.abs(tx_dpd[:128]),
label="TX sent",
linestyle=":")
- ax.plot(np.real(rxframe_aligned[:128]),
+ ax.plot(np.abs(rx_received[:128]),
label="RX received",
color="red")
- ax.set_title("Synchronized Signals")
- ax.set_xlabel("Samples")
- ax.set_ylabel("Real Part")
- ax.legend(loc=4)
-
- ax = plt.subplot(3,3,3)
- ax.plot(np.abs(txframe_aligned[:128]),
- label="TX Frame",
- linestyle=":",
- linewidth=0.5)
- ax.plot(np.abs(rxframe_aligned[:128]),
- label="RX Frame",
- linestyle="--",
- linewidth=0.5)
-
- rx_abs = np.abs(rxframe_aligned)
- rx_A = np.vstack([rx_abs,
- rx_abs ** 3,
- rx_abs ** 5,
- rx_abs ** 7,
- rx_abs ** 9,
- ]).T
- rx_dpd = np.sum(rx_A * self.coefs_am, axis=1)
- rx_dpd = rx_dpd * (
- np.median(np.abs(tx_choice)) / np.median(np.abs(rx_dpd)))
-
- ax.plot(np.abs(rx_dpd[:128]),
- label="RX DPD Frame",
- linestyle="-.",
- linewidth=0.5)
-
- tx_abs = np.abs(np.abs(txframe_aligned[:128]))
- tx_A = np.vstack([tx_abs,
- tx_abs ** 3,
- tx_abs ** 5,
- tx_abs ** 7,
- tx_abs ** 9,
- ]).T
- tx_dpd = np.sum(tx_A * new_coefs, axis=1)
- tx_dpd_norm = tx_dpd * (
- np.median(np.abs(tx_choice)) / np.median(np.abs(tx_dpd)))
-
- ax.plot(np.abs(tx_dpd_norm[:128]),
- label="TX DPD Frame Norm",
- linestyle="-.",
- linewidth=0.5)
- ax.legend(loc=4)
- ax.set_title("RX DPD")
+ ax.set_title("Synchronized Signals of Iteration {}"
+ .format(len(self.coefs_am_history)))
ax.set_xlabel("Samples")
ax.set_ylabel("Amplitude")
+ ax.text(0, 0, "TX (max {:01.3f}, mean {:01.3f}, " \
+ "median {:01.3f})".format(
+ np.max(np.abs(tx_dpd)),
+ np.mean(np.abs(tx_dpd)),
+ np.median(np.abs(tx_dpd))
+ ), size=8)
+ ax.legend(loc=4)
- ax = plt.subplot(3,3,4)
- ax.scatter(
- np.abs(tx_choice[:1024]),
- np.abs(rx_choice[:1024]),
- s=0.1)
- ax.plot(rx_range_dpd / self.coefs_am[0], rx_range, linewidth=0.25)
- ax.set_title("Amplifier Characteristic")
- ax.set_xlabel("TX Amplitude")
- ax.set_ylabel("RX Amplitude")
-
- ax = plt.subplot(3,3,5)
- ax.scatter(
- np.abs(tx_choice[:1024]),
- phase_diff_rad[:1024] * 180 / np.pi,
- s=0.1
- )
- ax.plot(tx_range, phase_range_dpd * 180 / np.pi, linewidth=0.25)
- ax.set_title("Amplifier Characteristic")
- ax.set_xlabel("TX Amplitude")
- ax.set_ylabel("Phase Difference [deg]")
-
- ax = plt.subplot(3,3,6)
+ ax = plt.subplot(4, 2, i_sub)
+ i_sub += 1
ccdf_min, ccdf_max = 0, 1
- tx_hist, ccdf_edges = np.histogram(np.abs(txframe_aligned),
- bins=60,
- range=(ccdf_min, ccdf_max))
- tx_hist_normalized = tx_hist.astype(float)/np.sum(tx_hist)
+ tx_hist, ccdf_edges = np.histogram(np.abs(tx_dpd),
+ bins=60,
+ range=(ccdf_min, ccdf_max))
+ tx_hist_normalized = tx_hist.astype(float) / np.sum(tx_hist)
ccdf = 1.0 - np.cumsum(tx_hist_normalized)
ax.semilogy(ccdf_edges[:-1], ccdf, label="CCDF")
ax.semilogy(ccdf_edges[:-1],
@@ -271,14 +230,51 @@ class Model:
label="Histogram",
drawstyle='steps')
ax.legend(loc=4)
- ax.set_ylim(1e-5,2)
+ ax.set_ylim(1e-5, 2)
ax.set_title("Complementary Cumulative Distribution Function")
ax.set_xlabel("TX Amplitude")
ax.set_ylabel("Ratio of Samples larger than x")
- ax = plt.subplot(3,3,7)
- coefs_history = np.array(self.coefs_history)
- for idx, coef_hist in enumerate(coefs_history.T):
+ ax = plt.subplot(4, 2, i_sub)
+ i_sub += 1
+ ax.semilogy(np.array(self.mses_am) + 1e-10, label="log(MSE)")
+ ax.semilogy(np.array(self.errs_am) + 1e-10, label="log(ERR)")
+ ax.legend(loc=4)
+ ax.set_title("MSE History")
+ ax.set_xlabel("Iterations")
+ ax.set_ylabel("MSE")
+
+ ax = plt.subplot(4, 2, i_sub)
+ i_sub += 1
+ ax.plot(self.tx_mers, label="TX MER")
+ ax.plot(self.rx_mers, label="RX MER")
+ ax.legend(loc=4)
+ ax.set_title("MER History")
+ ax.set_xlabel("Iterations")
+ ax.set_ylabel("MER")
+
+ ax = plt.subplot(4, 2, i_sub)
+ rx_range = np.linspace(0, 1, num=100, dtype=np.complex64)
+ rx_range_dpd = self.dpd_amplitude(rx_range)[0]
+ rx_range_dpd_new = self.dpd_amplitude(rx_range, new_coefs_am)[0]
+ i_sub += 1
+ ax.scatter(
+ np.abs(tx_choice),
+ np.abs(rx_choice),
+ s=0.1)
+ ax.plot(rx_range_dpd / self.coefs_am[0], rx_range, linewidth=0.25, label="current")
+ ax.plot(rx_range_dpd_new / self.coefs_am[0], rx_range, linewidth=0.25, label="next")
+ ax.set_ylim(0, 1)
+ ax.set_xlim(0, 1)
+ ax.legend()
+ ax.set_title("Amplifier Characteristic")
+ ax.set_xlabel("TX Amplitude")
+ ax.set_ylabel("RX Amplitude")
+
+ ax = plt.subplot(4, 2, i_sub)
+ i_sub += 1
+ coefs_am_history = np.array(self.coefs_am_history)
+ for idx, coef_hist in enumerate(coefs_am_history.T):
ax.plot(coef_hist,
label="Coef {}".format(idx),
linewidth=0.5)
@@ -287,10 +283,38 @@ class Model:
ax.set_xlabel("Iterations")
ax.set_ylabel("Coefficient Value")
- ax = plt.subplot(3,3,8)
- coefs_history = np.array(self.coefs_pm_history)
- for idx, coef_hist in enumerate(coefs_history.T):
- ax.plot(coef_hist,
+ phase_range = np.linspace(0, 1, num=100, dtype=np.complex64)
+ phase_range_dpd = self.dpd_phase(phase_range)[0]
+ phase_range_dpd_new = self.dpd_phase(phase_range,
+ coefs=new_coefs_pm)[0]
+ ax = plt.subplot(4, 2, i_sub)
+ i_sub += 1
+ ax.scatter(
+ np.abs(tx_choice),
+ np.rad2deg(phase_diff_choice),
+ s=0.1)
+ ax.plot(
+ np.abs(phase_range),
+ np.rad2deg(phase_range_dpd),
+ linewidth=0.25,
+ label="current")
+ ax.plot(
+ np.abs(phase_range),
+ np.rad2deg(phase_range_dpd_new),
+ linewidth=0.25,
+ label="next")
+ ax.set_ylim(-60, 60)
+ ax.set_xlim(0, 1)
+ ax.legend()
+ ax.set_title("Amplifier Characteristic")
+ ax.set_xlabel("TX Amplitude")
+ ax.set_ylabel("Phase Difference")
+
+ ax = plt.subplot(4, 2, i_sub)
+ i_sub += 1
+ coefs_pm_history = np.array(self.coefs_pm_history)
+ for idx, coef_phase_hist in enumerate(coefs_pm_history.T):
+ ax.plot(coef_phase_hist,
label="Coef {}".format(idx),
linewidth=0.5)
ax.legend(loc=4)
@@ -298,21 +322,12 @@ class Model:
ax.set_xlabel("Iterations")
ax.set_ylabel("Coefficient Value")
- ax = plt.subplot(3,3,9)
- coefs_history = np.array(self.coefs_history)
- ax.plot(self.mses, label="MSE")
- ax.plot(self.errs, label="ERR")
- ax.legend(loc=4)
- ax.set_title("MSE History")
- ax.set_xlabel("Iterations")
- ax.set_ylabel("MSE")
-
fig.tight_layout()
fig.savefig(fig_path)
fig.clf()
- self.coefs_am = new_coefs
- self.coefs_history.append(self.coefs_am)
+ self.coefs_am = new_coefs_am
+ self.coefs_am_history.append(self.coefs_am)
self.coefs_pm = new_coefs_pm
self.coefs_pm_history.append(self.coefs_pm)
return self.coefs_am, self.coefs_pm