From 895f50ff6ff2711ff2197eb3bd3924c40ca395dd Mon Sep 17 00:00:00 2001
From: Roman Zeyde <roman.zeyde@gmail.com>
Date: Sat, 12 Jul 2014 10:02:12 +0300
Subject: [PATCH] fix training scaling factor

---
 common.py |  6 +++---
 recv.py   | 11 +++++------
 2 files changed, 8 insertions(+), 9 deletions(-)

diff --git a/common.py b/common.py
index a781e39..50a6aa3 100644
--- a/common.py
+++ b/common.py
@@ -7,14 +7,14 @@ log = logging.getLogger(__name__)
 Fs = 32e3
 Ts = 1.0 / Fs
 
-frequencies = (1 + np.arange(1)) * 1e3
+frequencies = (1 + np.arange(2)) * 1e3
 carrier_index = 0
 Fc = frequencies[carrier_index]
 Tc = 1.0 / Fc
 
 symbols = np.array([complex(x, y)
-                   for x in np.linspace(-1, 1, 4)
-                   for y in np.linspace(-1, 1, 4)]) / np.sqrt(2)
+                   for x in np.linspace(-1, 1, 8)
+                   for y in np.linspace(-1, 1, 8)]) / np.sqrt(2)
 
 Tsym = 1e-3
 Nsym = int(Tsym / Ts)
diff --git a/recv.py b/recv.py
index dd58214..15fac2b 100755
--- a/recv.py
+++ b/recv.py
@@ -78,20 +78,19 @@ def receive_prefix(symbols):
 
 def train_receiver(symbols, freqs):
     filters = {}
-    full_scale = len(freqs)
+    scaling_factor = len(freqs)  # to avoid saturation
     training = np.array(train.equalizer)
-    expected = full_scale * training
     if pylab:
         pylab.figure()
 
     for i, freq in enumerate(freqs):
-        S = take(symbols, i, len(expected))
+        S = take(symbols, i, len(training))
 
-        filt = sigproc.train(S, expected)
+        filt = sigproc.train(S, training * scaling_factor)
         filters[freq] = filt
 
         S = list(filt(S))
-        y = np.array(S)
+        y = np.array(S) / scaling_factor
         if pylab:
             pylab.subplot(HEIGHT, WIDTH, i+1)
             show.constellation(y, 'Train: $F_c = {}Hz$'.format(freq))
@@ -100,7 +99,7 @@ def train_receiver(symbols, freqs):
         if not all(train_result == training):
             raise ValueError('#{} training failed on {} Hz'.format(i, freq))
 
-        noise = y - expected
+        noise = y - training
         Pnoise = sigproc.power(noise)
         log.info('%10.1f kHz: Noise sigma=%.4f, SNR=%.1f dB',
                  freq/1e3, Pnoise**0.5, 10*np.log10(1/Pnoise))