mirror of
https://github.com/romanz/amodem.git
synced 2026-02-07 18:08:03 +08:00
74 lines
2.0 KiB
Python
74 lines
2.0 KiB
Python
import numpy as np
|
|
from numpy import linalg
|
|
|
|
import common
|
|
|
|
def lfilter(b, a, x):
|
|
b = np.array(b) / a[0]
|
|
a = np.array(a[1:]) / a[0]
|
|
|
|
x_ = [0] * len(b)
|
|
y_ = [0] * len(a)
|
|
for v in x:
|
|
x_ = [v] + x_[:-1]
|
|
u = np.dot(x_, b)
|
|
u = u - np.dot(y_, a)
|
|
|
|
y_ = [u] + y_[1:]
|
|
yield u
|
|
|
|
def train(S, training):
|
|
A = np.array([ S[1:], S[:-1], training[:-1] ]).T
|
|
b = training[1:]
|
|
b0, b1, a1 = linalg.lstsq(A, b)[0]
|
|
return lambda x: lfilter(b=[b0, b1], a=[1, -a1], x=x)
|
|
|
|
class QAM(object):
|
|
def __init__(self, bits_per_symbol, radii):
|
|
self._enc = {}
|
|
index = 0
|
|
N = (2 ** bits_per_symbol) / len(radii)
|
|
for a in radii:
|
|
for i in range(N):
|
|
k = tuple(int(index & (1 << j) != 0) for j in range(bits_per_symbol))
|
|
v = np.exp(2j * i * np.pi / N)
|
|
self._enc[k] = v * a
|
|
index += 1
|
|
self._dec = {v: k for k, v in self._enc.items()}
|
|
self.points = self._enc.values()
|
|
self.bits_per_symbol = bits_per_symbol
|
|
|
|
def encode(self, bits):
|
|
trailing_bits = len(bits) % self.bits_per_symbol
|
|
if trailing_bits:
|
|
bits = bits + [0] * (self.bits_per_symbol - trailing_bits)
|
|
for i in range(0, len(bits), self.bits_per_symbol):
|
|
s = self._enc[ tuple(bits[i:i+self.bits_per_symbol]) ]
|
|
yield s
|
|
|
|
def decode(self, symbols):
|
|
keys = np.array(self._dec.keys())
|
|
for s in symbols:
|
|
index = np.argmin(np.abs(s - keys))
|
|
yield self._dec[ keys[index] ]
|
|
|
|
modulator = QAM(bits_per_symbol=2, radii=[1.0])
|
|
|
|
def clip(x, lims):
|
|
return min(max(x, lims[0]), lims[1])
|
|
|
|
def power(x):
|
|
return np.dot(x.conj(), x).real / len(x)
|
|
|
|
def exp_iwt(freq, n):
|
|
iwt = 2j * np.pi * freq * np.arange(n) * common.Ts
|
|
return np.exp(iwt)
|
|
|
|
def norm(x):
|
|
return np.sqrt(np.dot(x.conj(), x).real)
|
|
|
|
def coherence(x, freq):
|
|
n = len(x)
|
|
Hc = exp_iwt(-freq, n) / np.sqrt(0.5*n)
|
|
return np.dot(Hc, x) / norm(x)
|