⚠️ WARNING: This page is obsolete
Articles typically receive this designation when the technology they describe is no longer relevant, code
provided is later deemed to be of poor quality, or the topics discussed are better presented in future
articles. Articles like this are retained for the sake of preservation, but their content should be
critically assessed.
I had TKinter all wrong! While my initial tests with PyGame’s rapid ability to render Numpy arrays in the form of pixel maps proved impressive, it was only because I was comparing it to poor TK code. I don’t know what I was doing wrong, but when I decided to give TKinter one more shot I was blown away – it’s as smooth or smoother as PyGame. Forget PyGame! I’m rendering everything in raw TK from now on. This utilizes the Python Imaging Library (PIL) so it’s EXTREMELY flexible (supports fancy operations, alpha channels, etc).
The screenshot shows me running the script (below) generating random noise and “scrolling” it horizontally (like my spectrograph software does) at a fast rate smoothly (almost 90 FPS!). Basically, it launches a window, creates a canvas widget (which I’m told is faster to update than a label and reduces flickering that’s often associated with rapid redraws because it uses double-buffering). Also, it uses threading to handle the calculations/redraws without lagging the GUI. The code speaks for itself.
import Tkinter
from PIL import Image, ImageTk
import numpy
import time
class mainWindow():
times = 1
timestart = time.clock()
data = numpy.array(numpy.random.random((400, 500))*100, dtype=int)
def __init__(self):
self.root = Tkinter.Tk()
self.frame = Tkinter.Frame(self.root, width=500, height=400)
self.frame.pack()
self.canvas = Tkinter.Canvas(self.frame, width=500, height=400)
self.canvas.place(x=-2, y=-2)
self.root.after(0, self.start) # INCREASE THE 0 TO SLOW IT DOWN
self.root.mainloop()
def start(self):
global data
self.im = Image.fromstring('L', (self.data.shape[1],
self.data.shape[0]), self.data.astype('b').tostring())
self.photo = ImageTk.PhotoImage(image=self.im)
self.canvas.create_image(0, 0, image=self.photo, anchor=Tkinter.NW)
self.root.update()
self.times += 1
if self.times % 33 == 0:
print "%.02f FPS" % (self.times/(time.clock()-self.timestart))
self.root.after(10, self.start)
self.data = numpy.roll(self.data, -1, 1)
if __name__ == '__main__':
x = mainWindow()
⚠️ WARNING: This page is obsolete
Articles typically receive this designation when the technology they describe is no longer relevant, code
provided is later deemed to be of poor quality, or the topics discussed are better presented in future
articles. Articles like this are retained for the sake of preservation, but their content should be
critically assessed.
⚠️ SEE UPDATED POST: Signal Filtering in Python
While continuing my quest into the world of linear data analysis and signal processing, I came to a point where I wanted to emphasize variations in FFT traces. While I am keeping my original data for scientific reference, visually I want to represent it emphasizing variations rather than concentrating on trends. I wrote a detrending function which I’m sure will be useful for many applications:
def detrend(data,degree=10):
detrended=[None]*degree
for i in range(degree,len(data)-degree):
chunk=data[i-degree:i+degree]
chunk=sum(chunk)/len(chunk)
detrended.append(data[i]-chunk)
return detrended+[None]*degree
However, this method is extremely slow. I need to think of a way to accomplish this same thing much faster. [ponders]
UPDATE: It looks like I’ve once again re-invented the wheel. All of this has been done already, and FAR more efficiently I might add. For more see scipy.signal.detrend.html
import scipy.signal
ffty=scipy.signal.detrend(ffty)
⚠️ WARNING: This page is obsolete
Articles typically receive this designation when the technology they describe is no longer relevant, code
provided is later deemed to be of poor quality, or the topics discussed are better presented in future
articles. Articles like this are retained for the sake of preservation, but their content should be
critically assessed.
I’m attempting to thoroughly re-write the data assessment portions of my QRSS VD software, and rather than rushing to code it (like I did last time) I’m working hard on every step trying to optimize the code. I came across some notes I made about Fast Fourier Transformations from the first time I coded the software, and though I’d post some code I found helpful. Of particular satisfaction is an email I received from Alberto, I2PHD, the creator of Argo (the “gold standard” QRSS spectrograph software for Windows). In it he notes:
I think that [it is a mistake to] throw away the imaginary part of the FFT. What I do in Argo, in Spectran, in Winrad, in SDRadio and in all of my other programs is compute the magnitude of the [FFT] signal, then compute the logarithm of it, and only then I do a mapping of the colors on the screen with the result of this last computation.
Alberto, I2PHD (the creator of Argo)
UPDATE IN SEPTEMBER, 2020 (10 years later): I now understand that magnitude = sqrt(real^2 + imag^2) and this post is a bit embarrassing to read! Check out my .NET FFT library FftSharp for a more advanced discussion on this topic.
These concepts are simple to visualize when graphed. Here I’ve written a short Python script to listen to the microphone (which is being fed a 2kHz sine wave), perform the FFT, and graph the real FFT component, imaginary FFT component, and their sum. The output is:
Of particular interest to me is the beautiful complementary of the two curves. It makes me wonder what types of data can be extracted by the individual curves (or perhaps their difference?) down the road. I wonder if phase measurements would be useful in extracting weak carries from beneath the noise floor?
Here’s the code I used to generate the image above. Note that my microphone device was set to listen to my stereo output, and I generated a 2kHz sine wave using the command speaker-test -t sine -f 2000 on a PC running Linux. I hope you find it useful!
import numpy
import pyaudio
import pylab
import numpy
### RECORD AUDIO FROM MICROPHONE ###
rate = 44100
soundcard = 1 # CUSTOMIZE THIS!!!
p = pyaudio.PyAudio()
strm = p.open(format=pyaudio.paInt16, channels=1, rate=rate,
input_device_index=soundcard, input=True)
strm.read(1024) # prime the sound card this way
pcm = numpy.fromstring(strm.read(1024), dtype=numpy.int16)
### DO THE FFT ANALYSIS ###
fft = numpy.fft.fft(pcm)
fftr = 10*numpy.log10(abs(fft.real))[:len(pcm)/2]
ffti = 10*numpy.log10(abs(fft.imag))[:len(pcm)/2]
fftb = 10*numpy.log10(numpy.sqrt(fft.imag**2+fft.real**2))[:len(pcm)/2]
freq = numpy.fft.fftfreq(numpy.arange(len(pcm)).shape[-1])[:len(pcm)/2]
freq = freq*rate/1000 # make the frequency scale
### GRAPH THIS STUFF ###
pylab.subplot(411)
pylab.title("Original Data")
pylab.grid()
pylab.plot(numpy.arange(len(pcm))/float(rate)*1000, pcm, 'r-', alpha=1)
pylab.xlabel("Time (milliseconds)")
pylab.ylabel("Amplitude")
pylab.subplot(412)
pylab.title("Real FFT")
pylab.xlabel("Frequency (kHz)")
pylab.ylabel("Power")
pylab.grid()
pylab.plot(freq, fftr, 'b-', alpha=1)
pylab.subplot(413)
pylab.title("Imaginary FFT")
pylab.xlabel("Frequency (kHz)")
pylab.ylabel("Power")
pylab.grid()
pylab.plot(freq, ffti, 'g-', alpha=1)
pylab.subplot(414)
pylab.title("Real+Imaginary FFT")
pylab.xlabel("Frequency (kHz)")
pylab.ylabel("Power")
pylab.grid()
pylab.plot(freq, fftb, 'k-', alpha=1)
pylab.show()
After fighting for a while long with a “shifty baseline” of the FFT, I came to another understanding. Let me first address the problem. Taking the FFT of different regions of the 2kHz wave I got traces with the peak in the identical location, but the “baselines” completely different.
Like many things, I re-invented the wheel. Since I knew the PCM values weren’t changing, the only variable was the starting/stopping point of the linear sample. “Hard edges”, I imagined, must be the problem. I then wrote the following function to shape the PCM audio like a triangle, silencing the edges and sweeping the volume up toward the middle of the sample:
def shapeTriangle(data):
triangle=numpy.array(range(len(data)/2)+range(len(data)/2)[::-1])+1
return data*triangle
After shaping the data BEFORE I applied the FFT, I made the subsequent traces MUCH more acceptable. Observe:
Now that I’ve done all this experimentation/thinking, I remembered that this is nothing new! Everyone talks about shaping the wave to minimize hard edges before taking the FFT. They call it windowing. Another case of me re-inventing the wheel because I’m too lazy to read others’ work. However, in my defense, I learned a lot by trying all this stuff – far more than I would have learned simply by copying someone else’s code into my script. Experimentation is the key to discovery!
⚠️ WARNING: This page is obsolete
Articles typically receive this designation when the technology they describe is no longer relevant, code
provided is later deemed to be of poor quality, or the topics discussed are better presented in future
articles. Articles like this are retained for the sake of preservation, but their content should be
critically assessed.
⚠️ SEE UPDATED POST: Signal Filtering in Python
While I wrote a pervious post on linear data smoothing with python, those scripts were never fully polished. Fred (KJ4LFJ) asked me about this today and I felt bad I had nothing to send him. While I might add that the script below isn’t polished, at least it’s clean. I’ve been using this method for all of my smoothing recently. Funny enough, none of my code was clean enough to copy and paste, so I wrote this from scratch tonight. It’s a function to take a list in (any size) and smooth it with a triangle window (of any size, given by “degree”) and return the smoothed data with or without flanking copies of data to make it the identical length as before. The script also graphs the original data vs. smoothed traces of varying degrees. The output is below. I hope it helps whoever wants it!
import numpy
import pylab
def smoothTriangle(data, degree, dropVals=False):
"""
performs moving triangle smoothing with a variable degree.
note that if dropVals is False, output length will be identical
to input length, but with copies of data at the flanking regions
"""
triangle = numpy.array(range(degree)+[degree]+range(degree)[::-1])+1
smoothed = []
for i in range(degree, len(data)-degree*2):
point = data[i:i+len(triangle)]*triangle
smoothed.append(sum(point)/sum(triangle))
if dropVals:
return smoothed
smoothed = [smoothed[0]]*(degree+degree/2)+smoothed
while len(smoothed) < len(data):
smoothed.append(smoothed[-1])
return smoothed
### CREATE SOME DATA ###
data = numpy.random.random(100) # make 100 random numbers from 0-1
data = numpy.array(data*100, dtype=int) # make them integers from 1 to 100
for i in range(100):
data[i] = data[i]+i**((150-i)/80.0) # give it a funny trend
### GRAPH ORIGINAL/SMOOTHED DATA ###
pylab.plot(data, "k.-", label="original data", alpha=.3)
pylab.plot(smoothTriangle(data, 3), "-", label="smoothed d=3")
pylab.plot(smoothTriangle(data, 5), "-", label="smoothed d=5")
pylab.plot(smoothTriangle(data, 10), "-", label="smoothed d=10")
pylab.title("Moving Triangle Smoothing")
pylab.grid(alpha=.3)
pylab.axis([20, 80, 50, 300])
pylab.legend()
pylab.show()
⚠️ WARNING: This page is obsolete
Articles typically receive this designation when the technology they describe is no longer relevant, code
provided is later deemed to be of poor quality, or the topics discussed are better presented in future
articles. Articles like this are retained for the sake of preservation, but their content should be
critically assessed.
While thinking of ways to improve my QRSS VD high-definitions spectrograph software, I often wish I had a better way to display large spectrographs. Currently I’m using PIL (the Python Imaging Library) with TK and it’s slow as heck. I looked into the PyGame project, and it seems to be designed with speed in mind. I whipped-up this quick demo, and it’s a simple case audio spectrograph which takes in audio from your sound card and graphs it time vs. frequency. This method is far superior to the method I was using previously to display the data, because while QRSS VD can only update the entire GUI (500px by 8,000 px) every 3 seconds, early tests with PyGame suggests it can do it about 20 times a second (wow!). With less time/CPU going into the GUI, the program can be more responsivle and my software can be less of a drain.
import pygame
import numpy
import threading
import pyaudio
import scipy
import scipy.fftpack
import scipy.io.wavfile
import wave
rate = 12000 # try 5000 for HD data, 48000 for realtime
soundcard = 2
windowWidth = 500
fftsize = 512
currentCol = 0
scooter = []
overlap = 5 # 1 for raw, realtime - 8 or 16 for high-definition
def graphFFT(pcm):
global currentCol, data
ffty = scipy.fftpack.fft(pcm) # convert WAV to FFT
ffty = abs(ffty[0:len(ffty)/2])/500 # FFT is mirror-imaged
# ffty=(scipy.log(ffty))*30-50 # if you want uniform data
print "MIN:t%stMAX:t%s" % (min(ffty), max(ffty))
for i in range(len(ffty)):
if ffty[i] < 0:
ffty[i] = 0
if ffty[i] > 255:
ffty[i] = 255
scooter.append(ffty)
if len(scooter) < 6:
return
scooter.pop(0)
ffty = (scooter[0]+scooter[1]*2+scooter[2]*3+scooter[3]*2+scooter[4])/9
data = numpy.roll(data, -1, 0)
data[-1] = ffty[::-1]
currentCol += 1
if currentCol == windowWidth:
currentCol = 0
def record():
p = pyaudio.PyAudio()
inStream = p.open(format=pyaudio.paInt16, channels=1, rate=rate,
input_device_index=soundcard, input=True)
linear = [0]*fftsize
while True:
linear = linear[fftsize/overlap:]
pcm = numpy.fromstring(inStream.read(
fftsize/overlap), dtype=numpy.int16)
linear = numpy.append(linear, pcm)
graphFFT(linear)
pal = [(max((x-128)*2, 0), x, min(x*2, 255)) for x in xrange(256)]
print max(pal), min(pal)
data = numpy.array(numpy.zeros((windowWidth, fftsize/2)), dtype=int)
# data=Numeric.array(data) # for older PyGame that requires Numeric
pygame.init() # crank up PyGame
pygame.display.set_caption("Simple Spectrograph")
screen = pygame.display.set_mode((windowWidth, fftsize/2))
world = pygame.Surface((windowWidth, fftsize/2), depth=8) # MAIN SURFACE
world.set_palette(pal)
t_rec = threading.Thread(target=record) # make thread for record()
t_rec.daemon = True # daemon mode forces thread to quit with program
t_rec.start() # launch thread
clk = pygame.time.Clock()
while 1:
for event in pygame.event.get(): # check if we need to exit
if event.type == pygame.QUIT:
pygame.quit()
sys.exit()
pygame.surfarray.blit_array(world, data) # place data in window
screen.blit(world, (0, 0))
pygame.display.flip() # RENDER WINDOW
clk.tick(30) # limit to 30FPS
