gtLara · May 14, 2022 07:24 · Apr 9, 2022 · Apr 9, 2022 · Apr 9, 2022 · Mar 30, 2022
diff --git a/README.md b/README.md
@@ -4,12 +4,10 @@ My first experiment with Keras.
 
 ![](https://gist.github.com/endolith/1de9a8700f72b97974a2e93b0fba316a/raw/08d6ddf4ce234c7c6a559a628b009fe5c4bce6ea/FIR.png)
 
-Which is the same structure as a neural net (assuming no activation function),
-
-
+Which is the same structure as a neural net (assuming no activation function):
 
 ![](https://gist.github.com/endolith/1de9a8700f72b97974a2e93b0fba316a/raw/08d6ddf4ce234c7c6a559a628b009fe5c4bce6ea/ANN.png)
 
-so it should be able to learn the FIR coefficients by learning from chunks of signal before and after filtering, right?
+So it should be able to learn the FIR coefficients by learning from chunks of signal before and after filtering, right?
 
 **Conclusion:** Yep, it works great.
diff --git a/README.md b/README.md
@@ -2,13 +2,13 @@ My first experiment with Keras.
 
 **Hypothesis:** Each output sample of an FIR filter is just a sum of weighted input samples taken from a small chunk of the input:
 
-![](https://upload.wikimedia.org/wikipedia/commons/thumb/9/9b/FIR_Filter.svg/320px-FIR_Filter.svg.png)
+![](https://gist.github.com/endolith/1de9a8700f72b97974a2e93b0fba316a/raw/08d6ddf4ce234c7c6a559a628b009fe5c4bce6ea/FIR.png)
 
 Which is the same structure as a neural net (assuming no activation function),
 
 
 
-![](https://upload.wikimedia.org/wikipedia/commons/thumb/6/6a/Perceptron-unit.svg/266px-Perceptron-unit.svg.png)
+![](https://gist.github.com/endolith/1de9a8700f72b97974a2e93b0fba316a/raw/08d6ddf4ce234c7c6a559a628b009fe5c4bce6ea/ANN.png)
 
 so it should be able to learn the FIR coefficients by learning from chunks of signal before and after filtering, right?
 

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -106,7 +106,7 @@ def rolling_window(a, window):
 Node-level graph
 """
 
-# Broken, requires old keras.
+# Working version: https://github.com/endolith/ann-visualizer
 # from ann_visualizer.visualize import ann_viz
 # ann_viz(model, title="Learned FIR filter")
 

diff --git a/README.md b/README.md
@@ -1 +1,15 @@
-My first experiment with keras.
+My first experiment with Keras.  
+
+**Hypothesis:** Each output sample of an FIR filter is just a sum of weighted input samples taken from a small chunk of the input:
+
+![](https://upload.wikimedia.org/wikipedia/commons/thumb/9/9b/FIR_Filter.svg/320px-FIR_Filter.svg.png)
+
+Which is the same structure as a neural net (assuming no activation function),
+
+
+
+![](https://upload.wikimedia.org/wikipedia/commons/thumb/6/6a/Perceptron-unit.svg/266px-Perceptron-unit.svg.png)
+
+so it should be able to learn the FIR coefficients by learning from chunks of signal before and after filtering, right?
+
+**Conclusion:** Yep, it works great.
diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -52,7 +52,7 @@ def rolling_window(a, window):
 
 
 X = rolling_window(sig, numtaps)
-Y = filtered
+Y = filtered  # Filter outputs 1 sample for each chunk of input samples
 
 # plt.plot(X[0])
 # plt.plot(Y[0])

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -124,21 +124,14 @@ def rolling_window(a, window):
 class LossHistory(Callback):
     def on_train_begin(self, logs={}):
         self.losses = []
-#        self.n = 0
-#        maxx = np.amax(np.abs(model.get_weights()[0]))
-#        plt.imsave(str(self.n)+'.png', model.get_weights()[0],
-#                   cmap=cmap, vmin=-maxx, vmax=maxx)
-#        self.n+=1
+        # Could plot the convergence here
 
     def on_batch_end(self, batch, logs={}):
         self.losses.append(logs.get('loss'))
 
     def on_epoch_end(self, batch, logs={}):
         pass
-#        maxx = np.amax(np.abs(model.get_weights()[0]))
-#        plt.imsave(str(self.n)+'.png', model.get_weights()[0],
-#                   cmap=cmap, vmin=-maxx, vmax=maxx)
-#        self.n+=1
+        # Could plot the convergence here
 
 
 history = LossHistory()

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -85,11 +85,6 @@ def rolling_window(a, window):
 """
 Make block diagram of network (not from tutorial)
 """
-
-# Workaround for some packaging bug
-import os
-os.environ["PATH"] += os.pathsep + r'C:\Anaconda3\Library\bin\graphviz'
-
 from tensorflow.keras.utils import plot_model
 plot_model(model, to_file='model.png', show_shapes=True)
 

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -15,8 +15,10 @@
 Generate or load a signal to use as input data
 """
 # Only learns at the frequencies present in the signal
-# Learns at all frequencies with white noise
+# https://en.wikipedia.org/wiki/File:Short-beaked_Echidna.ogg
 sig, fs = read('echidna.wav')
+
+# Learns at all frequencies with white noise
 # sig, fs = np.random.randn(10000), 10000
 
 """

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -109,13 +109,14 @@ def rolling_window(a, window):
 Node-level graph
 """
 
-from ann_visualizer.visualize import ann_viz
-
 # Broken, requires old keras.
-# Replace with https://github.com/Dicksonchin93/keras-architecture-visualizer/
+# from ann_visualizer.visualize import ann_viz
 # ann_viz(model, title="Learned FIR filter")
 
-
+# https://github.com/Dicksonchin93/keras-architecture-visualizer/
+# from keras_architecture_visualizer import KerasArchitectureVisualizer
+# vis = KerasArchitectureVisualizer()
+# vis.visualize(model)
 
 # Compile model
 model.compile(loss='mean_squared_error',

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -25,8 +25,9 @@
 numtaps = 51
 # b = signal.firwin(numtaps, 1, fs=fs)
 # b = signal.firwin(numtaps, cutoff=[0.3, 0.5], window='blackmanharris',
-#                   pass_zero=False)
-b = signal.firwin(numtaps, [500, 1000], pass_zero=True, width=100, fs=fs)
+#                    pass_zero=False)
+b = signal.firwin(numtaps, cutoff=[6000, 11000], fs=fs,
+                  window='blackmanharris', pass_zero=False)
 
 # TODO: Use an IIR filter and have ANN approximate it as best it can
 

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -11,18 +11,6 @@
 import matplotlib.pyplot as plt
 from soundfile import read
 
-"""
-Create the FIR filter for the ANN to copy
-"""
-# TODO: Use an IIR filter and have ANN approximate it as best it can
-numtaps = 51
-#b = signal.firwin(numtaps, 1, fs=fs)
-b = signal.firwin(numtaps, cutoff=[0.3, 0.5], window='blackmanharris',
-                  pass_zero=False)
-#b = signal.firwin(numtaps, [500, 1000], pass_zero=True, width=100, fs=fs)
-
-
-
 """
 Generate or load a signal to use as input data
 """
@@ -31,6 +19,16 @@
 sig, fs = read('echidna.wav')
 # sig, fs = np.random.randn(10000), 10000
 
+"""
+Create the FIR filter for the ANN to copy
+"""
+numtaps = 51
+# b = signal.firwin(numtaps, 1, fs=fs)
+# b = signal.firwin(numtaps, cutoff=[0.3, 0.5], window='blackmanharris',
+#                   pass_zero=False)
+b = signal.firwin(numtaps, [500, 1000], pass_zero=True, width=100, fs=fs)
+
+# TODO: Use an IIR filter and have ANN approximate it as best it can
 
 """
 Training data is chunks of input and output of FIR filter

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -158,15 +158,17 @@ def on_epoch_end(self, batch, logs={}):
 
 final = model.get_weights()
 
-plt.figure('kernel')
-plt.plot(b, '.-', label='Filter')
-plt.plot(initial[0], '.-', label='Initial')
-plt.plot(final[0], '.-', label='Learned')
-plt.grid(True, color='0.7', linestyle='-', which='major')
-plt.grid(True, color='0.9', linestyle='-', which='minor')
-plt.title('Kernel')
-plt.legend()
-
+fig, (ax0, ax1) = plt.subplots(nrows=2, ncols=1, num='kernel', sharex=True)
+ax1.plot(b, '.-', label='Filter', alpha=0.5, c='gray')
+ax0.plot(initial[0], '.-', label='Initial')
+ax1.plot(final[0], '.-', label='Learned')
+ax0.grid(True, color='0.7', linestyle='-', which='major')
+ax0.grid(True, color='0.9', linestyle='-', which='minor')
+ax1.grid(True, color='0.7', linestyle='-', which='major')
+ax1.grid(True, color='0.9', linestyle='-', which='minor')
+ax0.set_title('Kernel')
+ax0.legend()
+ax1.legend()
 
 plt.figure('frequency response')
 w, h = signal.freqz(b, [1.0])

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -152,9 +152,9 @@ def on_epoch_end(self, batch, logs={}):
 model.fit(X, Y, epochs=35, batch_size=100, callbacks=[history])
 
 # evaluate the model
-#print("Evaluating...")
-#scores = model.evaluate(X, Y)
-#print(scores[1]*100)
+print("Evaluating...")
+scores = model.evaluate(X, Y)
+print(scores*100)  # percent??
 
 final = model.get_weights()
 

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -59,9 +59,18 @@ def rolling_window(a, window):
 
 """
 Create model
+
+Initializer matters because signal might have missing areas of spectrum, and
+model will not learn there. So if the initial guess is all zeros, those areas
+of the spectrum will stay silenced, while the passband is "built up" for
+frequencies that are present.
 """
 model = Sequential([
-    Dense(1, input_dim=numtaps, use_bias=False)
+    Dense(1, input_dim=numtaps, use_bias=False,
+          # kernel_initializer='random_normal',  # typical usage
+          # kernel_initializer='ones',  # boxcar window = running average
+          kernel_initializer='zeros',  # nothing (good for non-white input)
+          )
     ])
 
 model.summary()

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -28,8 +28,8 @@
 """
 # Only learns at the frequencies present in the signal
 # Learns at all frequencies with white noise
-#sig, fs = read('echidna.wav')
-sig = np.random.randn(10000)
+sig, fs = read('echidna.wav')
+# sig, fs = np.random.randn(10000), 10000
 
 
 """
@@ -161,16 +161,17 @@ def on_epoch_end(self, batch, logs={}):
 
 plt.figure('frequency response')
 w, h = signal.freqz(b, [1.0])
-plt.semilogx(w, 20*np.log10(abs(h)), label='Filter')
-
+plt.semilogx(w*fs/(2*np.pi), 20*np.log10(abs(h)), label='Filter',
+             alpha=0.5, c='gray')
 w, h = signal.freqz(initial[0], [1.0])
-plt.semilogx(w, 20*np.log10(abs(h)), label='Initial')
-
+plt.semilogx(w*fs/(2*np.pi), 20*np.log10(abs(h)), label='Initial')
 w, h = signal.freqz(final[0], [1.0])
-plt.semilogx(w, 20*np.log10(abs(h)), label='Learned')
-
+plt.semilogx(w*fs/(2*np.pi), 20*np.log10(abs(h)), label='Learned', alpha=0.5)
 plt.grid(True, color='0.7', linestyle='-', which='major')
 plt.grid(True, color='0.9', linestyle='-', which='minor')
+plt.xlabel('Frequency [Hz]')
+plt.ylabel('Response [dB]')
+plt.xlim(None, fs/2)
 plt.title('Frequency response')
 plt.legend()
 

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -21,9 +21,6 @@
                   pass_zero=False)
 #b = signal.firwin(numtaps, [500, 1000], pass_zero=True, width=100, fs=fs)
 
-# Plot filter response
-#w, h = signal.freqz(b, [1.0], 2000)
-#plt.plot(w/np.pi*fs/2, 20*np.log10(abs(h)), label='Filter')
 
 
 """

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -15,7 +15,7 @@
 Create the FIR filter for the ANN to copy
 """
 # TODO: Use an IIR filter and have ANN approximate it as best it can
-numtaps = 1001
+numtaps = 51
 #b = signal.firwin(numtaps, 1, fs=fs)
 b = signal.firwin(numtaps, cutoff=[0.3, 0.5], window='blackmanharris',
                   pass_zero=False)

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -152,16 +152,17 @@ def on_epoch_end(self, batch, logs={}):
 
 final = model.get_weights()
 
-plt.figure()
+plt.figure('kernel')
 plt.plot(b, '.-', label='Filter')
 plt.plot(initial[0], '.-', label='Initial')
 plt.plot(final[0], '.-', label='Learned')
 plt.grid(True, color='0.7', linestyle='-', which='major')
 plt.grid(True, color='0.9', linestyle='-', which='minor')
+plt.title('Kernel')
 plt.legend()
 
 
-plt.figure()
+plt.figure('frequency response')
 w, h = signal.freqz(b, [1.0])
 plt.semilogx(w, 20*np.log10(abs(h)), label='Filter')
 
@@ -173,10 +174,12 @@ def on_epoch_end(self, batch, logs={}):
 
 plt.grid(True, color='0.7', linestyle='-', which='major')
 plt.grid(True, color='0.9', linestyle='-', which='minor')
+plt.title('Frequency response')
 plt.legend()
 
-plt.figure()
+plt.figure('loss')
 plt.semilogy(history.losses)
 plt.xlabel('Batch')
 plt.ylabel('Loss')
 plt.grid(True, which="both")
+plt.title('Loss')
diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -3,9 +3,9 @@
 
 Created on Fri Aug  3 15:00:40 2018
 """
-from keras.models import Sequential
-from keras.layers import Dense
-from keras.callbacks import Callback
+from tensorflow.keras.models import Sequential
+from tensorflow.keras.layers import Dense
+from tensorflow.keras.callbacks import Callback
 import numpy as np
 from scipy import signal
 import matplotlib.pyplot as plt
@@ -83,7 +83,7 @@ def rolling_window(a, window):
 import os
 os.environ["PATH"] += os.pathsep + r'C:\Anaconda3\Library\bin\graphviz'
 
-from keras.utils import plot_model
+from tensorflow.keras.utils import plot_model
 plot_model(model, to_file='model.png', show_shapes=True)
 
 
@@ -94,8 +94,9 @@ def rolling_window(a, window):
 """
 import tensorflow as tf
 
-with tf.Session() as sess:
-    writer = tf.summary.FileWriter('logs', sess.graph)
+# TODO: This isn't working like it used to. Replace with TF2.0 conventions.
+with tf.compat.v1.Session() as sess:
+    writer = tf.compat.v1.summary.FileWriter('logs', sess.graph)
     writer.close()
 
 
@@ -105,7 +106,9 @@ def rolling_window(a, window):
 
 from ann_visualizer.visualize import ann_viz
 
-#ann_viz(model, title="Learned FIR filter")
+# Broken, requires old keras.
+# Replace with https://github.com/Dicksonchin93/keras-architecture-visualizer/
+# ann_viz(model, title="Learned FIR filter")
 
 
 

diff --git a/FIR_filter_NN.py b/FIR_filter_NN.py
@@ -0,0 +1,179 @@
+"""
+Train a neural network to learn an FIR filter.
+
+Created on Fri Aug  3 15:00:40 2018
+"""
+from keras.models import Sequential
+from keras.layers import Dense
+from keras.callbacks import Callback
+import numpy as np
+from scipy import signal
+import matplotlib.pyplot as plt
+from soundfile import read
+
+"""
+Create the FIR filter for the ANN to copy
+"""
+# TODO: Use an IIR filter and have ANN approximate it as best it can
+numtaps = 1001
+#b = signal.firwin(numtaps, 1, fs=fs)
+b = signal.firwin(numtaps, cutoff=[0.3, 0.5], window='blackmanharris',
+                  pass_zero=False)
+#b = signal.firwin(numtaps, [500, 1000], pass_zero=True, width=100, fs=fs)
+
+# Plot filter response
+#w, h = signal.freqz(b, [1.0], 2000)
+#plt.plot(w/np.pi*fs/2, 20*np.log10(abs(h)), label='Filter')
+
+
+"""
+Generate or load a signal to use as input data
+"""
+# Only learns at the frequencies present in the signal
+# Learns at all frequencies with white noise
+#sig, fs = read('echidna.wav')
+sig = np.random.randn(10000)
+
+
+"""
+Training data is chunks of input and output of FIR filter
+"""
+# filtered = signal.lfilter(b, 1.0, sig)
+filtered = signal.convolve(sig, b, mode='valid')
+
+
+def rolling_window(a, window):
+    """
+    Return chunks of signal `a` of size `window`, incremented by 1 each time.
+
+    https://gist.github.com/codehacken/708f19ae746784cef6e68b037af65788
+    """
+    shape = a.shape[:-1] + (a.shape[-1] - window + 1, window)
+    strides = a.strides + (a.strides[-1],)
+    return np.lib.stride_tricks.as_strided(a, shape=shape, strides=strides)
+
+
+X = rolling_window(sig, numtaps)
+Y = filtered
+
+# plt.plot(X[0])
+# plt.plot(Y[0])
+
+
+"""
+Create model
+"""
+model = Sequential([
+    Dense(1, input_dim=numtaps, use_bias=False)
+    ])
+
+model.summary()
+
+initial = model.get_weights()
+
+print('Initial weights:')
+print(initial)
+
+
+"""
+Make block diagram of network (not from tutorial)
+"""
+
+# Workaround for some packaging bug
+import os
+os.environ["PATH"] += os.pathsep + r'C:\Anaconda3\Library\bin\graphviz'
+
+from keras.utils import plot_model
+plot_model(model, to_file='model.png', show_shapes=True)
+
+
+"""
+Make graph diagram of network (not from tutorial)
+
+Viewable with `tensorboard --logdir="logs"`
+"""
+import tensorflow as tf
+
+with tf.Session() as sess:
+    writer = tf.summary.FileWriter('logs', sess.graph)
+    writer.close()
+
+
+"""
+Node-level graph
+"""
+
+from ann_visualizer.visualize import ann_viz
+
+#ann_viz(model, title="Learned FIR filter")
+
+
+
+# Compile model
+model.compile(loss='mean_squared_error',
+              optimizer='adam',
+              )
+
+
+class LossHistory(Callback):
+    def on_train_begin(self, logs={}):
+        self.losses = []
+#        self.n = 0
+#        maxx = np.amax(np.abs(model.get_weights()[0]))
+#        plt.imsave(str(self.n)+'.png', model.get_weights()[0],
+#                   cmap=cmap, vmin=-maxx, vmax=maxx)
+#        self.n+=1
+
+    def on_batch_end(self, batch, logs={}):
+        self.losses.append(logs.get('loss'))
+
+    def on_epoch_end(self, batch, logs={}):
+        pass
+#        maxx = np.amax(np.abs(model.get_weights()[0]))
+#        plt.imsave(str(self.n)+'.png', model.get_weights()[0],
+#                   cmap=cmap, vmin=-maxx, vmax=maxx)
+#        self.n+=1
+
+
+history = LossHistory()
+
+
+# Fit the model
+print("Fitting...")
+model.fit(X, Y, epochs=35, batch_size=100, callbacks=[history])
+
+# evaluate the model
+#print("Evaluating...")
+#scores = model.evaluate(X, Y)
+#print(scores[1]*100)
+
+final = model.get_weights()
+
+plt.figure()
+plt.plot(b, '.-', label='Filter')
+plt.plot(initial[0], '.-', label='Initial')
+plt.plot(final[0], '.-', label='Learned')
+plt.grid(True, color='0.7', linestyle='-', which='major')
+plt.grid(True, color='0.9', linestyle='-', which='minor')
+plt.legend()
+
+
+plt.figure()
+w, h = signal.freqz(b, [1.0])
+plt.semilogx(w, 20*np.log10(abs(h)), label='Filter')
+
+w, h = signal.freqz(initial[0], [1.0])
+plt.semilogx(w, 20*np.log10(abs(h)), label='Initial')
+
+w, h = signal.freqz(final[0], [1.0])
+plt.semilogx(w, 20*np.log10(abs(h)), label='Learned')
+
+plt.grid(True, color='0.7', linestyle='-', which='major')
+plt.grid(True, color='0.9', linestyle='-', which='minor')
+plt.legend()
+
+plt.figure()
+plt.semilogy(history.losses)
+plt.xlabel('Batch')
+plt.ylabel('Loss')
+plt.grid(True, which="both")
diff --git a/README.md b/README.md
@@ -0,0 +1 @@
+My first experiment with keras.
No results found