big update

plots/normal_vs_abnormal.py

@@ -0,0 +1,60 @@
+import numpy as np
+import matplotlib as mpl
+import matplotlib.pyplot as plt
+import librosa
+import librosa.display
+# set LaTeX font
+# ==============
+nice_fonts = {
+    # Use LaTeX to write all text
+    "text.usetex": True,
+    "font.family": "serif",
+    # Use 10pt font in plots, to match 10pt font in document
+    "axes.labelsize": 16,  # default 10
+    "font.size": 16,  # default 10
+    # Make the legend/label fonts a little smaller
+    "legend.fontsize": 14,  # default 8
+    "xtick.labelsize": 14,  # default 8
+    "ytick.labelsize": 14,  # default 8
+}
+mpl.rcParams.update(nice_fonts, )
+mpl.rcParams['axes.unicode_minus'] = False
+
+# TODO LaTex
+fan_normal = '../data/mimii/-6_dB_fan/id_04/normal/00000042.wav'
+fan_anomaly = '../data/mimii/-6_dB_fan/id_04/abnormal/00000048.wav'
+
+pump_normal = '../data/mimii/-6_dB_pump/id_04/normal/00000042.wav'
+pump_anomaly = '../data/mimii/-6_dB_pump/id_04/abnormal/00000042.wav'
+
+slider_normal = '../data/mimii/-6_dB_slider/id_04/normal/00000042.wav'
+slider_anomaly = '../data/mimii/-6_dB_slider/id_04/abnormal/00000042.wav'
+
+valve_normal = '../data/mimii/-6_dB_valve/id_04/normal/00000042.wav'
+valve_anomaly = '../data/mimii/-6_dB_valve/id_04/abnormal/00000042.wav'
+
+fig = plt.figure(figsize=(17, 5))
+for i, (p, title) in enumerate(zip([fan_normal, pump_normal, slider_normal, valve_normal, fan_anomaly, pump_anomaly, slider_anomaly, valve_anomaly],
+                          ['Fan', 'Pump', 'Slider', 'Valve', 'Fan', 'Pump', 'Slider', 'Valve'])):
+    plt.subplot(2, 4, i+1)
+    audio, sr = librosa.load(p, sr=None, mono=True)
+    # n_fft=1024, hop_length=256, n_mels=80, center=False, power=2.0
+    S = librosa.feature.melspectrogram(y=audio, sr=sr, n_fft=1024, hop_length=256, n_mels=80, center=False, power=2.0)
+    S_dB = librosa.power_to_db(S, ref=np.max)
+    librosa.display.specshow(S_dB, x_axis='s' if i == 4 else 'off', hop_length=256,
+                             y_axis='mel' if i==4  else 'off', sr=16000, cmap='viridis')
+    if i < 4:
+        plt.title(title)
+    else:
+        plt.title(title + ' malfunction')
+
+cbar_ax = fig.add_axes([0.94, 0.18, 0.015, 0.7])
+cmap = mpl.cm.viridis
+norm = mpl.colors.Normalize(vmin=0, vmax=-80)
+cb1 = mpl.colorbar.ColorbarBase(cbar_ax, cmap=cmap,
+                                norm=norm,
+                                orientation='vertical',
+                                format='%+2.0f dB')
+plt.tight_layout()
+fig.subplots_adjust(right=0.93)
+plt.savefig('normal_vs_abnormal.png')

plots/playground.py

@@ -0,0 +1,64 @@
+import librosa
+from matplotlib import pyplot as plt
+import numpy as np
+from models.utils import count_parameters
+from models.ae import SubSpecCAE, AE, GlobalCAE
+from cfg import ALL_DATASET_PATHS
+import librosa.display
+import librosa.feature
+import seaborn as sns
+
+purple = (126/255, 87/255, 194/255)
+params_subspecae = count_parameters(SubSpecCAE())
+params_ae = count_parameters(AE())
+params_globalcae = count_parameters(GlobalCAE())
+print(f'#Parameters SubSpecCAE: {params_subspecae}')
+print(f'#Parameters AE: {params_ae}')
+print(f'#SubSpecAe/#AE: {(params_subspecae/params_ae):.2f}')
+print(f'#GlobalCAE: {params_globalcae}')
+print(f'#GlobalCAE/#SubSpecCAE: {params_globalcae/params_subspecae}')
+
+
+fig = plt.figure(figsize=(10, 5))
+slider_normal = '../data/mimii/-6_dB_slider/id_04/normal/00000042.wav'
+audio, sr = librosa.load(slider_normal)
+waveplot = librosa.display.waveplot(audio, sr=sr)
+plt.tight_layout()
+plt.savefig('wavplot.png')
+
+ids = range(0)
+specs = []
+i = 0
+audios = list((ALL_DATASET_PATHS[1] / 'id_00' / 'normal').glob('*.wav'))
+print(str(audios[80]))
+audio, sr = librosa.load(str(audios[80]), sr=None)
+mel_spec = librosa.feature.melspectrogram(audio, sr=sr, n_fft=1024, hop_length=256, n_mels=80, center=False, power=2.0)
+mel_spec_db = librosa.power_to_db(mel_spec, ref=np.max)
+librosa.display.specshow(mel_spec_db, hop_length=256, x_axis='s', y_axis='mel', cmap='viridis', sr=sr)
+plt.savefig('notadjusted.png')
+plt.clf()
+
+centroids = []
+for p in audios:
+    audio, sr = librosa.load(str(p), sr=None)
+    spectral_centroids = librosa.feature.spectral_centroid(audio, sr=sr)[0]
+    centroids += spectral_centroids.tolist()
+
+sns.distplot(centroids, hist=True, kde=True,
+             color=purple,
+             hist_kws={'edgecolor':'black'},
+             kde_kws={'linewidth': 2})
+plt.xlabel('Occurence counts')
+plt.ylabel('Density')
+plt.title('Spectral centroid distribution')
+plt.tight_layout()
+plt.savefig('hist.png')
+
+def get_bands(centroids, n_mels=80):
+    std = np.std(centroids)
+    mu = np.mean(centroids)
+    return int((mu-3*std)/n_mels), int((mu+3*std)/n_mels)
+
+print(get_bands(centroids))
+
+