predict smoothing

[asr1617data.git] / experiments.py

diff --git a/experiments.py b/experiments.py
index 699c68c..833e1c5 100644 (file)
--- a/experiments.py
+++ b/experiments.py
@@ -10,24 +10,27 @@ import numpy as np
 
 # keras
 from keras.models import Sequential
-from keras.layers import Dense, Dropout  # , Activation
-from keras import backend
+from keras.layers import Dense  # , Dropout  # , Activation
+from keras import backend, utils
 
 # Testset ratio
 testset = 0.10
 samplerate = 16000
-verbosity = 1
+verbosity = 0
+
 
 def get_datafiles():
     files = glob.glob(os.path.join(os.getcwd(), 'textgrid', '*.TextGrid'))
     # Loop over all datafiles and make wavefile string
     for i, tg in enumerate(files):
         num = re.match('^.*/(\\d+).TextGrid$', tg).group(1)
-        yield (tg, 'wav/{:02d}.wav'.format(int(num)))
+        yield (tg, 'wav/{:02d}.wav'.format(int(num)), int(num))
+
 
 def label_from_annotation(ann):
     return 0 if ann.strip() == '' else 1
 
+
 def features_from_wav(tg, wavp, typ='mfcc', winlen=0.025, winstep=0.01):
     # Load textgrid
     tgob = pympi.TextGrid(tg)
@@ -39,7 +42,8 @@ def features_from_wav(tg, wavp, typ='mfcc', winlen=0.025, winstep=0.01):
         data = mfcc(sig, rate, winlen=winlen, winstep=winstep, numcep=13,
                     appendEnergy=True)
     elif typ == 'fbank':
-        (data, energy) = logfbank(sig, rate, winlen=winlen, winstep=winstep, nfilt=26)
+        (data, energy) = logfbank(
+            sig, rate, winlen=winlen, winstep=winstep, nfilt=26)
     else:
         raise ValueError("No such type")
 
@@ -62,18 +66,39 @@ def features_from_wav(tg, wavp, typ='mfcc', winlen=0.025, winstep=0.01):
         i += 1
     return (data, labels)
 
-def run(typ, winlen, winstep, modelfun, modelname):
+
+def singerfun(num, l):
+    if l == 1:
+        if 0 <= num <= 11:
+            return 1
+        elif 12 <= num <= 21:
+            return 2
+        elif 22 <= num <= 28:
+            return 3
+        else:
+            raise Exception("halp")
+    else:
+        return 0
+
+
+def run(typ, winlen, winstep, modelfun, modelname, multiclass=False):
     datas = []
     labels = []
 
-    for tg, wavp in get_datafiles():
+    for tg, wavp, num in get_datafiles():
         (d, l) = features_from_wav(
             tg, wavp, winlen=winlen, winstep=winstep, typ=typ)
         datas.append(d)
-        labels.append(l)
+        if multiclass:
+            labels.append(list(map(lambda x: singerfun(int(num), x), l)))
+        else:
+            labels.append(l)
 
     datas = np.concatenate(datas)
     labels = np.concatenate(labels)
+    print(np.unique(labels, return_counts=True))
+    if multiclass:
+        labels = utils.to_categorical(labels, num_classes=4)
 
     rng_state = np.random.get_state()
     np.random.shuffle(datas)
@@ -87,46 +112,65 @@ def run(typ, winlen, winstep, modelfun, modelname):
 
     model = modelfun(traindata)
 
-    #Train
+    # Train
     model.fit(traindata, trainlabels, epochs=10, batch_size=32, shuffle=False,
               verbose=verbosity)
 
-    #Test
+    # Test
     loss, acc = model.evaluate(testdata, testlabels, batch_size=32,
                                verbose=verbosity)
     print('{}\t{}\t{}\t{}\t{}\n'.format(
         winlen, winstep, modelname, loss, acc))
     return model
 
-def simplemodel(d):
-    model = Sequential()
-    model.add(
-        Dense(d.shape[1]*2, input_shape=(d.shape[1],), activation='relu'))
-    model.add(Dense(100, activation='relu'))
-    model.add(Dense(1, activation='sigmoid'))
-    model.compile(optimizer='rmsprop',
-                  loss='binary_crossentropy',
-                  metrics=['accuracy'])
-    return model
 
-def bottlemodel(d):
-    model = Sequential()
-    model.add(
-        Dense(d.shape[1]*2, input_shape=(d.shape[1],), activation='relu'))
-    model.add(Dense(13, activation='relu'))
-    model.add(Dense(1, activation='sigmoid'))
-    model.compile(optimizer='rmsprop',
-                  loss='binary_crossentropy',
-                  metrics=['accuracy'])
-    return model
+def bottlemodel(layers):
+    def fun(d):
+        model = Sequential()
+        model.add(Dense(layers, activation='relu', input_shape=(d.shape[1],)))
+        model.add(Dense(1, activation='sigmoid'))
+#    model.add(
+#        Dense(d.shape[1]*2, input_shape=(d.shape[1],), activation='relu'))
+#    model.add(Dense(13, activation='relu'))
+#    model.add(Dense(1, activation='sigmoid'))
+        model.compile(optimizer='rmsprop',
+                      loss='binary_crossentropy',
+                      metrics=['accuracy'])
+        return model
+    return fun
+
+
+def multimodel(layers):
+    def fun(d):
+        model = Sequential()
+#  model.add(Dense(d.shape[1]*2, input_shape=(d.shape[1],), activation='relu'))
+        model.add(Dense(layers, activation='relu', input_shape=(d.shape[1],)))
+        model.add(Dense(4, activation='softmax'))
+        model.compile(optimizer='rmsprop',
+                      loss='categorical_crossentropy',
+                      metrics=['accuracy'])
+        return model
+    return fun
+
+
+models = [
+    ('bottle3', bottlemodel(3), False),
+    ('bottle5', bottlemodel(5), False),
+    ('bottle8', bottlemodel(8), False),
+    ('bottle13', bottlemodel(13), False),
+    ('multi3', multimodel(3), True),
+    ('multi5', multimodel(5), True),
+    ('multi8', multimodel(8), True),
+    ('multi13', multimodel(13), True)]
 
 if __name__ == '__main__':
     print('winlen\twinstep\tmodel\tloss\taccuracy\n')
     with backend.get_session():
         for winlen, winstep in ((0.025, 0.01), (0.1, 0.04), (0.2, 0.08)):
-            for name, model in (('simple', simplemodel), ('bottle', bottlemodel)):
-                m = run('mfcc', winlen, winstep, model, name)
-                fproot = 'model_{}_{}_{}.json'.format(winlen, winstep, name)
+            for name, model, multi in models:
+                m = run('mfcc', winlen, winstep, model, name, multi)
+                fproot = 'model_{}_{}_{}'.format(winlen, winstep, name)
+                print(fproot);
                 with open('{}.json'.format(fproot), 'w') as f:
                     f.write(m.to_json())
                 m.save_weights('{}.hdf5'.format(fproot))