Added code to support training of a 'left_center_right' model.

core/utils.py

@@ -220,3 +220,14 @@ def one_hot_encoding_left_or_right(a, dtype='int'):
     b = np.zeros((a.size, a.max() + 1))
     b[np.arange(a.size), a] = 1
     return b
+
+
+def azimuth_to_left_center_right_onehot(labels, in_place=False):
+    if not in_place:
+        labels = labels.copy()
+    labels = labels.astype(dtype=np.int, copy=False)
+    labels[(labels > 180) & (labels < 360)] = 0  # left side
+    labels[(labels == 180) | (labels == 360)] = 1  # center
+    labels[(labels > 0) & (labels < 180)] = 2  # right side
+    labels = tf.one_hot(labels, depth=3)
+    return labels

scripts/train_model.py

 import argparse
 from datetime import datetime
-from core.utils import load_ircam_hrirs_data, split_dataset, generate_signals, generate_augmented_labelled_dataset
+from core.utils import load_ircam_hrirs_data, split_dataset, generate_signals, generate_augmented_labelled_dataset, \
+    azimuth_to_left_center_right_onehot
 import numpy as np
 import tensorflow as tf
 from core.logger import experiment_logger
@@ -57,7 +58,7 @@ if __name__ == '__main__':
             if not 'convolve_mode' in exp_config.keys() or exp_config['convolve_mode'] == 'valid':
                 # the valid part from the convolved sound is sound[len(hrir)-1:-len(hrir)+1]
                 # hence we have to do len(hrir)-1 left and right zero padding so that the dirac belong to the result
-                n_zeros = 512-1
+                n_zeros = 512 - 1
                 zeros = np.zeros(shape=(2, n_zeros))
                 sounds = np.concatenate([zeros, np.ones(shape=(2, 1)), zeros], axis=1)
                 training_sounds, test_sounds = np.array_split(sounds, 2, axis=0)
@@ -136,9 +137,23 @@ if __name__ == '__main__':
 
         opt = tf.keras.optimizers.RMSprop(learning_rate=exp_config['learning_rate'])
         # opt = tf.keras.optimizers.Adam(learning_rate=lr)
-        model.compile(optimizer=opt,
-                      loss=MeanAbsoluteAzimuthError(),  # tf.keras.losses.MeanSquaredError(),
-                      metrics=[])
+
+        if exp_config['model_name'] == 'default':
+            model.compile(optimizer=opt,
+                          loss=MeanAbsoluteAzimuthError(),  # tf.keras.losses.MeanSquaredError(),
+                          metrics=[])
+        elif exp_config['model_name'] == 'left_center_right':
+            model.compile(optimizer=opt,
+                          loss=tf.keras.losses.CategoricalCrossentropy(),  # tf.keras.losses.MeanSquaredError(),
+                          metrics=[  # tf.keras.metrics.TruePositives(name='tp'),
+                              # tf.keras.metrics.FalsePositives(name='fp'),
+                              # tf.keras.metrics.TrueNegatives(name='tn'),
+                              # tf.keras.metrics.FalseNegatives(name='fn'),
+                              tf.keras.metrics.Precision(name='precision'),
+                              tf.keras.metrics.Recall(name='recall'),
+                              tf.keras.metrics.CategoricalAccuracy(name='acc'),
+                              # tf.keras.metrics.AUC(name='auc'),
+                          ])
         # model.summary()
 
         # model = default_model_creator(lr=exp_config['learning_rate'])
@@ -165,6 +180,10 @@ if __name__ == '__main__':
                                                                             n_samples=n_samples,
                                                                             n_augment=1)
 
+            if exp_config['model_name'] == 'left_center_right':
+                training_labels = azimuth_to_left_center_right_onehot(training_labels, in_place=True)
+                test_labels = azimuth_to_left_center_right_onehot(test_labels, in_place=True)
+
             # Expand dimensions of arrays to comply with the stanrdized shape BWHF (Batch, Width, Height, Features).
             # Here, width=time axis, and Height=channel (left or right)
             training_signals = np.expand_dims(training_signals, axis=-1)