BPM on a balanced dataset m17_dataclean_300k_balanced

autosync/dataset.py

@@ -63,7 +63,8 @@ def fetch_dataset_slices(dir, test_percent=10, block_ms=32, framerate=16000):
         #feat = feat[:, :16]
         #feat = dct(feat)
         feat = feat[:, :14]
-        feat = feat / 1000.0
+        # Remap to approx. 0 - 1
+        #feat = feat / 1000.0 + 1.0
         print('Feats min/max: {}/{}'.format(
             np.amin(feat),
             np.amax(feat)

autosync/sync.py

@@ -17,7 +17,7 @@ def sync(movie, sub):
             meta_graph_def = tf.saved_model.loader.load(
                 sess,
                 [tf.saved_model.tag_constants.SERVING],
-                'saved_models/m13_dataclean_300k',
+                'saved_models/m17_dataclean_300k_balanced',
             )
             signature = meta_graph_def.signature_def
             signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY

autosync/train.py

@@ -32,7 +32,7 @@ def train(dir='.'):
     epochs = 200
     batch_size = 32
     # m6 is just feat normalized
-    model_id = 'm15_dataclean_300k'
+    model_id = 'm17_dataclean_300k_balanced'
     logsess = 'logs/' + model_id
 
     training_set = create_dataset(train_audio, train_subs)
@@ -229,7 +229,7 @@ def train(dir='.'):
     '''
 
 def main():
-    dir = ["datasets/Song*.avi", "datasets/Game*.mkv"]
+    dir = ["datasets/10*.avi", "datasets/Game*.mkv"]
     print('Beginning training, data dir: {}'.format(dir))
     train(dir=dir)
 

autosync/verify.py

@@ -19,7 +19,7 @@ def verify(dir):
             meta_graph_def = tf.saved_model.loader.load(
                 sess,
                 [tf.saved_model.tag_constants.SERVING],
-                'saved_models/m15_dataclean_300k',
+                'saved_models/m17_dataclean_300k_balanced',
             )
             signature = meta_graph_def.signature_def
             signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
@@ -55,7 +55,7 @@ def verify(dir):
     result_bin = np.zeros(result[:, 0].shape)
     result_bin[result[:, 0] > 0.5] = 1
 
-    result_bin_f = medfilt(result_bin, kernel_size=11)
+    result_bin_f = medfilt(result_bin, kernel_size=21)
     reconstruct_sub(result_bin_f, block_len_ms=32)
 
     plt.plot(xaxis, result[:, 0])