refactored the ml.rnn to ml.nn and added tests; updated the models su…

…bmodule
CPJKU · Mar 9, 2016 · be3b0f9 · be3b0f9
1 parent 4ef285d
commit be3b0f9
Show file tree

Hide file tree

Showing 20 changed files with 845 additions and 851 deletions.
diff --git a/bin/BeatDetector b/bin/BeatDetector
@@ -17,7 +17,7 @@ from madmom.processors import (SequentialProcessor, ParallelProcessor,
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.spectrogram import (LogarithmicFilteredSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor)
-from madmom.ml.rnn import RNNProcessor, average_predictions
+from madmom.ml.nn import NeuralNetwork, average_predictions
 from madmom.features import ActivationsProcessor
 from madmom.features.beats import BeatDetectionProcessor
 from madmom.features.tempo import TempoEstimationProcessor
@@ -95,7 +95,7 @@ def main():
     args.diff_ratio = 0.5
     args.positive_diffs = True
     args.stack_diffs = np.hstack
-    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].npz" %
+    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].pkl" %
                               MODELS_PATH)
 
     # print arguments
@@ -120,8 +120,8 @@ def main():
         # wrap everything in a ParallelProcessor and stack the features
         multi = ParallelProcessor(multi)
         stack = np.hstack
-        # process everything with an RNN and average the predictions
-        rnn = RNNProcessor(**vars(args))
+        # process everything with multiple RNNs and average the predictions
+        rnn = ParallelProcessor([NeuralNetwork.load(f) for f in args.nn_files])
         avg = average_predictions
         # sequentially process everything
         in_processor = [sig, multi, stack, rnn, avg]

diff --git a/bin/BeatTracker b/bin/BeatTracker
@@ -17,7 +17,7 @@ from madmom.processors import (SequentialProcessor, ParallelProcessor,
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.spectrogram import (LogarithmicFilteredSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor)
-from madmom.ml.rnn import RNNProcessor, average_predictions
+from madmom.ml.nn import NeuralNetwork, average_predictions
 from madmom.features import ActivationsProcessor
 from madmom.features.beats import BeatTrackingProcessor
 from madmom.features.tempo import TempoEstimationProcessor
@@ -94,7 +94,7 @@ def main():
     args.diff_ratio = 0.5
     args.positive_diffs = True
     args.stack_diffs = np.hstack
-    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].npz" %
+    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].pkl" %
                               MODELS_PATH)
 
     # print arguments
@@ -119,8 +119,8 @@ def main():
         # wrap everything in a ParallelProcessor and stack the features
         multi = ParallelProcessor(multi)
         stack = np.hstack
-        # process everything with an RNN and average the predictions
-        rnn = RNNProcessor(**vars(args))
+        # process everything with multiple RNNs and average the predictions
+        rnn = ParallelProcessor([NeuralNetwork.load(f) for f in args.nn_files])
         avg = average_predictions
         # sequentially process everything
         in_processor = [sig, multi, stack, rnn, avg]

diff --git a/bin/CRFBeatDetector b/bin/CRFBeatDetector
@@ -17,7 +17,7 @@ from madmom.processors import (SequentialProcessor, ParallelProcessor,
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.spectrogram import (LogarithmicFilteredSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor)
-from madmom.ml.rnn import RNNProcessor, average_predictions
+from madmom.ml.nn import NeuralNetwork, average_predictions
 from madmom.features import ActivationsProcessor
 from madmom.features.beats import CRFBeatDetectionProcessor
 from madmom.features.tempo import TempoEstimationProcessor
@@ -97,7 +97,7 @@ def main():
     args.diff_ratio = 0.5
     args.positive_diffs = True
     args.stack_diffs = np.hstack
-    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].npz" %
+    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].pkl" %
                               MODELS_PATH)
 
     # print arguments
@@ -122,8 +122,8 @@ def main():
         # wrap everything in a ParallelProcessor and stack the features
         multi = ParallelProcessor(multi)
         stack = np.hstack
-        # process everything with an RNN and average the predictions
-        rnn = RNNProcessor(**vars(args))
+        # process everything with multiple RNNs and average the predictions
+        rnn = ParallelProcessor([NeuralNetwork.load(f) for f in args.nn_files])
         avg = average_predictions
         # sequentially process everything
         in_processor = [sig, multi, stack, rnn, avg]

diff --git a/bin/DBNBeatTracker b/bin/DBNBeatTracker
@@ -17,7 +17,7 @@ from madmom.processors import (SequentialProcessor, ParallelProcessor,
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.spectrogram import (LogarithmicFilteredSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor)
-from madmom.ml.rnn import RNNProcessor, average_predictions
+from madmom.ml.nn import NeuralNetwork, average_predictions
 from madmom.features import ActivationsProcessor
 from madmom.features.beats import DBNBeatTrackingProcessor
 
@@ -95,7 +95,7 @@ def main():
     args.diff_ratio = 0.5
     args.positive_diffs = True
     args.stack_diffs = np.hstack
-    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].npz" %
+    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].pkl" %
                               MODELS_PATH)
 
     # print arguments
@@ -120,8 +120,8 @@ def main():
         # wrap everything in a ParallelProcessor and stack the features
         multi = ParallelProcessor(multi)
         stack = np.hstack
-        # process everything with an RNN and average the predictions
-        rnn = RNNProcessor(**vars(args))
+        # process everything with multiple RNNs and average the predictions
+        rnn = ParallelProcessor([NeuralNetwork.load(f) for f in args.nn_files])
         avg = average_predictions
         # sequentially process everything
         in_processor = [sig, multi, stack, rnn, avg]

diff --git a/bin/MMBeatTracker b/bin/MMBeatTracker
@@ -17,7 +17,7 @@ from madmom.processors import (SequentialProcessor, ParallelProcessor,
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.spectrogram import (LogarithmicFilteredSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor)
-from madmom.ml.rnn import RNNProcessor
+from madmom.ml.nn import NeuralNetwork
 from madmom.features import ActivationsProcessor
 from madmom.features.beats import (DBNBeatTrackingProcessor,
                                    MultiModelSelectionProcessor)
@@ -102,7 +102,7 @@ def main():
     args.diff_ratio = 0.5
     args.positive_diffs = True
     args.stack_diffs = np.hstack
-    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].npz" %
+    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].pkl" %
                               MODELS_PATH)
 
     # print arguments
@@ -127,8 +127,8 @@ def main():
         # wrap everything in a ParallelProcessor and stack the features
         multi = ParallelProcessor(multi)
         stack = np.hstack
-        # process everything with an RNN and select the best predictions
-        rnn = RNNProcessor(**vars(args))
+        # process everything with multiple RNNs and average the predictions
+        rnn = ParallelProcessor([NeuralNetwork.load(f) for f in args.nn_files])
         if args.nn_ref_files is None:
             # set the nn_ref_files the same as the nn_files, i.e. average them
             args.nn_ref_files = args.nn_files

diff --git a/bin/OnsetDetector b/bin/OnsetDetector
@@ -17,7 +17,7 @@ from madmom.processors import (SequentialProcessor, ParallelProcessor,
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.spectrogram import (LogarithmicFilteredSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor)
-from madmom.ml.rnn import RNNProcessor, average_predictions
+from madmom.ml.nn import NeuralNetwork, average_predictions
 from madmom.features import ActivationsProcessor
 from madmom.features.onsets import PeakPickingProcessor
 
@@ -85,7 +85,7 @@ def main():
     args.diff_ratio = 0.25
     args.positive_diffs = True
     args.stack_diffs = np.hstack
-    args.nn_files = glob.glob("%s/onsets/2013/onsets_brnn_[1-8].npz" %
+    args.nn_files = glob.glob("%s/onsets/2013/onsets_brnn_[1-8].pkl" %
                               MODELS_PATH)
     args.pre_max = 1. / args.fps
     args.post_max = 1. / args.fps
@@ -112,8 +112,8 @@ def main():
         # wrap everything in a ParallelProcessor and stack the features
         multi = ParallelProcessor(multi)
         stack = np.hstack
-        # process everything with an RNN and average the predictions
-        rnn = RNNProcessor(**vars(args))
+        # process everything with multiple RNNs and average the predictions
+        rnn = ParallelProcessor([NeuralNetwork.load(f) for f in args.nn_files])
         avg = average_predictions
         # sequentially process everything
         in_processor = [sig, multi, stack, rnn, avg]

diff --git a/bin/OnsetDetectorLL b/bin/OnsetDetectorLL
@@ -17,7 +17,7 @@ from madmom.processors import (SequentialProcessor, ParallelProcessor,
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.spectrogram import (LogarithmicFilteredSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor)
-from madmom.ml.rnn import RNNProcessor, average_predictions
+from madmom.ml.nn import NeuralNetwork, average_predictions
 from madmom.features import ActivationsProcessor
 from madmom.features.onsets import PeakPickingProcessor
 
@@ -96,7 +96,7 @@ def main():
     args.diff_ratio = 0.25
     args.positive_diffs = True
     args.stack_diffs = np.hstack
-    args.nn_files = glob.glob("%s/onsets/2013/onsets_rnn_[1-8].npz" %
+    args.nn_files = glob.glob("%s/onsets/2013/onsets_rnn_[1-8].pkl" %
                               MODELS_PATH)
     args.pre_max = 1. / args.fps
     args.post_max = 0
@@ -124,8 +124,8 @@ def main():
         # wrap everything in a ParallelProcessor and stack the features
         multi = ParallelProcessor(multi)
         stack = np.hstack
-        # process everything with an RNN and average the predictions
-        rnn = RNNProcessor(**vars(args))
+        # process everything with multiple RNNs and average the predictions
+        rnn = ParallelProcessor([NeuralNetwork.load(f) for f in args.nn_files])
         avg = average_predictions
         # sequentially process everything
         in_processor = [sig, multi, stack, rnn, avg]

diff --git a/bin/PianoTranscriptor b/bin/PianoTranscriptor
@@ -17,11 +17,10 @@ from madmom.processors import (SequentialProcessor, ParallelProcessor,
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.spectrogram import (LogarithmicFilteredSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor)
-from madmom.ml.rnn import RNNProcessor, average_predictions
+from madmom.ml.nn import NeuralNetwork
 from madmom.features import ActivationsProcessor
 from madmom.features.onsets import PeakPickingProcessor
-from madmom.features.notes import (write_midi, write_notes, write_mirex_format,
-                                   note_reshaper)
+from madmom.features.notes import write_midi, write_notes, write_mirex_format
 
 
 def main():
@@ -95,7 +94,7 @@ def main():
     args.diff_ratio = 0.5
     args.positive_diffs = True
     args.stack_diffs = np.hstack
-    args.nn_files = glob.glob("%s/notes/2013/notes_brnn.npz" % MODELS_PATH)
+    args.nn_file = "%s/notes/2013/notes_brnn.pkl" % MODELS_PATH
     args.pre_max = 1. / args.fps
     args.post_max = 1. / args.fps
 
@@ -125,12 +124,10 @@ def main():
         # wrap everything in a ParallelProcessor and stack the features
         multi = ParallelProcessor(multi)
         stack = np.hstack
-        # process everything with an RNN and average the predictions
-        rnn = RNNProcessor(**vars(args))
-        avg = average_predictions
-        reshape = note_reshaper
+        # process everything with a RNN
+        rnn = NeuralNetwork.load(args.nn_file)
         # sequentially process everything
-        in_processor = [sig, multi, stack, rnn, avg, reshape]
+        in_processor = [sig, multi, stack, rnn]
 
     # output processor
     if args.save:

diff --git a/bin/SuperFluxNN b/bin/SuperFluxNN
@@ -11,13 +11,13 @@ import glob
 import argparse
 
 from madmom import MODELS_PATH
-from madmom.processors import IOProcessor, io_arguments
+from madmom.processors import ParallelProcessor, IOProcessor, io_arguments
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.filters import FilterbankProcessor
 from madmom.audio.spectrogram import (LogarithmicSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor,
                                       SuperFluxProcessor)
-from madmom.ml.rnn import RNNProcessor, average_predictions
+from madmom.ml.nn import NeuralNetwork, average_predictions
 from madmom.features import ActivationsProcessor
 from madmom.features.onsets import SpectralOnsetProcessor, PeakPickingProcessor
 
@@ -82,7 +82,7 @@ def main():
     args.fps = 100
     args.online = False
     args.onset_method = 'superflux'
-    args.nn_files = glob.glob("%s/onsets/2014/onsets_brnn_pp_[1-8].npz" %
+    args.nn_files = glob.glob("%s/onsets/2014/onsets_brnn_pp_[1-8].pkl" %
                               MODELS_PATH)
 
     # print arguments
@@ -106,9 +106,10 @@ def main():
         # save the Onset activations to file
         out_processor = ActivationsProcessor(mode='w', **vars(args))
     else:
-        # perform RNN processing, averaging and peak-picking
-        rnn = RNNProcessor(**vars(args))
+        # process everything with multiple RNNs and average the predictions
+        rnn = ParallelProcessor([NeuralNetwork.load(f) for f in args.nn_files])
         avg = average_predictions
+        # perform peak picking, i.e. select the local maxima
         pp = PeakPickingProcessor(**vars(args))
         from madmom.utils import write_events as writer
         # sequentially process them

diff --git a/bin/TempoDetector b/bin/TempoDetector
@@ -17,7 +17,7 @@ from madmom.processors import (SequentialProcessor, ParallelProcessor,
 from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
 from madmom.audio.spectrogram import (LogarithmicFilteredSpectrogramProcessor,
                                       SpectrogramDifferenceProcessor)
-from madmom.ml.rnn import RNNProcessor, average_predictions
+from madmom.ml.nn import NeuralNetwork, average_predictions
 from madmom.features import ActivationsProcessor
 from madmom.features.tempo import TempoEstimationProcessor, write_tempo
 
@@ -90,7 +90,7 @@ def main():
     args.diff_ratio = 0.5
     args.positive_diffs = True
     args.stack_diffs = np.hstack
-    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].npz" %
+    args.nn_files = glob.glob("%s/beats/2013/beats_blstm_[1-8].pkl" %
                               MODELS_PATH)
 
     # print arguments
@@ -115,8 +115,8 @@ def main():
         # wrap everything in a ParallelProcessor and stack the features
         multi = ParallelProcessor(multi)
         stack = np.hstack
-        # process everything with an RNN and average the predictions
-        rnn = RNNProcessor(**vars(args))
+        # process everything with multiple RNNs and average the predictions
+        rnn = ParallelProcessor([NeuralNetwork.load(f) for f in args.nn_files])
         avg = average_predictions
         # sequentially process everything
         in_processor = [sig, multi, stack, rnn, avg]

diff --git a/madmom/ml/__init__.py b/madmom/ml/__init__.py
@@ -7,4 +7,4 @@
 from __future__ import absolute_import, division, print_function
 
 # import the submodules
-from . import rnn, hmm, gmm
+from . import nn, hmm, gmm