def __init__(self, model_name: str, chunk_size: int = -1, runner_cls: type = None):
self.window_audio = np.array([])
self.pr = inject_params(model_name)
self.mfccs = np.zeros((self.pr.n_features, self.pr.n_mfcc))
self.chunk_size = chunk_size
runner_cls = runner_cls or self.find_runner(model_name)
self.runner = runner_cls(model_name)
self.threshold_decoder = ThresholdDecoder(self.pr.threshold_config, pr.threshold_center)
def run(self):
args = self.args
if args.models:
data = TrainData.from_both(args.tags_file, args.tags_folder,
args.folder)
print('Data:', data)
filenames = sum(
data.train_files if args.use_train else data.test_files, [])
loader = CachedDataLoader(
partial(data.load,
args.use_train,
not args.use_train,
shuffle=False))
model_data = calc_stats(args.models, loader, args.use_train,
filenames)
else:
model_data = {
name: Stats.from_np_dict(data)
for name, data in np.load(args.input_file)
['data'].item().items()
}
for name, stats in model_data.items():
print('=== {} ===\n{}\n\n{}\n'.format(name, stats.counts_str(),
stats.summary_str()))
if args.output_file:
np.savez(args.output_file,
data={
name: stats.to_np_dict()
for name, stats in model_data.items()
})
else:
plt = load_plt()
decoder = ThresholdDecoder(pr.threshold_config,
pr.threshold_center)
thresholds = [
decoder.encode(i)
for i in np.linspace(0.0, 1.0, args.resolution)[1:-1]
]
for model_name, stats in model_data.items():
x = [stats.false_positives(i) for i in thresholds]
y = [stats.false_negatives(i) for i in thresholds]
plt.plot(x, y, marker='x', linestyle='-', label=model_name)
if args.labels:
for x, y, threshold in zip(x, y, thresholds):
plt.annotate('{:.4f}'.format(threshold), (x, y))
plt.legend()
plt.xlabel('False Positives')
plt.ylabel('False Negatives')
plt.show()
class Listener:
"""Listener that preprocesses audio into MFCC vectors and executes neural networks"""
def __init__(self,
model_name: str,
chunk_size: int = -1,
runner_cls: type = None):
self.window_audio = np.array([])
self.pr = inject_params(model_name)
self.mfccs = np.zeros((self.pr.n_features, self.pr.n_mfcc))
self.chunk_size = chunk_size
runner_cls = runner_cls or self.find_runner(model_name)
self.runner = runner_cls(model_name)
self.threshold_decoder = ThresholdDecoder(self.pr.threshold_config,
pr.threshold_center)
@staticmethod
def find_runner(model_name: str) -> Type[Runner]:
runners = {'.net': KerasRunner, '.pb': TensorFlowRunner}
ext = splitext(model_name)[-1]
if ext not in runners:
raise ValueError('File extension of ' + model_name + ' must be: ' +
str(list(runners)))
return runners[ext]
def clear(self):
self.window_audio = np.array([])
self.mfccs = np.zeros((self.pr.n_features, self.pr.n_mfcc))
def update_vectors(
self, stream: Union[BinaryIO, np.ndarray, bytes]) -> np.ndarray:
if isinstance(stream, np.ndarray):
buffer_audio = stream
else:
if isinstance(stream, (bytes, bytearray)):
chunk = stream
else:
chunk = stream.read(self.chunk_size)
if len(chunk) == 0:
raise EOFError
buffer_audio = buffer_to_audio(chunk)
self.window_audio = np.concatenate((self.window_audio, buffer_audio))
if len(self.window_audio) >= self.pr.window_samples:
new_features = vectorize_raw(self.window_audio)
self.window_audio = self.window_audio[len(new_features) *
self.pr.hop_samples:]
if len(new_features) > len(self.mfccs):
new_features = new_features[-len(self.mfccs):]
self.mfccs = np.concatenate(
(self.mfccs[len(new_features):], new_features))
return self.mfccs
def update(self, stream: Union[BinaryIO, np.ndarray, bytes]) -> float:
mfccs = self.update_vectors(stream)
if self.pr.use_delta:
mfccs = add_deltas(mfccs)
raw_output = self.runner.run(mfccs)
return self.threshold_decoder.decode(raw_output)
def main():
parser = create_parser(usage)
parser.add_argument(
'models',
nargs='*',
help='Either Keras (.net) or TensorFlow (.pb) models to test')
args = TrainData.parse_args(parser)
if not args.models and not args.input_file and args.folder:
args.input_file = args.folder
if bool(args.models) == bool(args.input_file):
parser.error('Please specify either a list of models or an input file')
if not args.output_file:
load_plt() # Error early if matplotlib not installed
import numpy as np
if args.models:
data = TrainData.from_both(args.tags_file, args.tags_folder,
args.folder)
print('Data:', data)
filenames = sum(
data.train_files if args.use_train else data.test_files, [])
loader = CachedDataLoader(
partial(data.load,
args.use_train,
not args.use_train,
shuffle=False))
model_data = calc_stats(args.models, loader, args.use_train, filenames)
else:
model_data = {
name: Stats.from_np_dict(data)
for name, data in np.load(args.input_file)['data'].item().items()
}
for name, stats in model_data.items():
print('=== {} ===\n{}\n\n{}\n'.format(name, stats.counts_str(),
stats.summary_str()))
if args.output_file:
np.savez(args.output_file,
data={
name: stats.to_np_dict()
for name, stats in model_data.items()
})
else:
plt = load_plt()
decoder = ThresholdDecoder(pr.threshold_config, pr.threshold_center)
thresholds = [
decoder.encode(i)
for i in np.linspace(0.0, 1.0, args.resolution)[1:-1]
]
for model_name, stats in model_data.items():
x = [stats.false_positives(i) for i in thresholds]
y = [stats.false_negatives(i) for i in thresholds]
plt.plot(x, y, marker='x', linestyle='-', label=model_name)
if args.labels:
for x, y, threshold in zip(x, y, thresholds):
plt.annotate('{:.4f}'.format(threshold), (x, y))
plt.legend()
plt.xlabel('False Positives')
plt.ylabel('False Negatives')
plt.show()