def test_dynamic_range_quantization(self):
self._train_and_export(
audio_spec.YAMNetSpec(keep_yamnet_and_custom_heads=True),
num_classes=5,
filename='basic_5_classes_training.tflite',
expected_model_size=5 * 1000 * 1000,
quantization_config=configs.QuantizationConfig.for_dynamic())
def test_from_esc50(self):
folder_path = self._get_folder_path('test_examples_helper')
headers = [
'filename', 'fold', 'target', 'category', 'esc10', 'src_file',
'take'
]
rows = []
rows.append(
['1-100032-A-0.wav', '1', '0', 'dog', 'True', '100032', 'A'])
rows.append([
'1-100210-B-36.wav', '2', '36', 'vacuum_cleaner', 'False',
'100210', 'B'
])
rows.append([
'1-100210-A-36.wav', '1', '36', 'vacuum_cleaner', 'False',
'100210', 'A'
])
write_csv(folder_path, 'meta', 'esc50.csv', headers, rows)
spec = audio_spec.YAMNetSpec()
loader = audio_dataloader.DataLoader.from_esc50(spec, folder_path)
self.assertEqual(len(loader), 3)
self.assertEqual(loader.index_to_label, ['dog', 'vacuum_cleaner'])
expected_results = {
'1-100032-A-0.wav': 0,
'1-100210-B-36.wav': 1,
'1-100210-A-36.wav': 1,
}
for full_path, label in loader._dataset:
filename = full_path.numpy().decode('utf-8').split('/')[-1]
self.assertEqual(expected_results[filename], label)
def test_create_model(self):
# Make sure that there is no naming conflicts in the graph.
spec = audio_spec.YAMNetSpec()
model = spec.create_model(10)
model = spec.create_model(10)
model = spec.create_model(10)
self.assertEqual(model.input_shape, (None, 1024))
self.assertEqual(model.output_shape, (None, 10))
def test_yamnet_single_head(self):
tflite_path = self._train_and_export(
audio_spec.YAMNetSpec(keep_yamnet_and_custom_heads=False),
num_classes=2,
filename='single_head.tflite',
expected_model_size=13 * 1000 * 1000)
self.assertEqual(
1, len(model_util.get_lite_runner(tflite_path).output_details))
def test_spec(self):
folder_path = self._get_folder_path('test_examples_helper')
write_sample(folder_path, 'unknown', '2s.wav', 44100, 2, value=1)
spec = audio_spec.YAMNetSpec()
audio_dataloader.DataLoader.from_folder(spec, folder_path)
spec = audio_spec.BrowserFFTSpec()
audio_dataloader.DataLoader.from_folder(spec, folder_path)
def test_spec(self):
folder_path = self._get_folder_path('test_examples_helper')
spec = audio_spec.YAMNetSpec()
audio_dataloader.DataLoader.from_esc50(spec, folder_path)
spec = audio_spec.BrowserFFTSpec()
with self.assertRaises(AssertionError):
audio_dataloader.DataLoader.from_esc50(spec, folder_path)
def test_no_metadata(self):
audio_spec.ENABLE_METADATA = False
tflite_path = self._train_and_export(
audio_spec.YAMNetSpec(keep_yamnet_and_custom_heads=True),
num_classes=2,
filename='two_heads.tflite',
expected_model_size=15 * 1000 * 1000)
self.assertEqual(
2, len(model_util.get_lite_runner(tflite_path).output_details))
with self.assertRaisesRegex(ValueError, 'The model does not have metadata'):
model_util.extract_tflite_metadata_json(tflite_path)
audio_spec.ENABLE_METADATA = True
def _test_preprocess(self, input_shape, input_count, output_shape,
output_count):
spec = audio_spec.YAMNetSpec()
wav_ds = tf.data.Dataset.from_tensor_slices([tf.ones(input_shape)] *
input_count)
label_ds = tf.data.Dataset.range(input_count).map(
lambda x: tf.cast(x, tf.int32))
ds = tf.data.Dataset.zip((wav_ds, label_ds))
ds = spec.preprocess_ds(ds)
chunks = output_count // input_count
cnt = 0
for item, label in ds:
cnt += 1
self.assertEqual(cnt, output_count)
# More thorough checks.
cnt = 0
for item, label in ds:
self.assertEqual(output_shape, item.shape)
self.assertEqual(label, cnt // chunks)
cnt += 1
def test_binary_classification(self):
self._train_and_export(
audio_spec.YAMNetSpec(keep_yamnet_and_custom_heads=True),
num_classes=2,
filename='binary_classification.tflite',
expected_model_size=15 * 1000 * 1000)
def test_yamnet_single_head(self):
tflite_path = self._train_and_export(
audio_spec.YAMNetSpec(keep_yamnet_and_custom_heads=False),
num_classes=2,
filename='single_head.tflite',
expected_model_size=13 * 1000 * 1000)
self.assertEqual(
1, len(model_util.get_lite_runner(tflite_path).output_details))
self.assertAllEqual(
[1, 2],
model_util.get_lite_runner(tflite_path).output_details[0]['shape'])
self.assertEqual(
model_util.extract_tflite_metadata_json(tflite_path), """{
"name": "yamnet/classification",
"description": "Recognizes sound events",
"version": "v1",
"subgraph_metadata": [
{
"input_tensor_metadata": [
{
"name": "audio_clip",
"description": "Input audio clip to be classified.",
"content": {
"content_properties_type": "AudioProperties",
"content_properties": {
"sample_rate": 16000,
"channels": 1
}
},
"stats": {
}
}
],
"output_tensor_metadata": [
{
"name": "custom",
"description": "Scores in range 0..1.0 for each output classes.",
"content": {
"content_properties_type": "FeatureProperties",
"content_properties": {
}
},
"stats": {
"max": [
1.0
],
"min": [
0.0
]
},
"associated_files": [
{
"name": "custom_labels.txt",
"description": "Labels for categories that the model can recognize.",
"type": "TENSOR_AXIS_LABELS"
}
]
}
]
}
],
"author": "TensorFlow Lite Model Maker",
"license": "Apache License. Version 2.0 http://www.apache.org/licenses/LICENSE-2.0.",
"min_parser_version": "1.3.0"
}
""")
def setUpClass(cls):
super(YAMNetSpecTest, cls).setUpClass()
cls._spec = audio_spec.YAMNetSpec()
def test_basic_training(self):
self._train_and_export(
audio_spec.YAMNetSpec(keep_yamnet_and_custom_heads=True),
num_classes=5,
filename='basic_5_classes_training.tflite',
expected_model_size=15 * 1000 * 1000)
def testYAMNet(self):
self._test_spec(audio_spec.YAMNetSpec(), YAMNetWithoutPreprcessing())
def test_from_esc50(self):
folder_path = self._get_folder_path('test_from_esc50')
headers = [
'filename', 'fold', 'target', 'category', 'esc10', 'src_file',
'take'
]
rows = []
rows.append(
['1-100032-A-0.wav', '1', '0', 'dog', 'True', '100032', 'A'])
rows.append([
'1-100210-B-36.wav', '2', '36', 'vacuum_cleaner', 'False',
'100210', 'B'
])
rows.append([
'1-100210-A-36.wav', '1', '36', 'vacuum_cleaner', 'False',
'100210', 'A'
])
write_csv(folder_path, 'meta', 'esc50.csv', headers, rows)
spec = audio_spec.YAMNetSpec()
loader = audio_dataloader.DataLoader.from_esc50(spec, folder_path)
self.assertEqual(len(loader), 3)
self.assertEqual(loader.index_to_label, ['dog', 'vacuum_cleaner'])
expected_results = {
'1-100032-A-0.wav': 0,
'1-100210-B-36.wav': 1,
'1-100210-A-36.wav': 1,
}
for full_path, label in loader._dataset:
filename = full_path.numpy().decode('utf-8').split('/')[-1]
self.assertEqual(expected_results[filename], label)
# fitlered dataset
with self.assertRaisesRegexp(ValueError, 'No audio files found'):
loader = audio_dataloader.DataLoader.from_esc50(spec,
folder_path,
folds=[
3,
])
with self.assertRaisesRegexp(ValueError, 'No audio files found'):
loader = audio_dataloader.DataLoader.from_esc50(
spec, folder_path, categories=['unknown'])
loader = audio_dataloader.DataLoader.from_esc50(spec,
folder_path,
folds=[
1,
])
self.assertEqual(len(loader), 2)
loader = audio_dataloader.DataLoader.from_esc50(
spec, folder_path, categories=['vacuum_cleaner'])
self.assertEqual(len(loader), 2)
loader = audio_dataloader.DataLoader.from_esc50(
spec, folder_path, folds=[
1,
], categories=['vacuum_cleaner'])
self.assertEqual(len(loader), 1)
loader = audio_dataloader.DataLoader.from_esc50(
spec, folder_path, folds=[1, 2], categories=['vacuum_cleaner'])
self.assertEqual(len(loader), 2)
