def test_SumAllComponents(self):
factorization = SpleeterFactorization(self.dp, n_temporal_segments=1,
composition_fn=None,
model_name='spleeter:5stems',
spleeter_sources_path=spleeter_sources_path)
all_components = factorization.compose_model_input()
self.assertTrue(np.allclose(all_components, self.reference, atol=10**5))
def test_TemporalSegmentation(self):
n_segments = 7
factorization = SpleeterFactorization(self.dp, n_temporal_segments=n_segments,
composition_fn=None,
model_name='spleeter:5stems',
spleeter_sources_path=spleeter_sources_path)
all_components = factorization.compose_model_input()
leng = len(all_components) # to deal with ignored samples at the end
self.assertTrue(np.allclose(all_components, self.reference[:leng], atol=10 ** 5))
self.assertEqual(n_segments * 5, factorization.get_number_components()) # nr. sources = 5
def test_AnalysisWindow(self):
start = 35000
leng = 27333
reference = self.reference[start:start+leng]
factorization = SpleeterFactorization(self.dp, n_temporal_segments=1,
composition_fn=None,
model_name='spleeter:5stems',
spleeter_sources_path=spleeter_sources_path)
factorization.set_analysis_window(start, leng)
all_components = factorization.compose_model_input()
self.assertTrue(np.allclose(all_components, reference, atol=10 ** 5))
outputs = model(x)
top_tag_per_snippet = torch.argmax(outputs.detach().cpu(), axis=1)
print("top_tag_per_snippet", len(top_tag_per_snippet), top_tag_per_snippet)
sorted_args = torch.argsort(outputs.detach().cpu().mean(axis=0), descending=True)
print([tags[t] for t in sorted_args[0:3]])
top_idx = sorted_args[0].item()
sorted_snippets = torch.argsort(outputs[:, top_idx].detach().cpu()).numpy()
print("top idx:", top_idx)
print("top segments:", sorted_snippets)
predict_fn = create_predict_fn(model, config)
spleeter_factorization = SpleeterFactorization(data_provider,
n_temporal_segments=n_segments,
composition_fn=composition_fn,
model_name='spleeter:5stems')
explainer = lime_audio.LimeAudioExplainer(verbose=True, absolute_feature_sort=False)
for sn in range(len(snippet_starts)):
if args.use_global_tag:
labels = [top_idx]
else:
snippet_tag = top_tag_per_snippet[sn].item()
labels = [snippet_tag]
print("processing {}_{}".format(sample, sn))
explanation_name = "{}/{}_cls{}_sntag{}_nc{}_sn{}_seg{}_smp{}_nd{}".format(config.model_type, sample, top_idx, labels[0], config.batch_size, sn, n_segments, num_samples, n_display_components)
from audioLIME.factorization import SpleeterFactorization
from audioLIME import lime_audio
import soundfile as sf
import os
from examples.sota_utils import prepare_config, get_predict_fn
if __name__ == '__main__':
audio_path = '/share/home/verena/data/samples/3_Hop Along-SisterCities.stem.mp4_sn0_original.wav'
path_sota = '/home/verena/deployment/sota-music-tagging-models/'
config = prepare_config("fcn", 29 * 16000)
predict_fn = get_predict_fn(config)
data_provider = RawAudioProvider(audio_path)
spleeter_factorization = SpleeterFactorization(data_provider,
n_temporal_segments=10,
composition_fn=None,
model_name='spleeter:5stems')
explainer = lime_audio.LimeAudioExplainer(verbose=True, absolute_feature_sort=False)
explanation = explainer.explain_instance(factorization=spleeter_factorization,
predict_fn=predict_fn,
top_labels=1,
num_samples=16384,
batch_size=32
)
label = list(explanation.local_exp.keys())[0]
top_components, component_indeces = explanation.get_sorted_components(label,
positive_components=True,
negative_components=False,