def predict_one(items, outs, model_names, config):
data, beattheta, bpm_label, song_name = items
theta, idxs = numpy_fft(outs[None, :] * 2 * np.pi, 100)
plot_beat_and_bpm2(
# theta[0],
# {"beat": outs, "beat_label": beattheta},
{
"before": outs,
"after": theta[0]
},
song_name=song_name,
model_names=model_names,
time_range=(0, 6000),
)
k = 200
sin = np.sin(outs * 2 * np.pi)
spectrogram = [
np.abs(np.fft.fft(sin[i:i + k]))[:k // 2]
for i in range(outs.shape[0] - k)
]
spectrogram = np.stack(spectrogram)
# plt.plot(np.abs(np.fft.fft(sin))[:3000])
print(np.argmax(spectrogram[4]))
plt.imshow(spectrogram.T, aspect=40, origin="lower")
# plt.plot(spectrogram[5])
plt.show()
def evaluate(items, outs, config, trial=None):
# eval_beat = EvaluateBeat(estimated_type="index", target_type="activation")
eval_downbeat = EvaluateDownbeat(
estimated_type="activation", target_type="activation")
data, beat_label, downbeattheta_label = items
theta, idxs = numpy_fft(outs * 2 * np.pi, 100)
# act = (outs[:, :-1] - outs[:, 1:]) / (2 * np.pi)
# act[act <= 0] = 1e-5
beat_metrics = eval_downbeat.batch(outs, downbeattheta_label)
ret = {"beat": beat_metrics}
return ret
def predict_one(items, outs, model_names, config):
data, beattheta, downbeat_label, song_name = items
beat, downbeat = outs
plt.plot(beat)
plt.plot(downbeat)
plt.show()
theta, idxs = numpy_fft(beat[None, :] * 2 * np.pi, 400)
plot_beat_and_bpm2(
# theta[0],
{"beat": beat, "downbeat_label": downbeat_label, "downbeat": downbeat},
# {"downbeat": theta[0]/2/np.pi, "label": beattheta},
song_name=song_name,
model_names=model_names,
time_range=(0, 6000),
)
def predict_one(items, outs, model_names, config):
data, bpm_label, beattheta, song_name = items
beattheta_out, diffbeattheta = outs
diffbeattheta = medfilt(diffbeattheta, kernel_size=21)
beattheta_out = np.argmax(beattheta_out.T, axis=-1) / 6000
theta = numpy_fft(beattheta_out[None, :] * 2 * np.pi, 100)
theta /= 2 * np.pi
activation = beattheta_out[:-1] - beattheta_out[1:]
activation[activation <= 0] = 1e-6
bpm_out = diffbeattheta * 3000 / np.pi
# plot_beat_and_bpm(
# theta[0], beattheta, bpm_label, bpm_out, song_name, model_names,
# )
plt.plot(np.sin(beattheta_out * 2 * np.pi))
# plt.plot(theta[0], alpha=0.7)
plt.plot(np.sin(beattheta * 2 * np.pi), alpha=0.7)
plt.show()