def model_predict(model, input_track, sample_size, trans_fw=None, trans_bw=None):
dim = sample_size
n_batches = int(len(input_track) / dim) - 1
pred_batches = input_track[0:n_batches * dim].reshape((-1, dim))
pred_batches_shifted = input_track[dim // 2:n_batches * dim + dim // 2].reshape((-1, dim))
print("FFT transforming...")
if trans_fw is not None:
pred_batches, pred_batches_shifted = map(lambda x: parallel_apply_along_axis(trans_fw, 0, x),
(pred_batches, pred_batches_shifted))
if trans_bw is None:
trans_bw = lambda x: x
xfp = x_fade_profile(dim)
print("Predicting and transforming...")
pred_batches, pred_batches_shifted = map(lambda x: parallel_apply_along_axis(trans_bw, 0, model.predict(x)),
(pred_batches, pred_batches_shifted))
print("Applying x-fade and mixing...")
x0 = np.array([xfp * batch for batch in pred_batches]).reshape(-1)
x1 = np.array([xfp * batch for batch in pred_batches_shifted]).reshape(-1)
return mix_at(x0, x1, dim // 2)
def forward_stft(data, stft_sample_size=4096, stft_stride=256, window=None):
stft_ix = np.arange(0, len(data) - stft_sample_size, stft_stride)
stft_samples = []
for ix in stft_ix:
sample = data[ix:ix + stft_sample_size]
stft_samples.append(sample)
stft_samples = np.array(stft_samples)
if window is not None:
if len(window) != stft_sample_size:
raise ValueError(f"Number of samples in window parameter should match the sample size ({stft_sample_size})")
stft_samples *= np.outer(np.ones(stft_samples.shape[0]), window)
transformed = parallel_apply_along_axis(rfft, 0, stft_samples)
return transformed
def backward_stft(data, stft_stride, window=None):
n_stft_samples = data.shape[0]
stft_sample_size = data.shape[1]
overlap_factor = stft_stride / stft_sample_size
stft_samples = parallel_apply_along_axis(irfft, 0, data)
stft_samples *= overlap_factor
out = np.zeros(stft_stride * n_stft_samples + stft_sample_size)
for ii, sample in enumerate(stft_samples):
i_begin = ii*stft_stride
i_end = i_begin + stft_sample_size
out[i_begin:i_end] += sample
return out
def apply_fourier_on_input(input_track, window_size):
n_batches = int(len(input_track) / window_size)
batches = input_track[0:n_batches*window_size].reshape((-1, window_size))
return parallel_apply_along_axis(rfft, 0, batches)
def decoder_predict(decoder_model, encoded_x, trans_bw):
return parallel_apply_along_axis(trans_bw, 0, decoder_model.predict(encoded_x)).reshape(-1)
