storytold/storyteller-ml
2
0
Fork 0
mirror of https://github.com/storytold/storyteller-ml.git synced 2026-10-09 00:09:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

use librosa to load more file types

Browse Source
This commit is contained in:
Paul Sajna
2023-05-12 23:03:32 -07:00
parent 2f46a8d3fd
commit da48f55c96
1 changed files with 8 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
stem_splitter/mdx_net_main/infer.py
+8 -13
View File
@@ -18,7 +18,6 @@ class MDXNetMain():
self.window = torch.hann_window(self.n_fft)
self.freq_pad = torch.zeros([1, self.dim_c, self.n_bins-self.dim_f, self.dim_t])
self.adjust = 1
self.device="cpu"
self.ort_ = ort.InferenceSession("models/UVR_MDXNET_Main.onnx")
self.model_run = lambda spek:self.ort_.run(None, {'input': spek})[0]
@@ -44,7 +43,7 @@ class MDXNetMain():
def run_model(self, mix, is_ckpt=False, is_match_mix=False):
spek = self.stft(mix)*self.adjust
spek[:, :, :3, :] *= 0
spec_pred = self.model_run(spek.cpu().float().numpy())
spec_pred = self.model_run(spek.float().numpy())
vocal = self.istft(torch.from_numpy(spec_pred)).transpose(0,1).reshape(2, -1).numpy()
return vocal
@@ -58,7 +57,11 @@ def normalize_two_stem(wave, mix, is_normalize=False):
return wave, mix
def process(input_file_path, vocal_output_path, instrumental_output_path):
wav = torchaudio.load(input_file_path)[0]
wav, _ = librosa.load(input_file_path, sr=44100, mono=False)
if wav.ndim == 1:
wav = np.array([wav, wav])
if wav.ndim != 2:
raise Exception("Only mono or stereo files are supported")
mdx = MDXNetMain()
full_size = ((wav.shape[1] // mdx.chunk_size)+1)*mdx.chunk_size
full_vocal = np.ndarray([2, full_size])
@@ -66,9 +69,9 @@ def process(input_file_path, vocal_output_path, instrumental_output_path):
for i in range(0, wav.shape[1], mdx.chunk_size):
chunk_wav_pad = np.zeros([2, mdx.chunk_size])
if i + mdx.chunk_size < wav.shape[1]:
chunk_wav_pad[:2, :mdx.chunk_size] = wav[:2, i:i+mdx.chunk_size].numpy()
chunk_wav_pad[:2, :mdx.chunk_size] = wav[:2, i:i+mdx.chunk_size]
else:
chunk_wav_pad[:2, i+mdx.chunk_size:full_size] = wav[:2, i*mdx.chunk_size:full_size].numpy()
chunk_wav_pad[:2, i+mdx.chunk_size:full_size] = wav[:2, i*mdx.chunk_size:full_size]
chunk_vocal = mdx.run_model(chunk_wav_pad)
chunk_wav_pad, chunk_vocal = normalize_two_stem(chunk_wav_pad, chunk_vocal)
chunk_inst = chunk_wav_pad - chunk_vocal
@@ -80,11 +83,3 @@ def process(input_file_path, vocal_output_path, instrumental_output_path):
torchaudio.save(vocal_output_path, torch.from_numpy(full_vocal).float(), 44100)
torchaudio.save(instrumental_output_path, torch.from_numpy(full_instrumental).float(), 44100)
if __name__ == "__main__":
process(
"/mnt/nas/media/Music/Days N' Daze/Rogue Taxidemy [REMASTER] [WAV] [LOSSLESS]/06 Blue Jays_1644.wav",
"vocals.wav",
"inst.wav"
)