def make_conversion(root,
result_dir,
checkpoint,
ut_min=91,
ut_max=100,
sp_min=91,
sp_max=100):
alpha = 0.42
n_fft = 1024
root = Path(root)
result_dir = Path(result_dir)
dicts = torch.load(checkpoint, map_location='cpu')
model = AE(dicts['config']['model'], train=False)
model.load_state_dict(dicts['model'])
model = model.eval()
for s in range(sp_min, sp_max + 1):
sp = f'jvs{s:03}'
sp_root = result_dir / sp
sp_root.mkdir(parents=True, exist_ok=True)
sp_dict_path = sp_root / 'sp_dict.pt'
if not sp_dict_path.is_file():
nonparas = list(
(root / sp /
'nonpara30/wav24kHz16bit').glob('BASIC5000_*.mcep.npy'))
index = max(enumerate(nonparas),
key=lambda p: p[1].stat().st_size)[0]
ref_mcep = nonparas[index]
ref_f0 = ref_mcep.parent / ref_mcep.stem.replace(
'.mcep', '.f0.npy')
sp_dict = extract_from(model, ref_mcep, ref_f0, sp_dict_path)
else:
sp_dict = torch.load(sp_dict_path)
for s2 in range(sp_min, sp_max + 1):
sp2 = f'jvs{s2:03}'
sp2_root = result_dir / sp2
sp2_root.mkdir(parents=True, exist_ok=True)
target_root = sp_root / sp2
target_root.mkdir(parents=True, exist_ok=True)
for u in range(ut_min, ut_max + 1):
src_mcep = root / sp2 / 'parallel100/wav24kHz16bit' / f'VOICEACTRESS100_{u:03}.mcep.npy'
src_f0 = root / sp2 / 'parallel100/wav24kHz16bit' / f'VOICEACTRESS100_{u:03}.f0.npy'
src_c0 = root / sp2 / 'parallel100/wav24kHz16bit' / f'VOICEACTRESS100_{u:03}.c0.npy'
src_ap = root / sp2 / 'parallel100/wav24kHz16bit' / f'VOICEACTRESS100_{u:03}.ap.npy'
src_dict_path = sp2_root / f'VOICEACTRESS100_{u:03}.pt'
if not src_dict_path.is_file():
src_dict = prep_content(model, src_mcep, src_dict_path)
else:
src_dict = torch.load(src_dict_path)
converted_mcep = model.reconstruct_mcep(
src_dict['c'], src_dict['q'], sp_dict['k'],
sp_dict['v']).squeeze().numpy()
tgt_mcep = target_root / f'VOICEACTRESS100_{u:03}.mcep.npy'
np.save(tgt_mcep, converted_mcep)
f0 = np.load(src_f0).astype(np.float64)
f0 = convert_f0(f0, sp_dict)
ap = np.load(src_ap).astype(np.float64)
ap = reconstruct_ap(ap)
c0 = np.load(src_c0).astype(np.float64)
assert (
c0.shape[0] <= converted_mcep.shape[-1]
), f'{s}->{s2}/{u}, {c0.shape[0]} <= {converted_mcep.shape[-1]}'
mcep = np.hstack(
[c0[:, None],
converted_mcep[:, :c0.shape[0]].T]).astype(np.float64)
sp = pysptk.mc2sp(np.ascontiguousarray(mcep), alpha, n_fft)
wav = pyworld.synthesize(f0, sp, ap, 16000)
tgt_wav = target_root / f'VOICEACTRESS100_{u:03}.wav'
wavfile.write(tgt_wav, 16000, (wav * 32768).astype(np.int16))
print(tgt_wav, flush=True)
def invert(pred):
return np.abs(1 - pred)
def save_prediction(pred, out_csv='prediction.csv'):
with open(out_csv, 'w') as f:
f.write('id,label\n')
for i, p in enumerate(pred):
f.write(f'{i},{p}\n')
print(f'Save prediction to {out_csv}.')
# load model
model = AE().cuda()
model.load_state_dict(torch.load(model_path))
model.eval()
# 準備 data
trainX = np.load(npy_path)
# 預測答案
latents = inference(X=trainX, model=model)
pred, X_embedded = predict(latents)
# 將預測結果存檔，上傳 kaggle
#save_prediction(pred, 'prediction.csv')
# 由於是 unsupervised 的二分類問題，我們只在乎有沒有成功將圖片分成兩群
# 如果上面的檔案上傳 kaggle 後正確率不足 0.5，只要將 label 反過來就行了
save_prediction(invert(pred), out)
