Add files via upload

infer/infer-pm-index256.py

@@ -0,0 +1,147 @@
+'''
+
+对源特征进行检索
+'''
+import torch, pdb, os,parselmouth
+os.environ["CUDA_VISIBLE_DEVICES"]="0"
+import numpy as np
+import soundfile as sf
+# from models import SynthesizerTrn256#hifigan_nonsf
+# from infer_pack.models import SynthesizerTrn256NSF as SynthesizerTrn256#hifigan_nsf
+from infer_pack.models import SynthesizerTrnMs256NSFsid as SynthesizerTrn256#hifigan_nsf
+# from infer_pack.models import SynthesizerTrnMs256NSFsid_sim as SynthesizerTrn256#hifigan_nsf
+# from models import SynthesizerTrn256NSFsim as SynthesizerTrn256#hifigan_nsf
+# from models import SynthesizerTrn256NSFsimFlow as SynthesizerTrn256#hifigan_nsf
+
+
+from scipy.io import wavfile
+from fairseq import checkpoint_utils
+# import pyworld
+import librosa
+import torch.nn.functional as F
+import scipy.signal as signal
+# import torchcrepe
+from time import time as ttime
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model_path = r"E:\codes\py39\vits_vc_gpu_train\hubert_base.pt"#
+print("load model(s) from {}".format(model_path))
+models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
+    [model_path],
+    suffix="",
+)
+model = models[0]
+model = model.to(device)
+model = model.half()
+model.eval()
+
+# net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],183,256,is_half=True)#hifigan#512#256
+# net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],109,256,is_half=True)#hifigan#512#256
+net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],183,256,is_half=True)#hifigan#512#256#no_dropout
+# net_g = SynthesizerTrn256(1025,32,192,192,768,2,3,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],0)#ts3
+# net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2],512,[16,16,4],0)#hifigan-ps-sr
+#
+# net_g = SynthesizerTrn(1025, 32, 192, 192, 768, 2, 6, 3, 0.1, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [5,5], 512, [15,15], 0)#ms
+# net_g = SynthesizerTrn(1025, 32, 192, 192, 768, 2, 6, 3, 0.1, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10,10], 512, [16,16], 0)#idwt2
+
+# weights=torch.load("infer/ft-mi_1k-noD.pt")
+# weights=torch.load("infer/ft-mi-freeze-vocoder-flow-enc_q_1k.pt")
+# weights=torch.load("infer/ft-mi-freeze-vocoder_true_1k.pt")
+# weights=torch.load("infer/ft-mi-sim1k.pt")
+weights=torch.load("infer/ft-mi-no_opt-no_dropout.pt")
+print(net_g.load_state_dict(weights,strict=True))
+
+net_g.eval().to(device)
+net_g.half()
+def get_f0(x, p_len,f0_up_key=0):
+
+    time_step = 160 / 16000 * 1000
+    f0_min = 50
+    f0_max = 1100
+    f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+    f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+
+    f0 = parselmouth.Sound(x, 16000).to_pitch_ac(
+        time_step=time_step / 1000, voicing_threshold=0.6,
+        pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
+
+    pad_size=(p_len - len(f0) + 1) // 2
+    if(pad_size>0 or p_len - len(f0) - pad_size>0):
+        f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
+    f0 *= pow(2, f0_up_key / 12)
+    f0bak = f0.copy()
+
+    f0_mel = 1127 * np.log(1 + f0 / 700)
+    f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1
+    f0_mel[f0_mel <= 1] = 1
+    f0_mel[f0_mel > 255] = 255
+    # f0_mel[f0_mel > 188] = 188
+    f0_coarse = np.rint(f0_mel).astype(np.int)
+    return f0_coarse, f0bak
+
+import faiss
+index=faiss.read_index("infer/added_IVF512_Flat_mi_baseline_src_feat.index")
+big_npy=np.load("infer/big_src_feature_mi.npy")
+ta0=ta1=ta2=0
+for idx,name in enumerate(["冬之花clip1.wav",]):##
+    wav_path = "todo-songs/%s" % name#
+    f0_up_key=-2#
+    audio, sampling_rate = sf.read(wav_path)
+    if len(audio.shape) > 1:
+        audio = librosa.to_mono(audio.transpose(1, 0))
+    if sampling_rate != 16000:
+        audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
+
+
+    feats = torch.from_numpy(audio).float()
+    if feats.dim() == 2:  # double channels
+        feats = feats.mean(-1)
+    assert feats.dim() == 1, feats.dim()
+    feats = feats.view(1, -1)
+    padding_mask = torch.BoolTensor(feats.shape).fill_(False)
+    inputs = {
+        "source": feats.half().to(device),
+        "padding_mask": padding_mask.to(device),
+        "output_layer": 9,  # layer 9
+    }
+    torch.cuda.synchronize()
+    t0=ttime()
+    with torch.no_grad():
+        logits = model.extract_features(**inputs)
+        feats = model.final_proj(logits[0])
+
+    ####索引优化
+    npy = feats[0].cpu().numpy().astype("float32")
+    D, I = index.search(npy, 1)
+    feats = torch.from_numpy(big_npy[I.squeeze()].astype("float16")).unsqueeze(0).to(device)
+
+    feats=F.interpolate(feats.permute(0,2,1),scale_factor=2).permute(0,2,1)
+    torch.cuda.synchronize()
+    t1=ttime()
+    # p_len = min(feats.shape[1],10000,pitch.shape[0])#太大了爆显存
+    p_len = min(feats.shape[1],10000)#
+    pitch, pitchf = get_f0(audio, p_len,f0_up_key)
+    p_len = min(feats.shape[1],10000,pitch.shape[0])#太大了爆显存
+    torch.cuda.synchronize()
+    t2=ttime()
+    feats = feats[:,:p_len, :]
+    pitch = pitch[:p_len]
+    pitchf = pitchf[:p_len]
+    p_len = torch.LongTensor([p_len]).to(device)
+    pitch = torch.LongTensor(pitch).unsqueeze(0).to(device)
+    sid=torch.LongTensor([0]).to(device)
+    pitchf = torch.FloatTensor(pitchf).unsqueeze(0).to(device)
+    with torch.no_grad():
+        audio = net_g.infer(feats, p_len,pitch,pitchf,sid)[0][0, 0].data.cpu().float().numpy()#nsf
+    torch.cuda.synchronize()
+    t3=ttime()
+    ta0+=(t1-t0)
+    ta1+=(t2-t1)
+    ta2+=(t3-t2)
+    # wavfile.write("ft-mi_1k-index256-noD-%s.wav"%name, 40000, audio)##
+    # wavfile.write("ft-mi-freeze-vocoder-flow-enc_q_1k-%s.wav"%name, 40000, audio)##
+    # wavfile.write("ft-mi-sim1k-%s.wav"%name, 40000, audio)##
+    wavfile.write("ft-mi-no_opt-no_dropout-%s.wav"%name, 40000, audio)##
+
+
+print(ta0,ta1,ta2)#

infer/train-index.py

@@ -0,0 +1,36 @@
+'''
+格式：直接cid为自带的index位；aid放不下了，通过字典来查，反正就5w个
+'''
+import faiss,numpy as np,os
+
+# ###########如果是原始特征要先写save
+inp_root=r"E:\codes\py39\dataset\mi\2-co256"
+npys=[]
+for name in sorted(list(os.listdir(inp_root))):
+    phone=np.load("%s/%s"%(inp_root,name))
+    npys.append(phone)
+big_npy=np.concatenate(npys,0)
+print(big_npy.shape)#(6196072, 192)#fp32#4.43G
+np.save("infer/big_src_feature_mi.npy",big_npy)
+
+##################train+add
+# big_npy=np.load("/bili-coeus/jupyter/jupyterhub-liujing04/vits_ch/inference_f0/big_src_feature_mi.npy")
+print(big_npy.shape)
+index = faiss.index_factory(256, "IVF512,Flat")#mi
+print("training")
+index_ivf = faiss.extract_index_ivf(index)#
+index_ivf.nprobe = 9
+index.train(big_npy)
+faiss.write_index(index, 'infer/trained_IVF512_Flat_mi_baseline_src_feat.index')
+print("adding")
+index.add(big_npy)
+faiss.write_index(index,"infer/added_IVF512_Flat_mi_baseline_src_feat.index")
+'''
+大小（都是FP32）
+big_src_feature 2.95G
+    (3098036, 256)
+big_emb         4.43G
+    (6196072, 192)
+big_emb双倍是因为求特征要repeat后再加pitch
+
+'''

infer/trans_weights.py

@@ -0,0 +1,11 @@
+import torch,pdb
+
+# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-suc\G_1000.pth")["model"]#sim_nsf#
+# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-freeze-vocoder-flow-enc_q\G_1000.pth")["model"]#sim_nsf#
+# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-freeze-vocoder\G_1000.pth")["model"]#sim_nsf#
+# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-test\G_1000.pth")["model"]#sim_nsf#
+a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-no_opt-no_dropout\G_1000.pth")["model"]#sim_nsf#
+for key in a.keys():a[key]=a[key].half()
+# torch.save(a,"ft-mi-freeze-vocoder_true_1k.pt")#
+# torch.save(a,"ft-mi-sim1k.pt")#
+torch.save(a,"ft-mi-no_opt-no_dropout.pt")#