train 1-2b

infer/lib/train/process_ckpt.py

@@ -1,7 +1,6 @@
 import torch, traceback, os, sys
 
-now_dir = os.getcwd()
-sys.path.append(now_dir)
+
 from collections import OrderedDict
 from i18n.i18n import I18nAuto
 

infer/lib/train/utils.py

@@ -362,9 +362,9 @@ def get_hparams(init=True):
         os.makedirs(experiment_dir)
 
     if args.version == "v1" or args.sample_rate == "40k":
-        config_path = "configs/%s.json" % args.sample_rate
+        config_path = "configs/v1/%s.json" % args.sample_rate
     else:
-        config_path = "configs/%s_v2.json" % args.sample_rate
+        config_path = "configs/v2/%s.json" % args.sample_rate
     config_save_path = os.path.join(experiment_dir, "config.json")
     if init:
         with open(config_path, "r") as f:

infer/modules/train/extract/extract_f0_print.py

@@ -0,0 +1,168 @@
+import os, traceback, sys, parselmouth
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+from lib.audio import load_audio
+import pyworld
+import numpy as np, logging
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+from multiprocessing import Process
+
+exp_dir = sys.argv[1]
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+
+
+n_p = int(sys.argv[2])
+f0method = sys.argv[3]
+
+
+class FeatureInput(object):
+    def __init__(self, samplerate=16000, hop_size=160):
+        self.fs = samplerate
+        self.hop = hop_size
+
+        self.f0_bin = 256
+        self.f0_max = 1100.0
+        self.f0_min = 50.0
+        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+
+    def compute_f0(self, path, f0_method):
+        x = load_audio(path, self.fs)
+        p_len = x.shape[0] // self.hop
+        if f0_method == "pm":
+            time_step = 160 / 16000 * 1000
+            f0_min = 50
+            f0_max = 1100
+            f0 = (
+                parselmouth.Sound(x, self.fs)
+                .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"
+                )
+        elif f0_method == "harvest":
+            f0, t = pyworld.harvest(
+                x.astype(np.double),
+                fs=self.fs,
+                f0_ceil=self.f0_max,
+                f0_floor=self.f0_min,
+                frame_period=1000 * self.hop / self.fs,
+            )
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
+        elif f0_method == "dio":
+            f0, t = pyworld.dio(
+                x.astype(np.double),
+                fs=self.fs,
+                f0_ceil=self.f0_max,
+                f0_floor=self.f0_min,
+                frame_period=1000 * self.hop / self.fs,
+            )
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
+        elif f0_method == "rmvpe":
+            if hasattr(self, "model_rmvpe") == False:
+                from lib.rmvpe import RMVPE
+
+                print("loading rmvpe model")
+                self.model_rmvpe = RMVPE(
+                    "assets/rmvpe/rmvpe.pt", is_half=False, device="cpu"
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+        return f0
+
+    def coarse_f0(self, f0):
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
+            self.f0_bin - 2
+        ) / (self.f0_mel_max - self.f0_mel_min) + 1
+
+        # use 0 or 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
+        f0_coarse = np.rint(f0_mel).astype(int)
+        assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+            f0_coarse.max(),
+            f0_coarse.min(),
+        )
+        return f0_coarse
+
+    def go(self, paths, f0_method):
+        if len(paths) == 0:
+            printt("no-f0-todo")
+        else:
+            printt("todo-f0-%s" % len(paths))
+            n = max(len(paths) // 5, 1)  # 每个进程最多打印5条
+            for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
+                try:
+                    if idx % n == 0:
+                        printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
+                    if (
+                        os.path.exists(opt_path1 + ".npy") == True
+                        and os.path.exists(opt_path2 + ".npy") == True
+                    ):
+                        continue
+                    featur_pit = self.compute_f0(inp_path, f0_method)
+                    np.save(
+                        opt_path2,
+                        featur_pit,
+                        allow_pickle=False,
+                    )  # nsf
+                    coarse_pit = self.coarse_f0(featur_pit)
+                    np.save(
+                        opt_path1,
+                        coarse_pit,
+                        allow_pickle=False,
+                    )  # ori
+                except:
+                    printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
+
+
+if __name__ == "__main__":
+    # exp_dir=r"E:\codes\py39\dataset\mi-test"
+    # n_p=16
+    # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+    printt(sys.argv)
+    featureInput = FeatureInput()
+    paths = []
+    inp_root = "%s/1_16k_wavs" % (exp_dir)
+    opt_root1 = "%s/2a_f0" % (exp_dir)
+    opt_root2 = "%s/2b-f0nsf" % (exp_dir)
+
+    os.makedirs(opt_root1, exist_ok=True)
+    os.makedirs(opt_root2, exist_ok=True)
+    for name in sorted(list(os.listdir(inp_root))):
+        inp_path = "%s/%s" % (inp_root, name)
+        if "spec" in inp_path:
+            continue
+        opt_path1 = "%s/%s" % (opt_root1, name)
+        opt_path2 = "%s/%s" % (opt_root2, name)
+        paths.append([inp_path, opt_path1, opt_path2])
+
+    ps = []
+    for i in range(n_p):
+        p = Process(
+            target=featureInput.go,
+            args=(
+                paths[i::n_p],
+                f0method,
+            ),
+        )
+        ps.append(p)
+        p.start()
+    for i in range(n_p):
+        ps[i].join()

infer/modules/train/extract/extract_f0_rmvpe.py

@@ -0,0 +1,134 @@
+import os, traceback, sys, parselmouth
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+from lib.audio import load_audio
+import pyworld
+import numpy as np, logging
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+
+n_part = int(sys.argv[1])
+i_part = int(sys.argv[2])
+i_gpu = sys.argv[3]
+os.environ["CUDA_VISIBLE_DEVICES"] = str(i_gpu)
+exp_dir = sys.argv[4]
+is_half = sys.argv[5]
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+
+
+class FeatureInput(object):
+    def __init__(self, samplerate=16000, hop_size=160):
+        self.fs = samplerate
+        self.hop = hop_size
+
+        self.f0_bin = 256
+        self.f0_max = 1100.0
+        self.f0_min = 50.0
+        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+
+    def compute_f0(self, path, f0_method):
+        x = load_audio(path, self.fs)
+        # p_len = x.shape[0] // self.hop
+        if f0_method == "rmvpe":
+            if hasattr(self, "model_rmvpe") == False:
+                from lib.rmvpe import RMVPE
+
+                print("loading rmvpe model")
+                self.model_rmvpe = RMVPE(
+                    "assets/rmvpe/rmvpe.pt", is_half=is_half, device="cuda"
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+        return f0
+
+    def coarse_f0(self, f0):
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
+            self.f0_bin - 2
+        ) / (self.f0_mel_max - self.f0_mel_min) + 1
+
+        # use 0 or 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
+        f0_coarse = np.rint(f0_mel).astype(int)
+        assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+            f0_coarse.max(),
+            f0_coarse.min(),
+        )
+        return f0_coarse
+
+    def go(self, paths, f0_method):
+        if len(paths) == 0:
+            printt("no-f0-todo")
+        else:
+            printt("todo-f0-%s" % len(paths))
+            n = max(len(paths) // 5, 1)  # 每个进程最多打印5条
+            for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
+                try:
+                    if idx % n == 0:
+                        printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
+                    if (
+                        os.path.exists(opt_path1 + ".npy") == True
+                        and os.path.exists(opt_path2 + ".npy") == True
+                    ):
+                        continue
+                    featur_pit = self.compute_f0(inp_path, f0_method)
+                    np.save(
+                        opt_path2,
+                        featur_pit,
+                        allow_pickle=False,
+                    )  # nsf
+                    coarse_pit = self.coarse_f0(featur_pit)
+                    np.save(
+                        opt_path1,
+                        coarse_pit,
+                        allow_pickle=False,
+                    )  # ori
+                except:
+                    printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
+
+
+if __name__ == "__main__":
+    # exp_dir=r"E:\codes\py39\dataset\mi-test"
+    # n_p=16
+    # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+    printt(sys.argv)
+    featureInput = FeatureInput()
+    paths = []
+    inp_root = "%s/1_16k_wavs" % (exp_dir)
+    opt_root1 = "%s/2a_f0" % (exp_dir)
+    opt_root2 = "%s/2b-f0nsf" % (exp_dir)
+
+    os.makedirs(opt_root1, exist_ok=True)
+    os.makedirs(opt_root2, exist_ok=True)
+    for name in sorted(list(os.listdir(inp_root))):
+        inp_path = "%s/%s" % (inp_root, name)
+        if "spec" in inp_path:
+            continue
+        opt_path1 = "%s/%s" % (opt_root1, name)
+        opt_path2 = "%s/%s" % (opt_root2, name)
+        paths.append([inp_path, opt_path1, opt_path2])
+    try:
+        featureInput.go(paths[i_part::n_part], "rmvpe")
+    except:
+        printt("f0_all_fail-%s" % (traceback.format_exc()))
+    # ps = []
+    # for i in range(n_p):
+    #     p = Process(
+    #         target=featureInput.go,
+    #         args=(
+    #             paths[i::n_p],
+    #             f0method,
+    #         ),
+    #     )
+    #     ps.append(p)
+    #     p.start()
+    # for i in range(n_p):
+    #     ps[i].join()

infer/modules/train/extract/extract_f0_rmvpe_dml.py

@@ -0,0 +1,132 @@
+import os, traceback, sys, parselmouth
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+from lib.audio import load_audio
+import pyworld
+import numpy as np, logging
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+
+exp_dir = sys.argv[1]
+import torch_directml
+
+device = torch_directml.device(torch_directml.default_device())
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+
+
+class FeatureInput(object):
+    def __init__(self, samplerate=16000, hop_size=160):
+        self.fs = samplerate
+        self.hop = hop_size
+
+        self.f0_bin = 256
+        self.f0_max = 1100.0
+        self.f0_min = 50.0
+        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+
+    def compute_f0(self, path, f0_method):
+        x = load_audio(path, self.fs)
+        # p_len = x.shape[0] // self.hop
+        if f0_method == "rmvpe":
+            if hasattr(self, "model_rmvpe") == False:
+                from lib.rmvpe import RMVPE
+
+                print("loading rmvpe model")
+                self.model_rmvpe = RMVPE(
+                    "assets/rmvpe/rmvpe.pt", is_half=False, device=device
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+        return f0
+
+    def coarse_f0(self, f0):
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
+            self.f0_bin - 2
+        ) / (self.f0_mel_max - self.f0_mel_min) + 1
+
+        # use 0 or 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
+        f0_coarse = np.rint(f0_mel).astype(int)
+        assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+            f0_coarse.max(),
+            f0_coarse.min(),
+        )
+        return f0_coarse
+
+    def go(self, paths, f0_method):
+        if len(paths) == 0:
+            printt("no-f0-todo")
+        else:
+            printt("todo-f0-%s" % len(paths))
+            n = max(len(paths) // 5, 1)  # 每个进程最多打印5条
+            for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
+                try:
+                    if idx % n == 0:
+                        printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
+                    if (
+                        os.path.exists(opt_path1 + ".npy") == True
+                        and os.path.exists(opt_path2 + ".npy") == True
+                    ):
+                        continue
+                    featur_pit = self.compute_f0(inp_path, f0_method)
+                    np.save(
+                        opt_path2,
+                        featur_pit,
+                        allow_pickle=False,
+                    )  # nsf
+                    coarse_pit = self.coarse_f0(featur_pit)
+                    np.save(
+                        opt_path1,
+                        coarse_pit,
+                        allow_pickle=False,
+                    )  # ori
+                except:
+                    printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
+
+
+if __name__ == "__main__":
+    # exp_dir=r"E:\codes\py39\dataset\mi-test"
+    # n_p=16
+    # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+    printt(sys.argv)
+    featureInput = FeatureInput()
+    paths = []
+    inp_root = "%s/1_16k_wavs" % (exp_dir)
+    opt_root1 = "%s/2a_f0" % (exp_dir)
+    opt_root2 = "%s/2b-f0nsf" % (exp_dir)
+
+    os.makedirs(opt_root1, exist_ok=True)
+    os.makedirs(opt_root2, exist_ok=True)
+    for name in sorted(list(os.listdir(inp_root))):
+        inp_path = "%s/%s" % (inp_root, name)
+        if "spec" in inp_path:
+            continue
+        opt_path1 = "%s/%s" % (opt_root1, name)
+        opt_path2 = "%s/%s" % (opt_root2, name)
+        paths.append([inp_path, opt_path1, opt_path2])
+    try:
+        featureInput.go(paths, "rmvpe")
+    except:
+        printt("f0_all_fail-%s" % (traceback.format_exc()))
+    # ps = []
+    # for i in range(n_p):
+    #     p = Process(
+    #         target=featureInput.go,
+    #         args=(
+    #             paths[i::n_p],
+    #             f0method,
+    #         ),
+    #     )
+    #     ps.append(p)
+    #     p.start()
+    # for i in range(n_p):
+    #     ps[i].join()

infer/modules/train/extract_feature_print.py

@@ -0,0 +1,135 @@
+import os, sys, traceback
+
+os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
+os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"
+
+device = sys.argv[1]
+n_part = int(sys.argv[2])
+i_part = int(sys.argv[3])
+if len(sys.argv) == 6:
+    exp_dir = sys.argv[4]
+    version = sys.argv[5]
+else:
+    i_gpu = sys.argv[4]
+    exp_dir = sys.argv[5]
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(i_gpu)
+    version = sys.argv[6]
+import torch
+import torch.nn.functional as F
+import soundfile as sf
+import numpy as np
+import fairseq
+
+if "privateuseone" not in device:
+    device = "cpu"
+    if torch.cuda.is_available():
+        device = "cuda"
+    elif torch.backends.mps.is_available():
+        device = "mps"
+else:
+    import torch_directml
+
+    device = torch_directml.device(torch_directml.default_device())
+
+    def forward_dml(ctx, x, scale):
+        ctx.scale = scale
+        res = x.clone().detach()
+        return res
+
+    fairseq.modules.grad_multiply.GradMultiply.forward = forward_dml
+
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+
+
+printt(sys.argv)
+model_path = "assets/hubert/hubert_base.pt"
+
+printt(exp_dir)
+wavPath = "%s/1_16k_wavs" % exp_dir
+outPath = (
+    "%s/3_feature256" % exp_dir if version == "v1" else "%s/3_feature768" % exp_dir
+)
+os.makedirs(outPath, exist_ok=True)
+
+
+# wave must be 16k, hop_size=320
+def readwave(wav_path, normalize=False):
+    wav, sr = sf.read(wav_path)
+    assert sr == 16000
+    feats = torch.from_numpy(wav).float()
+    if feats.dim() == 2:  # double channels
+        feats = feats.mean(-1)
+    assert feats.dim() == 1, feats.dim()
+    if normalize:
+        with torch.no_grad():
+            feats = F.layer_norm(feats, feats.shape)
+    feats = feats.view(1, -1)
+    return feats
+
+
+# HuBERT model
+printt("load model(s) from {}".format(model_path))
+# if hubert model is exist
+if os.access(model_path, os.F_OK) == False:
+    printt(
+        "Error: Extracting is shut down because %s does not exist, you may download it from https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main"
+        % model_path
+    )
+    exit(0)
+models, saved_cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+    [model_path],
+    suffix="",
+)
+model = models[0]
+model = model.to(device)
+printt("move model to %s" % device)
+if device not in ["mps", "cpu"]:
+    model = model.half()
+model.eval()
+
+todo = sorted(list(os.listdir(wavPath)))[i_part::n_part]
+n = max(1, len(todo) // 10)  # 最多打印十条
+if len(todo) == 0:
+    printt("no-feature-todo")
+else:
+    printt("all-feature-%s" % len(todo))
+    for idx, file in enumerate(todo):
+        try:
+            if file.endswith(".wav"):
+                wav_path = "%s/%s" % (wavPath, file)
+                out_path = "%s/%s" % (outPath, file.replace("wav", "npy"))
+
+                if os.path.exists(out_path):
+                    continue
+
+                feats = readwave(wav_path, normalize=saved_cfg.task.normalize)
+                padding_mask = torch.BoolTensor(feats.shape).fill_(False)
+                inputs = {
+                    "source": feats.half().to(device)
+                    if device not in ["mps", "cpu"]
+                    else feats.to(device),
+                    "padding_mask": padding_mask.to(device),
+                    "output_layer": 9 if version == "v1" else 12,  # layer 9
+                }
+                with torch.no_grad():
+                    logits = model.extract_features(**inputs)
+                    feats = (
+                        model.final_proj(logits[0]) if version == "v1" else logits[0]
+                    )
+
+                feats = feats.squeeze(0).float().cpu().numpy()
+                if np.isnan(feats).sum() == 0:
+                    np.save(out_path, feats, allow_pickle=False)
+                else:
+                    printt("%s-contains nan" % file)
+                if idx % n == 0:
+                    printt("now-%s,all-%s,%s,%s" % (len(todo), idx, file, feats.shape))
+        except:
+            printt(traceback.format_exc())
+    printt("all-feature-done")

infer/modules/train/train.py

@@ -3,7 +3,7 @@ import os, sys
 now_dir = os.getcwd()
 sys.path.append(os.path.join(now_dir))
 
-from lib.train import utils
+from infer.lib.train import utils
 import datetime
 
 hps = utils.get_hparams()
@@ -22,10 +22,10 @@ import torch.multiprocessing as mp
 import torch.distributed as dist
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.cuda.amp import autocast, GradScaler
-from lib.infer_pack import commons
+from infer.lib.infer_pack import commons
 from time import sleep
 from time import time as ttime
-from lib.train.data_utils import (
+from infer.lib.train.data_utils import (
     TextAudioLoaderMultiNSFsid,
     TextAudioLoader,
     TextAudioCollateMultiNSFsid,
@@ -34,20 +34,25 @@ from lib.train.data_utils import (
 )
 
 if hps.version == "v1":
-    from lib.infer_pack.models import (
+    from infer.lib.infer_pack.models import (
         SynthesizerTrnMs256NSFsid as RVC_Model_f0,
         SynthesizerTrnMs256NSFsid_nono as RVC_Model_nof0,
         MultiPeriodDiscriminator,
     )
 else:
-    from lib.infer_pack.models import (
+    from infer.lib.infer_pack.models import (
         SynthesizerTrnMs768NSFsid as RVC_Model_f0,
         SynthesizerTrnMs768NSFsid_nono as RVC_Model_nof0,
         MultiPeriodDiscriminatorV2 as MultiPeriodDiscriminator,
     )
-from lib.train.losses import generator_loss, discriminator_loss, feature_loss, kl_loss
-from lib.train.mel_processing import mel_spectrogram_torch, spec_to_mel_torch
-from lib.train.process_ckpt import savee
+from infer.lib.train.losses import (
+    generator_loss,
+    discriminator_loss,
+    feature_loss,
+    kl_loss,
+)
+from infer.lib.train.mel_processing import mel_spectrogram_torch, spec_to_mel_torch
+from infer.lib.train.process_ckpt import savee
 
 global_step = 0