Delete dreamvoice/src/feats/.ipynb_checkpoints

dreamvoice/src/feats/.ipynb_checkpoints/contentvec-checkpoint.py

@@ -1,42 +0,0 @@
-import torch
-import librosa
-from fairseq import checkpoint_utils
-import torch.nn.functional as F
-
-
-def get_model(vec_path):
-    print("load model(s) from {}".format(vec_path))
-    models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
-        [vec_path],
-        suffix="",
-    )
-    model = models[0]
-    model.eval()
-    return model
-
-
-@torch.no_grad()
-def get_content(hmodel, wav_16k_tensor, device='cuda', layer=12):
-    # print(layer)
-    wav_16k_tensor = wav_16k_tensor.to(device)
-    # so that the output shape will be len(audio//320)
-    wav_16k_tensor = F.pad(wav_16k_tensor, ((400 - 320) // 2, (400 - 320) // 2))
-    feats = wav_16k_tensor
-    padding_mask = torch.BoolTensor(feats.shape).fill_(False)
-    inputs = {
-        "source": feats.to(wav_16k_tensor.device),
-        "padding_mask": padding_mask.to(wav_16k_tensor.device),
-        "output_layer": layer
-    }
-    logits = hmodel.extract_features(**inputs)[0]
-    # feats = hmodel.final_proj(logits[0])
-    return logits
-
-
-if __name__ == '__main__':
-    audio, sr = librosa.load('test.wav', sr=16000)
-    audio = audio[:100*320]
-    model = get_model('../../ckpts/checkpoint_best_legacy_500.pt')
-    model = model.cuda()
-    content = get_content(model, torch.tensor([audio]))
-    print(content)

dreamvoice/src/feats/.ipynb_checkpoints/contentvec_hf-checkpoint.py

@@ -1,40 +0,0 @@
-from transformers import HubertModel
-import torch.nn as nn
-import torch
-import torch.nn.functional as F
-import librosa
-
-
-class HubertModelWithFinalProj(HubertModel):
-    def __init__(self, config):
-        super().__init__(config)
-
-        # The final projection layer is only used for backward compatibility.
-        # Following https://github.com/auspicious3000/contentvec/issues/6
-        # Remove this layer is necessary to achieve the desired outcome.
-        self.final_proj = nn.Linear(config.hidden_size, config.classifier_proj_size)
-
-
-def get_content_model(config='lengyue233/content-vec-best'):
-    model = HubertModelWithFinalProj.from_pretrained(config)
-    model.eval()
-    return model
-
-
-@torch.no_grad()
-def get_content(model, wav_16k_tensor, device='cuda'):
-    # print(layer)
-    wav_16k_tensor = wav_16k_tensor.to(device)
-    # so that the output shape will be len(audio//320)
-    wav_16k_tensor = F.pad(wav_16k_tensor, ((400 - 320) // 2, (400 - 320) // 2))
-    logits = model(wav_16k_tensor)['last_hidden_state']
-    return logits
-
-
-if __name__ == '__main__':
-    model = get_content_model().cuda()
-    audio, sr = librosa.load('test.wav', sr=16000)
-    audio = audio[:100*320]
-    audio = torch.tensor([audio])
-    content = get_content(model, audio, 'cuda')
-    print(content)

dreamvoice/src/feats/.ipynb_checkpoints/hubert_model-checkpoint.py

@@ -1,24 +0,0 @@
-import torch, torchaudio
-from .hubert.hubert import HubertSoft
-from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
-import librosa
-
-
-def get_soft_model(model_path):
-    hubert = HubertSoft()
-    # Load checkpoint (either hubert_soft or hubert_discrete)
-    # hubert = torch.hub.load("bshall/hubert:main", "hubert_soft", trust_repo=True)
-    checkpoint = torch.load(model_path)
-    consume_prefix_in_state_dict_if_present(checkpoint["hubert"], "module.")
-    hubert.load_state_dict(checkpoint["hubert"])
-    hubert.eval()
-    return hubert
-
-
-@torch.no_grad()
-def get_hubert_soft_content(hmodel, wav_16k_tensor, device='cuda'):
-    wav_16k_tensor = wav_16k_tensor.to(device).unsqueeze(1)
-    # print(wav_16k_tensor.shape)
-    units = hmodel.units(wav_16k_tensor)
-    # print(units.shape)
-    return units.cpu()

dreamvoice/src/feats/.ipynb_checkpoints/test-checkpoint.py

@@ -1,22 +0,0 @@
-import torch, torchaudio
-from hubert.hubert import HubertSoft
-from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
-import librosa
-
-
-def get_soft_model(model_path):
-    hubert = HubertSoft()
-    # Load checkpoint (either hubert_soft or hubert_discrete)
-    # hubert = torch.hub.load("bshall/hubert:main", "hubert_soft", trust_repo=True)
-    checkpoint = torch.load(model_path)
-    consume_prefix_in_state_dict_if_present(checkpoint["hubert"], "module.")
-    hubert.load_state_dict(checkpoint["hubert"])
-    hubert.eval()
-    return model
-
-
-@torch.no_grad()
-def get_hubert_soft_content(hmodel, wav_16k_tensor, device='cuda'):
-    wav_16k_tensor = wav_16k_tensor.to(device)
-    units = hmodel.units(wav_16k_tensor)
-    return units.cpu()