add instruction for downloading content vec

fish_diffusion/modules/feature_extractors/content_vec.py

@@ -1,6 +1,8 @@
 import logging
+import os
 
 import torch
+from loguru import logger
 
 from .base import BaseFeatureExtractor
 from .builder import FEATURE_EXTRACTORS
@@ -19,6 +21,20 @@ def __init__(
     ):
         super().__init__()
 
+        if os.path.exists(checkpoint_path) is False:
+            logger.error(f"Checkpoint {checkpoint_path} does not exist")
+            logger.error(
+                f"If you are trying to use the pretrained ContentVec, you can either:"
+            )
+            logger.error(
+                f"1. Download the ContentVec from https://github.com/fishaudio/fish-diffusion/releases/tag/v1.12"
+            )
+            logger.error(
+                f"2. Run `python tools/download_nsf_hifigan.py --content-vec` to download the ContentVec model"
+            )
+
+            raise FileNotFoundError(f"Checkpoint {checkpoint_path} does not exist")
+
         models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
             [checkpoint_path], suffix=""
         )

tools/diffusion/clean_speaker_embeddings.py

@@ -1,9 +1,10 @@
 import torch
 
 data = torch.load(
-    "checkpoints/epoch=198-step=260000-valid_loss=0.18.ckpt", map_location="cpu"
+    "logs/DiffSVC/9ddsi2gk/checkpoints/epoch=88-step=300000-valid_loss=0.08.ckpt",
+    map_location="cpu",
 )
 
 del data["state_dict"]["model.speaker_encoder.embedding.weight"]
 
-torch.save(data, "checkpoints/epoch=198-step=260000-valid_loss=0.18-fixed.ckpt")
+torch.save(data, "checkpoints/content-vec-pretrained-v1.ckpt")

tools/diffusion/train.py

@@ -59,7 +59,7 @@
             i.replace(".naive_noise_predictor.", ".") for i in missing_keys
         )
 
-        assert len(unexpected_keys) == 0
+        assert len(unexpected_keys) == 0, f"Unexpected keys: {unexpected_keys}"
 
         if args.only_train_speaker_embeddings:
             for name, param in model.named_parameters():