fix description input for 1.5

levo_inference.py

@@ -16,6 +16,7 @@ from codeclm.models import builders
 from codeclm.models import CodecLM
 
 from separator import Separator
+from generate import check_language_by_text
 
 
 class LeVoInference(torch.nn.Module):
@@ -80,7 +81,11 @@ class LeVoInference(torch.nn.Module):
             melody_is_wav = True
         elif genre is not None and auto_prompt_path is not None:
             auto_prompt = torch.load(auto_prompt_path)
-            prompt_token = auto_prompt[genre][np.random.randint(0, len(auto_prompt[genre]))]
+            if genre == 'Auto':
+                lang = check_language_by_text(lyric)
+                prompt_token = auto_prompt['Auto'][lang][np.random.randint(0, len(auto_prompt['Auto'][lang]))]
+            else:
+                prompt_token = auto_prompt[genre][np.random.randint(0, len(auto_prompt[genre]))]
             pmt_wav = prompt_token[:,[0],:]
             vocal_wav = prompt_token[:,[1],:]
             bgm_wav = prompt_token[:,[2],:]
@@ -91,6 +96,8 @@ class LeVoInference(torch.nn.Module):
             bgm_wav = None
             melody_is_wav = True
 
+        description = description if description else '.'
+        description = '[Musicality-very-high]' + ', ' + description
         generate_inp = {
             'lyrics': [lyric.replace("  ", " ")],
             'descriptions': [description],

tools/new_prompt.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:db5490d73565f95047fa826cd43d2e1416f3b15e61bebd57c193813b752cb988
-size 2102124
+oid sha256:ec5113243a3726adf86b46c0cc5838211d7cee25449f2b780f26b24963eb5ad9
+size 4745412