cleanup: remove dead MAGNeT/MBD loader code, drop redundant dbn=False, remove loopgen-repo reference copy

.gitignore

@@ -1,4 +1,3 @@
 __pycache__/
 *.pyc
 .DS_Store
-loopgen-repo/

app.py

@@ -173,7 +173,7 @@ def process_fn(description: str, min_length: float, max_length: float, seed: int
         if loopgen is None:
             loopgen = LoopGen.get_pretrained(MAGNET_MODEL, device=DEVICE)
         if file2beats is None:
-            file2beats = File2Beats(checkpoint_path=beat_checkpoint_path, device=DEVICE, dbn=False)
+            file2beats = File2Beats(checkpoint_path=beat_checkpoint_path, device=DEVICE)
 
     torch.manual_seed(seed)
     np.random.seed(seed)

src/audiocraft/models/loaders.py

@@ -119,60 +119,3 @@ def load_lm_model(file_or_url_or_id: tp.Union[Path, str], device='cpu', cache_di
     return model
 
 
-def load_lm_model_magnet(file_or_url_or_id: tp.Union[Path, str], compression_model_frame_rate: int,
-                         device='cpu', cache_dir: tp.Optional[str] = None):
-    pkg = load_lm_model_ckpt(file_or_url_or_id, cache_dir=cache_dir)
-    cfg = OmegaConf.create(pkg['xp.cfg'])
-    cfg.device = str(device)
-    if cfg.device == 'cpu':
-        cfg.dtype = 'float32'
-    else:
-        cfg.dtype = 'float16'
-    _delete_param(cfg, 'conditioners.args.merge_text_conditions_p')
-    _delete_param(cfg, 'conditioners.args.drop_desc_p')
-
-    cfg.transformer_lm.compression_model_framerate = compression_model_frame_rate
-    cfg.transformer_lm.segment_duration = cfg.dataset.segment_duration
-    cfg.transformer_lm.span_len = cfg.masking.span_len
-
-    # MAGNeT models v1 support only xformers backend.
-    from src.audiocraft.modules.transformer import set_efficient_attention_backend
-    if cfg.transformer_lm.memory_efficient:
-        set_efficient_attention_backend("xformers")
-
-    model = builders.get_lm_model(cfg)
-    model.load_state_dict(pkg['best_state'])
-    model.eval()
-    model.cfg = cfg
-    return model
-
-
-def load_mbd_ckpt(file_or_url_or_id: tp.Union[Path, str],
-                  filename: tp.Optional[str] = None,
-                  cache_dir: tp.Optional[str] = None):
-    return _get_state_dict(file_or_url_or_id, filename=filename, cache_dir=cache_dir)
-
-
-def load_diffusion_models(file_or_url_or_id: tp.Union[Path, str],
-                          device='cpu',
-                          filename: tp.Optional[str] = None,
-                          cache_dir: tp.Optional[str] = None):
-    pkg = load_mbd_ckpt(file_or_url_or_id, filename=filename, cache_dir=cache_dir)
-    models = []
-    processors = []
-    cfgs = []
-    sample_rate = pkg['sample_rate']
-    for i in range(pkg['n_bands']):
-        cfg = pkg[i]['cfg']
-        model = builders.get_diffusion_model(cfg)
-        model_dict = pkg[i]['model_state']
-        model.load_state_dict(model_dict)
-        model.to(device)
-        processor = builders.get_processor(cfg=cfg.processor, sample_rate=sample_rate)
-        processor_dict = pkg[i]['processor_state']
-        processor.load_state_dict(processor_dict)
-        processor.to(device)
-        models.append(model)
-        processors.append(processor)
-        cfgs.append(cfg)
-    return models, processors, cfgs

src/audiocraft/models/magnet.py

@@ -9,10 +9,8 @@ Main model for using MAGNeT. This will combine all the required components
 and provide easy access to the generation API.
 """
 import typing as tp
-import torch
 
 from .genmodel import BaseGenModel
-from .loaders import load_compression_model, load_lm_model_magnet
 
 
 class MAGNeT(BaseGenModel):
@@ -26,37 +24,6 @@ class MAGNeT(BaseGenModel):
         self.duration = self.lm.cfg.dataset.segment_duration
         self.set_generation_params()
 
-    @staticmethod
-    def get_pretrained(name: str = 'facebook/magnet-small-10secs', device=None):
-        """Return pretrained model, we provide six models:
-        - facebook/magnet-small-10secs (300M), text to music, 10-second audio samples.
-          # see: https://huggingface.co/facebook/magnet-small-10secs
-        - facebook/magnet-medium-10secs (1.5B), text to music, 10-second audio samples.
-          # see: https://huggingface.co/facebook/magnet-medium-10secs
-        - facebook/magnet-small-30secs (300M), text to music, 30-second audio samples.
-          # see: https://huggingface.co/facebook/magnet-small-30secs
-        - facebook/magnet-medium-30secs (1.5B), text to music, 30-second audio samples.
-          # see: https://huggingface.co/facebook/magnet-medium-30secs
-        - facebook/audio-magnet-small (300M), text to sound-effect (10-second samples).
-          # see: https://huggingface.co/facebook/audio-magnet-small
-        - facebook/audio-magnet-medium (1.5B), text to sound-effect (10-second samples).
-          # see: https://huggingface.co/facebook/audio-magnet-medium
-        """
-        if device is None:
-            if torch.cuda.device_count():
-                device = 'cuda'
-            else:
-                device = 'cpu'
-
-        compression_model = load_compression_model(name, device=device)
-        lm = load_lm_model_magnet(name, compression_model_frame_rate=int(compression_model.frame_rate), device=device)
-
-        if 'self_wav' in lm.condition_provider.conditioners:
-            lm.condition_provider.conditioners['self_wav'].match_len_on_eval = True
-
-        kwargs = {'name': name, 'compression_model': compression_model, 'lm': lm}
-        return MAGNeT(**kwargs)
-
     def set_generation_params(self, use_sampling: bool = True, top_k: int = 0,
                               top_p: float = 0.9, temperature: float = 3.0,
                               max_cfg_coef: float = 10.0, min_cfg_coef: float = 1.0,