Delete evaluation/QP/qp_pred.py

evaluation/QP/qp_pred.py

@@ -1,73 +0,0 @@
-import os
-from huggingface_hub import hf_hub_download
-
-def ensure_checkpoint_exists(checkpoint_path, repo_id="LIUM/TTSEval", subfolder="evaluation/QP"):
-    # Check if the file exists locally
-    if not os.path.exists(checkpoint_path):
-        print(f"File {checkpoint_path} not found. Downloading from Hugging Face...")
-        
-        # This downloads the file and returns the local path to the cached version
-        # Note: checkpoint_path here should be the filename within the HF repo
-        filename = os.path.basename(checkpoint_path)
-        
-        downloaded_path = hf_hub_download(
-            repo_id=repo_id,
-            filename=filename,
-            subfolder=subfolder,
-            repo_type="model"
-        )
-        return downloaded_path
-    
-    return checkpoint_path
-
-def _load_sslmos_model(state):
-    """
-    Charge le backbone fairseq + LossPredictor, puis restaure ton checkpoint finetuné.
-    Retourne (net, target_sr).
-    """
-    device = state["device"]
-    target_sr = state.get("qp_target_sr", 24000)
-    ssl_out_dim = state.get("qp_ssl_out_dim", 9216)
-
-    base_model_ckpt = state["qp_fairseq_base_model"]          
-    finetuned_ckpt = state["qp_finetuned_checkpoint"]
-    # Update paths if necessary
-    base_model_ckpt = ensure_checkpoint_exists(base_model_ckpt)
-    finetuned_ckpt = ensure_checkpoint_exists(finetuned_ckpt)
-
-    # --- patch torch.load (comme dans ton script) pour éviter weights_only=True par défaut ---
-    torch.serialization.add_safe_globals([__import__("argparse").Namespace])
-    original_torch_load = torch.load
-
-    def patched_torch_load(*args, **kwargs):
-        if "weights_only" not in kwargs:
-            kwargs["weights_only"] = False
-        return original_torch_load(*args, **kwargs)
-
-    torch.load = patched_torch_load
-    models, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task([base_model_ckpt])
-    torch.load = original_torch_load
-
-    ssl_model = models[0]
-    ssl_model.remove_pretraining_modules()
-    for p in ssl_model.parameters():
-        p.requires_grad = False
-
-    net = LossPredictor(ssl_model, ssl_out_dim).to(device)
-    net.load_state_dict(torch.load(finetuned_ckpt, map_location=device))
-    net.eval()
-
-    return net, target_sr
-
-
-def _load_and_resample_wav(wav_path: str, target_sr: int, device):
-    wav, sr = torchaudio.load(wav_path)  # wav: [C, T]
-    # mono
-    if wav.size(0) > 1:
-        wav = wav.mean(dim=0, keepdim=True)
-    # resample
-    if sr != target_sr:
-        wav = torchaudio.functional.resample(wav, orig_freq=sr, new_freq=target_sr)
-    # shape attendue: [B, C, T] -> ici B=1
-    wav = wav.unsqueeze(0).to(device)  # [1, 1, T]
-    return wav