lt-tensor 0.0.1a40__tar.gz → 0.0.1a41__tar.gz

{lt_tensor-0.0.1a40 → lt_tensor-0.0.1a41}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lt-tensor
-Version: 0.0.1a40
+Version: 0.0.1a41
 Summary: General utilities for PyTorch and others. Built for general use.
 Home-page: https://github.com/gr1336/lt-tensor/
 Author: gr1336

{lt_tensor-0.0.1a40 → lt_tensor-0.0.1a41}/lt_tensor/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.0.1a40"
+__version__ = "0.0.1a41"
 
 from . import (
     lr_schedulers,

{lt_tensor-0.0.1a40 → lt_tensor-0.0.1a41}/lt_tensor/losses.py

@@ -140,7 +140,6 @@ class MultiMelScaleLoss(Model):
         onesided: Optional[bool] = None,
         std: int = 4,
         mean: int = -4,
-        use_istft_norm: bool = True,
         use_pitch_loss: bool = True,
         use_rms_loss: bool = True,
         norm_pitch_fn: Callable[[Tensor], Tensor] = normalize_minmax,
@@ -165,7 +164,6 @@ class MultiMelScaleLoss(Model):
         self.loss_rms_fn = loss_rms_fn
         self.lambda_mel = lambda_mel
         self.weight = weight
-        self.use_istft_norm = use_istft_norm
         self.use_pitch_loss = use_pitch_loss
         self.use_rms_loss = use_rms_loss
         self.lambda_pitch = lambda_pitch
@@ -251,36 +249,26 @@ class MultiMelScaleLoss(Model):
     def forward(
         self, input_wave: torch.Tensor, target_wave: torch.Tensor
     ) -> torch.Tensor:
-        assert self.use_istft_norm or input_wave.shape[-1] == target_wave.shape[-1], (
-            f"Size mismatch! input_wave {input_wave.shape[-1]} must match target_wave: {target_wave.shape[-1]}. "
-            "Alternatively 'use_istft_norm' can be set to Trie with will automatically force the audio to that size."
-        )
         target_wave = target_wave.to(input_wave.device)
         losses = 0.0
         for M in self.mel_spectrograms:
-            if self.use_istft_norm and input_proc.shape[-1] != target_proc.shape[-1]:
-                input_proc = M.istft_norm(input_wave, length=target_wave.shape[-1])
-                target_proc = M.istft_norm(target_wave, length=target_wave.shape[-1])
-            else:
-                input_proc, target_proc = input_wave, target_wave
-
-            x_mels = M(input_proc)
-            y_mels = M(target_proc)
+            x_mels = M.compute_mel(input_wave)
+            y_mels = M.compute_mel(target_wave)
 
             loss = self.loss_mel_fn(x_mels.squeeze(), y_mels.squeeze())
             losses += loss * self.lambda_mel
 
             # pitch/f0 loss
             if self.use_pitch_loss:
-                x_pitch = self.norm_pitch_fn(M.compute_pitch(input_proc))
-                y_pitch = self.norm_pitch_fn(M.compute_pitch(target_proc))
+                x_pitch = self.norm_pitch_fn(M.compute_pitch(input_wave))
+                y_pitch = self.norm_pitch_fn(M.compute_pitch(target_wave))
                 f0_loss = self.loss_pitch_fn(x_pitch, y_pitch)
                 losses += f0_loss * self.lambda_pitch
 
             # energy/rms loss
             if self.use_rms_loss:
-                x_rms = self.norm_rms(M.compute_rms(input_proc, x_mels))
-                y_rms = self.norm_rms(M.compute_rms(target_proc, y_mels))
+                x_rms = self.norm_rms(M.compute_rms(input_wave, x_mels))
+                y_rms = self.norm_rms(M.compute_rms(target_wave, y_mels))
                 rms_loss = self.loss_rms_fn(x_rms, y_rms)
                 losses += rms_loss * self.lambda_rms
 

lt_tensor-0.0.1a41/lt_tensor/model_zoo/audio_models/__init__.py

@@ -0,0 +1,3 @@
+from . import diffwave, istft, hifigan, bigvgan, bemaganv2
+
+__all__ = ["diffwave", "istft", "hifigan", "bigvgan", "bemaganv2"]

{lt_tensor-0.0.1a40 → lt_tensor-0.0.1a41}/lt_tensor/processors/audio.py

@@ -698,66 +698,4 @@ class AudioProcessor(Model):
         rms_batch_dim: bool = False,
         spec_phase_batch_dim: bool = False,
     ):
-        results = {
-            "wave": None,
-            "mel": None,
-            "pitch": None,
-            "rms": None,
-            "spec": None,
-            "phase": None,
-        }
-        results["wave"] = (
-            x.squeeze()
-            if isinstance(x, Tensor)
-            else self.load_audio(x, istft_norm=True).squeeze()
-        )
-        results["mel"] = self.compute_mel_librosa(
-            wave=(
-                results["wave"]
-                if results["wave"].ndim == 3
-                else results["wave"].unsqueeze(0)
-            ),
-            spectral_norm=spectral_norm,
-        ).squeeze()
-        try:
-            results["pitch"] = self.compute_pitch(results["wave"]).squeeze()
-        except Exception as e:
-            results["pitch"] = e
-        try:
-            results["rms"] = self.compute_rms(results["wave"], results["mel"]).squeeze()
-        except Exception as e:
-            results["rms"] = e
-        try:
-            sp_ph = self.stft(results["wave"], return_complex=False)
-            spec, phase = sp_ph.split(1, -1)
-            results["spec"] = spec.squeeze()
-            results["phase"] = phase.squeeze()
-        except Exception as e:
-            results["spec"] = e
-            results["phase"] = e
-
-        if (add_batch_to_all or wave_batch_dim) and results["wave"].ndim == 1:
-            results["wave"] = results["wave"].unsqueeze(0)
-        if (add_batch_to_all or mel_batch_dim) and results["mel"].ndim == 2:
-            results["mel"] = results["mel"].unsqueeze(0)
-        if (
-            isinstance(results["rms"], Tensor)
-            and (add_batch_to_all or rms_batch_dim)
-            and results["rms"].ndim == 1
-        ):
-            results["rms"] = results["rms"].unsqueeze(0)
-        if (
-            isinstance(results["pitch"], Tensor)
-            and (add_batch_to_all or pitch_batch_dim)
-            and results["pitch"].ndim == 1
-        ):
-            results["pitch"] = results["pitch"].unsqueeze(0)
-        if (
-            isinstance(results["spec"], Tensor)
-            and (add_batch_to_all or spec_phase_batch_dim)
-            and results["spec"].ndim == 2
-        ):
-            results["spec"] = results["spec"].unsqueeze(0)
-            results["phase"] = results["phase"].unsqueeze(0)
-
-        return results
+        pass

{lt_tensor-0.0.1a40 → lt_tensor-0.0.1a41}/lt_tensor.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lt-tensor
-Version: 0.0.1a40
+Version: 0.0.1a41
 Summary: General utilities for PyTorch and others. Built for general use.
 Home-page: https://github.com/gr1336/lt-tensor/
 Author: gr1336

{lt_tensor-0.0.1a40 → lt_tensor-0.0.1a41}/setup.py

@@ -4,7 +4,7 @@ with open("README.md", "r", encoding="utf-8") as f:
     long_description = f.read()
 
 setup(
-    version="0.0.1a40",
+    version="0.0.1a41",
     name="lt-tensor",
     description="General utilities for PyTorch and others. Built for general use.",
     long_description=long_description,

lt_tensor-0.0.1a40/lt_tensor/model_zoo/audio_models/__init__.py

@@ -1,3 +0,0 @@
-from . import diffwave, istft, hifigan, bigvgan
-
-__all__ = ["diffwave", "istft", "hifigan", "bigvgan"]