lt-tensor 0.0.1a36__py3-none-any.whl → 0.0.1a37__py3-none-any.whl

lt_tensor/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.0.1a36"
+__version__ = "0.0.1a37"
 
 from . import (
     lr_schedulers,

lt_tensor/processors/audio.py

@@ -10,6 +10,7 @@ from lt_utils.type_utils import is_file, is_array
 from lt_utils.file_ops import FileScan, get_file_name, path_to_str
 from torchaudio.functional import detect_pitch_frequency
 import torch.nn.functional as F
+from librosa.filters import mel as _mel_filter_bank
 
 DEFAULT_DEVICE = torch.tensor([0]).device
 
@@ -25,7 +26,7 @@ class AudioProcessorConfig(ModelConfig):
     f_min: float = 0
     f_max: Optional[float] = None
     center: bool = True
-    mel_scale: Literal["htk" "slaney"] = "htk"
+    mel_scale: Literal["htk", "slaney"] = "htk"
     std: int = 4
     mean: int = -4
     n_iter: int = 32
@@ -33,6 +34,7 @@ class AudioProcessorConfig(ModelConfig):
     normalized: bool = False
     onesided: Optional[bool] = None
     n_stft: int = None
+    mel_default: Literal["torch", "librosa"] = "librosa"
 
     def __init__(
         self,
@@ -49,6 +51,7 @@ class AudioProcessorConfig(ModelConfig):
         mean: int = -4,
         normalized: bool = False,
         onesided: Optional[bool] = None,
+        mel_default: Literal["torch", "librosa"] = "librosa",
         *args,
         **kwargs,
     ):
@@ -66,6 +69,7 @@ class AudioProcessorConfig(ModelConfig):
             "mean": mean,
             "normalized": normalized,
             "onesided": onesided,
+            "mel_default": mel_default,
         }
         super().__init__(**settings)
         self.post_process()
@@ -88,14 +92,10 @@ def _comp_rms_helper(i: int, audio: Tensor, mel: Optional[Tensor]):
 
 
 class AudioProcessor(Model):
-    def __init__(
-        self,
-        config: AudioProcessorConfig = AudioProcessorConfig(),
-        window: Optional[Tensor] = None,
-    ):
+    def __init__(self, config: AudioProcessorConfig = AudioProcessorConfig()):
         super().__init__()
         self.cfg = config
-        self._mel_spec = torchaudio.transforms.MelSpectrogram(
+        self._mel_spec_torch = torchaudio.transforms.MelSpectrogram(
             sample_rate=self.cfg.sample_rate,
             n_mels=self.cfg.n_mels,
             n_fft=self.cfg.n_fft,
@@ -107,6 +107,7 @@ class AudioProcessor(Model):
             mel_scale=self.cfg.mel_scale,
             normalized=self.cfg.normalized,
         )
+
         self._mel_rscale = torchaudio.transforms.InverseMelScale(
             n_stft=self.cfg.n_stft,
             n_mels=self.cfg.n_mels,
@@ -115,34 +116,119 @@ class AudioProcessor(Model):
             f_max=self.cfg.f_max,
             mel_scale=self.cfg.mel_scale,
         )
-
+        self.mel_lib_padding = (self.cfg.n_fft - self.cfg.hop_length) // 2
         self.register_buffer(
             "window",
-            (torch.hann_window(self.cfg.win_length) if window is None else window),
+            torch.hann_window(self.cfg.win_length),
         )
+        self.register_buffer(
+            "mel_filter_bank",
+            torch.from_numpy(
+                _mel_filter_bank(
+                    sr=self.cfg.sample_rate,
+                    n_fft=self.cfg.n_fft,
+                    n_mels=self.cfg.n_mels,
+                    fmin=self.cfg.f_min,
+                    fmax=self.cfg.f_max,
+                )
+            ).float(),
+        )
+
+    def spectral_norm(self, x: Tensor, c: int = 1, eps: float = 1e-5):
+        return torch.log(torch.clamp(x, min=eps) * c)
+
+    def spectral_de_norm(self, x: Tensor, c: int = 1):
+        return torch.exp(x) / c
+
+    def log_norm(
+        self,
+        entry: Tensor,
+        eps: float = 1e-5,
+        mean: Optional[Number] = None,
+        std: Optional[Number] = None,
+    ) -> Tensor:
+        mean = default(mean, self.cfg.mean)
+        std = default(std, self.cfg.std)
+        return (torch.log(eps + entry.unsqueeze(0)) - mean) / std
 
     def compute_mel(
         self,
         wave: Tensor,
-        eps: float = 1e-5,
-        raw_mel_only: bool = False,
-        *,
-        _recall: bool = False,
+        method: Optional[Literal["torch", "librosa"]] = None,
+        apply_norm: bool = False,
+        eps: Optional[float] = None,
+        **kwargs,
+    ) -> Tensor:
+        method = default(method, self.cfg.mel_default)
+        if method == "torch":
+            return self.compute_mel_torch(
+                wave,
+                log_norm=apply_norm,
+                eps=eps,
+                mean=kwargs.get("mean", None),
+                std=kwargs.get("std", None),
+            )
+        return self.compute_mel_librosa(
+            wave,
+            log_norm=apply_norm,
+            eps=eps,
+        )
+
+    def compute_mel_torch(
+        self,
+        wave: Tensor,
+        log_norm: bool = False,
+        eps: Optional[float] = None,
+        mean: Optional[Number] = None,
+        std: Optional[Number] = None,
+        *args,
+        **kwargs,
     ) -> Tensor:
         """Returns: (M, T) or (B, M, T) if batched"""
         try:
-            mel_tensor = self._mel_spec(wave.to(self.device))  # [M, T]
-            if not raw_mel_only:
-                mel_tensor = (
-                    torch.log(eps + mel_tensor.unsqueeze(0)) - self.cfg.mean
-                ) / self.cfg.std
-            return mel_tensor.squeeze()
+            mel_tensor = self._mel_spec_torch.forward(wave.to(self.device))  # [M, T]
 
         except RuntimeError as e:
-            if not _recall:
-                self._mel_spec.to(self.device)
-                return self.compute_mel(wave, raw_mel_only, eps, _recall=True)
-            raise e
+            mel_tensor = self._mel_spec_torch.forward(wave.to(self.device))  # [M, T]
+        if log_norm:
+            return self.log_norm(mel_tensor, eps, mean, std).squeeze()
+        return mel_tensor.squeeze()
+
+    def compute_mel_librosa(
+        self,
+        wave: Tensor,
+        eps: float = 1e-5,
+        spectral_norm: bool = False,
+        *args,
+        **kwargs,
+    ):
+        wave = torch.nn.functional.pad(
+            wave.unsqueeze(1),
+            (self.mel_lib_padding, self.mel_lib_padding),
+            mode="reflect",
+        ).squeeze(1)
+        spec = torch.stft(
+            wave,
+            self.cfg.n_fft,
+            hop_length=self.cfg.hop_length,
+            win_length=self.cfg.win_length,
+            window=self.window,
+            center=self.cfg.center,
+            pad_mode="reflect",
+            normalized=False,
+            onesided=True,
+            return_complex=True,
+        )
+        spec = torch.sqrt(torch.view_as_real(spec).pow(2).sum(-1) + 1e-12)
+        try:
+            results = torch.matmul(self.mel_filter_bank, spec)
+        except RuntimeError:
+            self.mel_filter_bank = self.mel_filter_bank.to(self.device)
+            self.window = self.window.to(self.device)
+            results = torch.matmul(self.mel_filter_bank, spec)
+        if spectral_norm:
+            return self.spectral_norm(results, eps=eps).squeeze()
+        return results.squeeze()
 
     def compute_inverse_mel(self, melspec: Tensor, *, _recall=False):
         try:
@@ -382,7 +468,7 @@ class AudioProcessor(Model):
             antialias=antialias,
         )
 
-    def istft(
+    def istft_spec_phase(
         self,
         spec: Tensor,
         phase: Tensor,
@@ -394,34 +480,91 @@ class AudioProcessor(Model):
         normalized: Optional[bool] = None,
         onesided: Optional[bool] = None,
         return_complex: bool = False,
-        *,
-        _recall: bool = False,
     ):
-        if win_length is not None and win_length != self.cfg.win_length:
-            window = torch.hann_window(win_length, device=spec.device)
-        else:
-            window = self.window
+        """Util for models that needs to reconstruct the audio using inverse stft"""
+        window = (
+            torch.hann_window(win_length, device=spec.device)
+            if win_length is not None and win_length != self.cfg.win_length
+            else self.window.to(spec.device)
+        )
+        return torch.istft(
+            spec * torch.exp(phase * 1j),
+            n_fft=default(n_fft, self.cfg.n_fft),
+            hop_length=default(hop_length, self.cfg.hop_length),
+            win_length=default(win_length, self.cfg.win_length),
+            window=window,
+            center=center,
+            normalized=default(normalized, self.cfg.normalized),
+            onesided=default(onesided, self.cfg.onesided),
+            length=length,
+            return_complex=return_complex,
+        )
 
-        try:
-            return torch.istft(
-                spec * torch.exp(phase * 1j),
-                n_fft=default(n_fft, self.cfg.n_fft),
-                hop_length=default(hop_length, self.cfg.hop_length),
-                win_length=default(win_length, self.cfg.win_length),
-                window=window,
-                center=center,
-                normalized=default(normalized, self.cfg.normalized),
-                onesided=default(onesided, self.cfg.onesided),
-                length=length,
-                return_complex=return_complex,
-            )
-        except RuntimeError as e:
-            if not _recall and spec.device != self.window.device:
-                self.window = self.window.to(spec.device)
-                return self.istft(
-                    spec, phase, n_fft, hop_length, win_length, length, _recall=True
-                )
-            raise e
+    def istft(
+        self,
+        wave: Tensor,
+        n_fft: Optional[int] = None,
+        hop_length: Optional[int] = None,
+        win_length: Optional[int] = None,
+        length: Optional[int] = None,
+        center: bool = True,
+        normalized: Optional[bool] = None,
+        onesided: Optional[bool] = None,
+        return_complex: bool = False,
+    ):
+        window = (
+            torch.hann_window(win_length, device=wave.device)
+            if win_length is not None and win_length != self.cfg.win_length
+            else self.window.to(wave.device)
+        )
+        if not torch.is_complex(wave):
+            wave = wave * 1j
+        return torch.istft(
+            wave,
+            n_fft=default(n_fft, self.cfg.n_fft),
+            hop_length=default(hop_length, self.cfg.hop_length),
+            win_length=default(win_length, self.cfg.win_length),
+            window=window,
+            center=center,
+            normalized=default(normalized, self.cfg.normalized),
+            onesided=default(onesided, self.cfg.onesided),
+            length=length,
+            return_complex=return_complex,
+        )
+
+    def stft(
+        self,
+        wave: Tensor,
+        center: bool = True,
+        n_fft: Optional[int] = None,
+        hop_length: Optional[int] = None,
+        win_length: Optional[int] = None,
+        normalized: Optional[bool] = None,
+        onesided: Optional[bool] = None,
+        return_complex: bool = True,
+    ):
+
+        window = (
+            torch.hann_window(win_length, device=wave.device)
+            if win_length is not None and win_length != self.cfg.win_length
+            else self.window.to(wave.device)
+        )
+
+        results = torch.stft(
+            input=wave,
+            n_fft=default(n_fft, self.cfg.n_fft),
+            hop_length=default(hop_length, self.cfg.hop_length),
+            win_length=default(win_length, self.cfg.win_length),
+            window=window,
+            center=center,
+            pad_mode="reflect",
+            normalized=default(normalized, self.cfg.normalized),
+            onesided=default(onesided, self.cfg.onesided),
+            return_complex=True,  # always, then if we need a not complex type we use view as real.
+        )
+        if not return_complex:
+            return torch.view_as_real(results)
+        return results
 
     def istft_norm(
         self,
@@ -435,11 +578,11 @@ class AudioProcessor(Model):
         onesided: Optional[bool] = None,
         return_complex: bool = False,
     ):
-
-        if win_length is not None and win_length != self.cfg.win_length:
-            window = torch.hann_window(win_length, device=wave.device)
-        else:
-            window = self.window
+        window = (
+            torch.hann_window(win_length, device=wave.device)
+            if win_length is not None and win_length != self.cfg.win_length
+            else self.window.to(wave.device)
+        )
         spectrogram = torch.stft(
             input=wave,
             n_fft=default(n_fft, self.cfg.n_fft),
@@ -473,15 +616,15 @@ class AudioProcessor(Model):
     def load_audio(
         self,
         path: PathLike,
-        top_db: float = 30,
+        top_db: Optional[float] = None,
         normalize: bool = False,
+        mono: bool = True,
         *,
-        ref: float | Callable[[np.ndarray], Any] = np.max,
-        frame_length: int = 2048,
+        sample_rate: Optional[float] = None,
         hop_length: int = 512,
-        mono: bool = True,
-        offset: float = 0.0,
+        frame_length: int = 2048,
         duration: Optional[float] = None,
+        offset: float = 0.0,
         dtype: Any = np.float32,
         res_type: str = "soxr_hq",
         fix: bool = True,
@@ -491,29 +634,32 @@ class AudioProcessor(Model):
         norm_axis: int = 0,
         norm_threshold: Optional[float] = None,
         norm_fill: Optional[bool] = None,
+        ref: float | Callable[[np.ndarray], Any] = np.max,
     ) -> Tensor:
         is_file(path, True)
+        sample_rate = default(sample_rate, self.cfg.sample_rate)
         wave, sr = librosa.load(
             str(path),
-            sr=self.cfg.sample_rate,
+            sr=sample_rate,
             mono=mono,
             offset=offset,
             duration=duration,
             dtype=dtype,
             res_type=res_type,
         )
-        wave, _ = librosa.effects.trim(
-            wave,
-            top_db=top_db,
-            ref=ref,
-            frame_length=frame_length,
-            hop_length=hop_length,
-        )
-        if sr != self.cfg.sample_rate:
+        if top_db is not None:
+            wave, _ = librosa.effects.trim(
+                wave,
+                top_db=top_db,
+                ref=ref,
+                frame_length=frame_length,
+                hop_length=hop_length,
+            )
+        if sr != sample_rate:
             wave = librosa.resample(
                 wave,
                 orig_sr=sr,
-                target_sr=self.cfg.sample_rate,
+                target_sr=sample_rate,
                 res_type=res_type,
                 fix=fix,
                 scale=scale,
@@ -553,10 +699,6 @@ class AudioProcessor(Model):
             maximum,
         )
 
-    def stft_loss(self, signal: Tensor, ground: Tensor, magnitude: float = 1.0):
-        ground = F.interpolate(ground, signal.shape[-1]).to(signal.device)
-        return F.l1_loss(signal.squeeze(), ground.squeeze()) * magnitude
-
     def forward(
         self,
         *inputs: Union[Tensor, float],

{lt_tensor-0.0.1a36.dist-info → lt_tensor-0.0.1a37.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lt-tensor
-Version: 0.0.1a36
+Version: 0.0.1a37
 Summary: General utilities for PyTorch and others. Built for general use.
 Home-page: https://github.com/gr1336/lt-tensor/
 Author: gr1336
@@ -18,7 +18,7 @@ Requires-Dist: tokenizers
 Requires-Dist: pyyaml>=6.0.0
 Requires-Dist: numba>0.60.0
 Requires-Dist: lt-utils>=0.0.4
-Requires-Dist: librosa==0.11.*
+Requires-Dist: librosa<1,>=0.10.2.post1
 Requires-Dist: einops
 Requires-Dist: plotly
 Requires-Dist: scipy

{lt_tensor-0.0.1a36.dist-info → lt_tensor-0.0.1a37.dist-info}/RECORD

@@ -1,4 +1,4 @@
-lt_tensor/__init__.py,sha256=nBbiGH1byHU0aTTKKorRj8MIEO2oEMBXl7kt5DOCatU,441
+lt_tensor/__init__.py,sha256=CFVK5h2Y-p3xFJ6mCW8dI1FOFeObsOyDjyUqJtxmkmg,441
 lt_tensor/config_templates.py,sha256=F9UvL8paAjkSvio890kp8WznpYeI50pYnm9iqQroBxk,2797
 lt_tensor/losses.py,sha256=Heco_WyoC1HkNkcJEircOAzS9umusATHiNAG-FKGyzc,8918
 lt_tensor/lr_schedulers.py,sha256=6_vcfaPHrozfH3wvmNEdKSFYl6iTIijYoHL8vuG-45U,7651
@@ -35,9 +35,9 @@ lt_tensor/model_zoo/losses/CQT/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5
 lt_tensor/model_zoo/losses/CQT/transforms.py,sha256=Vkid0J9dqLnlINfyyUlQf-qB3gOQAgU7W9j7xLOjDFw,13218
 lt_tensor/model_zoo/losses/CQT/utils.py,sha256=twGw6FVD7V5Ksfx_1BUEN3EP1tAS6wo-9LL3VnuHB8c,16751
 lt_tensor/processors/__init__.py,sha256=Pvxhh0KR65zLCgUd53_k5Z0y5JWWcO0ZBXFK9rv0o5w,109
-lt_tensor/processors/audio.py,sha256=3YzyEpMwh124rb1KMAly62qweeruF200BnM-vQIbzy0,18645
-lt_tensor-0.0.1a36.dist-info/licenses/LICENSE,sha256=TbiyJWLgNqqgqhfCnrGwFIxy7EqGNrIZZcKhHrefcuU,11354
-lt_tensor-0.0.1a36.dist-info/METADATA,sha256=mTmnoWn8EG48j_VOM3rr_8RLLgaxB5pWZE1tkPdFrac,1062
-lt_tensor-0.0.1a36.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-lt_tensor-0.0.1a36.dist-info/top_level.txt,sha256=35FuhFeXnUyvHWdbVHGPh0hS8euofafnJ_GJAVSF4Kk,10
-lt_tensor-0.0.1a36.dist-info/RECORD,,
+lt_tensor/processors/audio.py,sha256=QadO6e7uXRkheNU8ba-SNw72HPD1XvR-6VJltoF8YRA,23535
+lt_tensor-0.0.1a37.dist-info/licenses/LICENSE,sha256=TbiyJWLgNqqgqhfCnrGwFIxy7EqGNrIZZcKhHrefcuU,11354
+lt_tensor-0.0.1a37.dist-info/METADATA,sha256=6EkGRk9fT_wsvl_pqKZ0S8I-x1Awkm9pvr3MKnW8OPM,1071
+lt_tensor-0.0.1a37.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+lt_tensor-0.0.1a37.dist-info/top_level.txt,sha256=35FuhFeXnUyvHWdbVHGPh0hS8euofafnJ_GJAVSF4Kk,10
+lt_tensor-0.0.1a37.dist-info/RECORD,,