Merge pull request #161 from sp-nitech/zerodf [skip ci]

Improve computational efficiency of FIR

Author: takenori-y

.github/workflows/ci.yml

@@ -7,7 +7,7 @@ on:
 
   pull_request:
     branches:
-      - '**'
+      - "**"
 
 jobs:
   test:

README.md

@@ -4,7 +4,7 @@
 It provides various speech signal processing modules as PyTorch layers,
 allowing users to integrate classic signal processing algorithms directly into neural network architectures and optimize them through backpropagation.
 
-[![Manual](https://img.shields.io/badge/docs-stable-blue.svg)](https://sp-nitech.github.io/diffsptk/3.4.0/)
+[![Manual](https://img.shields.io/badge/docs-stable-blue.svg)](https://sp-nitech.github.io/diffsptk/stable/)
 [![Downloads](https://static.pepy.tech/badge/diffsptk)](https://pepy.tech/project/diffsptk)
 [![ClickPy](https://img.shields.io/badge/downloads-clickpy-yellow.svg)](https://clickpy.clickhouse.com/dashboard/diffsptk)
 [![Python Version](https://img.shields.io/pypi/pyversions/diffsptk.svg)](https://pypi.python.org/pypi/diffsptk)
@@ -22,7 +22,7 @@ allowing users to integrate classic signal processing algorithms directly into n
 
 ## Documentation
 
-- [**Reference Manual**](https://sp-nitech.github.io/diffsptk/3.4.0/) - Detailed API documentation and module specifications.
+- [**Reference Manual**](https://sp-nitech.github.io/diffsptk/stable/) - Detailed API documentation and module specifications.
 - [**Interactive Tutorial**](https://colab.research.google.com/drive/1xAoUKqXadvJXJ7RzN0OceB6y7q5i7Sn6?usp=drive_link) (Google Colab) - Hands-on examples to get started with `diffsptk` in your browser.
 - [**Conference Paper**](https://www.isca-archive.org/ssw_2023/yoshimura23_ssw.html) - Technical background and implementation details available on the ISCA Archive.
 

diffsptk/functional.py

@@ -3248,7 +3248,12 @@ def zcross(
 
 
 def zerodf(
-    x: Tensor, b: Tensor, frame_period: int = 80, ignore_gain: bool = False
+    x: Tensor,
+    b: Tensor,
+    frame_period: int = 80,
+    ignore_gain: bool = False,
+    zeroth_index: int = 0,
+    mode: str = "direct",
 ) -> Tensor:
     """Apply an all-zero digital filter.
 
@@ -3266,12 +3271,23 @@ def zerodf(
     ignore_gain : bool
         If True, perform filtering without the gain.
 
+    zeroth_index : int >= 0
+        The index of the zeroth coefficient in the filter coefficients.
+
+    mode : ['direct', 'efficient']
+        The implementation mode for time-varying convolution.
+
     Returns
     -------
     out : Tensor [shape=(..., T)]
         The output signal.
 
     """
     return nn.AllZeroDigitalFilter._func(
-        x, b, frame_period=frame_period, ignore_gain=ignore_gain
+        x,
+        b,
+        frame_period=frame_period,
+        ignore_gain=ignore_gain,
+        zeroth_index=zeroth_index,
+        mode=mode,
     )

diffsptk/modules/mglsadf.py

@@ -35,6 +35,7 @@
 from .mgc2sp import MelGeneralizedCepstrumToSpectrum
 from .root_pol import PolynomialToRoots
 from .stft import ShortTimeFourierTransform
+from .zerodf import AllZeroDigitalFilter
 
 
 def is_array_like(x: Any) -> bool:
@@ -277,18 +278,18 @@ def __init__(
         if alpha == 0 and gamma == 0:
             cep_order = filter_order
 
-        # Prepare padding module.
         if self.phase == "minimum":
-            padding = (cep_order, 0)
+            cep_orders = (cep_order, 0)
         elif self.phase == "maximum":
-            padding = (0, cep_order)
+            cep_orders = (0, cep_order)
         elif self.phase == "zero":
-            padding = (cep_order, cep_order)
+            cep_orders = (cep_order, cep_order)
         elif self.phase == "mixed":
-            padding = cep_order if is_array_like(cep_order) else (cep_order, cep_order)
+            cep_orders = (
+                cep_order if is_array_like(cep_order) else (cep_order, cep_order)
+            )
         else:
             raise ValueError(f"phase {phase} is not supported.")
-        self.pad = nn.ConstantPad1d(padding, 0)
 
         # Prepare frequency transformation module.
         if self.phase == "mixed":
@@ -297,7 +298,7 @@ def __init__(
                 self.mgc2c.append(
                     MelGeneralizedCepstrumToMelGeneralizedCepstrum(
                         filter_order[i],
-                        padding[i],
+                        cep_orders[i],
                         in_alpha=alpha,
                         in_gamma=gamma,
                         n_fft=n_fft,
@@ -318,6 +319,16 @@ def __init__(
 
         self.linear_intpl = LinearInterpolation(frame_period)
 
+        self.zerodf = AllZeroDigitalFilter(
+            sum(cep_orders),
+            frame_period,
+            ignore_gain=False,
+            zeroth_index=cep_orders[1],
+            mode="efficient",
+            device=device,
+            dtype=dtype,
+        )
+
         cp = mp.taylor(mp.exp, 0, taylor_order)
         cp = np.array([float(x) for x in cp])
         weights = cp[1:] / cp[:-1]
@@ -341,29 +352,25 @@ def forward(
             c_min = self.mgc2c[0](mc_min)
             c_max = self.mgc2c[1](mc_max)
             c0 = c_min[..., :1] + c_max[..., :1]
-            c1_min = c_min[..., 1:].flip(-1)
+            c1_min = c_min[..., 1:]
             c0_dummy = torch.zeros_like(c0)
-            c1_max = c_max[..., 1:]
-            c = torch.cat([c1_min, c0_dummy, c1_max], dim=-1)
+            c1_max = c_max[..., 1:].flip(-1)
+            c = torch.cat([c1_max, c0_dummy, c1_min], dim=-1)
         else:
             c = self.mgc2c(mc)
             c0, c = remove_gain(c, value=0, return_gain=True)
             if self.phase == "minimum":
-                c = c.flip(-1)
-            elif self.phase == "maximum":
                 pass
+            elif self.phase == "maximum":
+                c = c.flip(-1)
             elif self.phase == "zero":
                 c = mirror(c, half=True)
             else:
                 raise RuntimeError
 
-        c = self.linear_intpl(c)
-
         y = x * self.a[0]
         for i in range(1, len(self.a)):
-            x = self.pad(x)
-            x = x.unfold(-1, c.size(-1), 1)
-            x = (x * c).sum(-1) * self.weights[i]
+            x = self.zerodf(x, c) * self.weights[i]
             y += x * self.a[i]
 
         if not self.ignore_gain:
@@ -389,28 +396,26 @@ def __init__(
     ) -> None:
         super().__init__()
 
+        self.frame_period = frame_period
         self.ignore_gain = ignore_gain
         self.phase = phase
         self.n_fft = n_fft
 
-        # Prepare padding module.
-        taps = ir_length - 1
         if self.phase == "minimum":
-            padding = (taps, 0)
+            ir_orders = (ir_length - 1, 0)
         elif self.phase == "maximum":
-            padding = (0, taps)
+            ir_orders = (0, ir_length - 1)
         elif self.phase == "zero":
-            padding = (taps, taps)
+            ir_orders = (ir_length - 1, ir_length - 1)
         elif self.phase == "mixed":
-            padding = (
+            ir_orders = (
                 (ir_length[0] - 1, ir_length[1] - 1)
                 if is_array_like(ir_length)
-                else (taps, taps)
+                else (ir_length - 1, ir_length - 1)
             )
         else:
             raise ValueError(f"phase {phase} is not supported.")
-        self.pad = nn.ConstantPad1d(padding, 0)
-        self.padding = padding
+        self.ir_orders = ir_orders
 
         if self.phase in ("minimum", "maximum"):
             self.mgc2ir = MelGeneralizedCepstrumToMelGeneralizedCepstrum(
@@ -444,7 +449,7 @@ def __init__(
                 self.mgc2c.append(
                     MelGeneralizedCepstrumToMelGeneralizedCepstrum(
                         filter_order[i],
-                        padding[i],
+                        ir_orders[i],
                         in_alpha=alpha,
                         in_gamma=gamma,
                         n_fft=n_fft,
@@ -458,7 +463,15 @@ def __init__(
         else:
             raise ValueError(f"phase {phase} is not supported.")
 
-        self.linear_intpl = LinearInterpolation(frame_period)
+        self.zerodf = AllZeroDigitalFilter(
+            sum(ir_orders),
+            frame_period,
+            ignore_gain=False,
+            zeroth_index=ir_orders[1],
+            mode="efficient",
+            device=device,
+            dtype=dtype,
+        )
 
     def forward(
         self,
@@ -467,9 +480,13 @@ def forward(
     ) -> torch.Tensor:
         if self.phase == "minimum":
             h = self.mgc2ir(mc)
-            h = h.flip(-1)
+            if self.ignore_gain:
+                h = h / h[..., :1]
         elif self.phase == "maximum":
             h = self.mgc2ir(mc)
+            if self.ignore_gain:
+                h = h / h[..., :1]
+            h = h.flip(-1)
         elif self.phase == "zero":
             c = self.mgc2c(mc)
             c[..., 1:] *= 0.5
@@ -485,25 +502,16 @@ def forward(
                 c0 = torch.zeros_like(c_min[..., :1])
             else:
                 c0 = c_min[..., :1] + c_max[..., :1]
-            c = torch.cat([c_min[..., 1:].flip(-1), c0, c_max[..., 1:]], dim=-1)
+            c = torch.cat([c_max[..., 1:].flip(-1), c0, c_min[..., 1:]], dim=-1)
             c = F.pad(c, (0, self.n_fft - c.size(-1)))
-            c = torch.roll(c, -self.padding[0], dims=-1)
+            shift = self.ir_orders[1]
+            c = torch.roll(c, -shift, dims=-1)
             h = self.c2ir(c)
-            h = torch.roll(h, self.padding[0], dims=-1)[..., : sum(self.padding) + 1]
+            h = torch.roll(h, shift, dims=-1)[..., : sum(self.ir_orders) + 1]
         else:
             raise RuntimeError
 
-        h = self.linear_intpl(h)
-
-        if self.ignore_gain:
-            if self.phase == "minimum":
-                h = h / h[..., -1:]
-            elif self.phase == "maximum":
-                h = h / h[..., :1]
-
-        x = self.pad(x)
-        x = x.unfold(-1, h.size(-1), 1)
-        y = (x * h).sum(-1)
+        y = self.zerodf(x, h)
         return y