refactor quantize

diffsptk/functional.py

@@ -86,6 +86,44 @@ def delay(x, start=0, keeplen=False, dim=-1):
     return fm.Delay._forward(x, start=start, keeplen=keeplen, dim=dim)
 
 
+def dequantize(y, abs_max=1, n_bit=8, quantizer="mid-rise"):
+    """Dequantize input.
+
+    Parameters
+    ----------
+    y : Tensor [shape=(...,)]
+        Quantized input.
+
+    abs_max : float > 0
+        Absolute maximum value of input.
+
+    n_bit : int >= 1
+        Number of quantization bits.
+
+    quantizer : ['mid-rise', 'mid-tread']
+        Quantizer.
+
+    Returns
+    -------
+    Tensor [shape=(...,)]
+        Dequantized input.
+
+    Examples
+    --------
+    >>> x = diffsptk.ramp(-4, 4)
+    >>> x
+    tensor([-4., -3., -2., -1.,  0.,  1.,  2.,  3.,  4.])
+    >>> y = diffsptk.functional.quantize(x, 4, 2)
+    >>> z = diffsptk.functional.dequantize(y, 4, 2)
+    >>> z
+    tensor([-3., -3., -1., -1.,  1.,  1.,  3.,  3.,  3.])
+
+    """
+    return fm.InverseUniformQuantization_.forward(
+        y, abs_max=abs_max, n_bit=n_bit, quantizer=quantizer
+    )
+
+
 def grpdelay(b=None, a=None, *, fft_length=512, alpha=1, gamma=1, **kwargs):
     """Compute group delay.
 
@@ -190,6 +228,41 @@ def phase(b=None, a=None, *, fft_length=512, unwrap=False):
     return fm.Phase._forward(b, a, fft_length=fft_length, unwrap=unwrap)
 
 
+def quantize(x, abs_max=1, n_bit=8, quantizer="mid-rise"):
+    """Quantize input.
+
+    Parameters
+    ----------
+    x : Tensor [shape=(...,)]
+        Input.
+
+    abs_max : float > 0
+        Absolute maximum value of input.
+
+    n_bit : int >= 1
+        Number of quantization bits.
+
+    quantizer : ['mid-rise', 'mid-tread']
+        Quantizer.
+
+    Returns
+    -------
+    Tensor [shape=(...,)]
+        Quantized input.
+
+    Examples
+    --------
+    >>> x = diffsptk.ramp(-4, 4)
+    >>> y = diffsptk.functional.quantize(4, 2)
+    >>> y
+    tensor([0, 0, 1, 1, 2, 2, 3, 3, 3], dtype=torch.int32)
+
+    """
+    return fm.UniformQuantization._forward(
+        x, abs_max=abs_max, n_bit=n_bit, quantizer=quantizer
+    )
+
+
 def spec(
     b=None, a=None, *, fft_length=512, eps=0, relative_floor=None, out_format="power"
 ):

diffsptk/misc/utils.py

@@ -78,6 +78,14 @@ def to_3d(x):
     return y
 
 
+def check(x, names):
+    if type(x) is int and 0 <= x <= len(names):
+        x = list(names)[x]
+    if x not in names:
+        raise ValueError(f"Unsupported value: {x}")
+    return x
+
+
 def reflect(x):
     d = x.size(-1)
     y = x.view(-1, d)

diffsptk/modules/dequantize.py

@@ -17,17 +17,20 @@
 import torch
 import torch.nn as nn
 
+from ..misc.utils import check
+from .quantize import _quantization_level
+
 
 class InverseUniformQuantization(nn.Module):
     """See `this page <https://sp-nitech.github.io/sptk/latest/main/dequantize.html>`_
     for details.
 
     Parameters
     ----------
-    abs_max : float > 0 [scalar]
+    abs_max : float > 0
         Absolute maximum value of input.
 
-    n_bit : int >= 1 [scalar]
+    n_bit : int >= 1
         Number of quantization bits.
 
     quantizer : ['mid-rise', 'mid-tread']
@@ -39,19 +42,11 @@ def __init__(self, abs_max=1, n_bit=8, quantizer="mid-rise"):
         super(InverseUniformQuantization, self).__init__()
 
         self.abs_max = abs_max
-        self.quantizer = quantizer
+        self.n_bit = n_bit
+        self.quantizer = check(quantizer, _quantization_level)
 
         assert 0 < self.abs_max
-        assert 1 <= n_bit
-
-        if quantizer == 0 or quantizer == "mid-rise":
-            self.level = int(2**n_bit)
-            self.quantizer = "mid-rise"
-        elif quantizer == 1 or quantizer == "mid-tread":
-            self.level = int(2**n_bit) - 1
-            self.quantizer = "mid-tread"
-        else:
-            raise ValueError("quantizer {quantizer} is not supported")
+        assert 1 <= self.n_bit
 
     def forward(self, y):
         """Dequantize input.
@@ -63,7 +58,7 @@ def forward(self, y):
 
         Returns
         -------
-        x : Tensor [shape=(...,)]
+        Tensor [shape=(...,)]
             Dequantized input.
 
         Examples
@@ -78,13 +73,21 @@ def forward(self, y):
         tensor([-3., -3., -1., -1.,  1.,  1.,  3.,  3.,  3.])
 
         """
-        if self.quantizer == "mid-rise":
-            y = y - (self.level // 2 - 0.5)
-        elif self.quantizer == "mid-tread":
-            y = y - (self.level - 1) // 2
+        return self._forward(y, self.abs_max, self.n_bit, self.quantizer)
+
+    @staticmethod
+    def _forward(y, abs_max, n_bit, quantizer):
+        try:
+            level = _quantization_level[quantizer](n_bit)
+        except KeyError:
+            raise ValueError(f"quantizer {quantizer} is not supported")
+
+        if quantizer == "mid-rise":
+            y = y - (level // 2 - 0.5)
+        elif quantizer == "mid-tread":
+            y = y - (level - 1) // 2
         else:
-            raise RuntimeError
-
-        x = y * (2 * self.abs_max / self.level)
-        x = torch.clip(x, min=-self.abs_max, max=self.abs_max)
+            raise ValueError(f"quantizer {quantizer} is not supported")
+        x = y * (2 * abs_max / level)
+        x = torch.clip(x, min=-abs_max, max=abs_max)
         return x

diffsptk/modules/quantize.py

@@ -17,6 +17,13 @@
 import torch
 import torch.nn as nn
 
+from ..misc.utils import check
+
+_quantization_level = {
+    "mid-rise": lambda n_bit: 1 << n_bit,
+    "mid-tread": lambda n_bit: (1 << n_bit) - 1,
+}
+
 
 class Floor(torch.autograd.Function):
     @staticmethod
@@ -59,19 +66,11 @@ def __init__(self, abs_max=1, n_bit=8, quantizer="mid-rise"):
         super(UniformQuantization, self).__init__()
 
         self.abs_max = abs_max
-        self.quantizer = quantizer
+        self.n_bit = n_bit
+        self.quantizer = check(quantizer, _quantization_level)
 
         assert 0 < self.abs_max
-        assert 1 <= n_bit
-
-        if quantizer == 0 or quantizer == "mid-rise":
-            self.level = int(2**n_bit)
-            self.quantizer = "mid-rise"
-        elif quantizer == 1 or quantizer == "mid-tread":
-            self.level = int(2**n_bit) - 1
-            self.quantizer = "mid-tread"
-        else:
-            raise ValueError("quantizer {quantizer} is not supported")
+        assert 1 <= self.n_bit
 
     def forward(self, x):
         """Quantize input.
@@ -83,7 +82,7 @@ def forward(self, x):
 
         Returns
         -------
-        y : Tensor [shape=(...,)]
+        Tensor [shape=(...,)]
             Quantized input.
 
         Examples
@@ -95,15 +94,23 @@ def forward(self, x):
         tensor([0, 0, 1, 1, 2, 2, 3, 3, 3], dtype=torch.int32)
 
         """
-        x = x * self.level / (2 * self.abs_max)
-        if self.quantizer == "mid-rise":
-            x = x + self.level // 2
+        return self._forward(x, self.abs_max, self.n_bit, self.quantizer)
+
+    @staticmethod
+    def _forward(x, abs_max, n_bit, quantizer):
+        try:
+            level = _quantization_level[quantizer](n_bit)
+        except KeyError:
+            raise ValueError(f"quantizer {quantizer} is not supported")
+
+        x = x * (level / (2 * abs_max))
+        if quantizer == "mid-rise":
+            x += level // 2
             y = Floor.apply(x)
-        elif self.quantizer == "mid-tread":
-            x = x + (self.level - 1) // 2
+        elif quantizer == "mid-tread":
+            x += (level - 1) // 2
             y = Round.apply(x)
         else:
-            raise RuntimeError
-
-        y = torch.clip(y, min=0, max=self.level - 1)
+            raise ValueError(f"quantizer {quantizer} is not supported")
+        y = torch.clip(y, min=0, max=level - 1)
         return y

diffsptk/modules/spec.py

@@ -17,6 +17,7 @@
 import torch
 import torch.nn as nn
 
+from ..misc.utils import check
 from ..misc.utils import remove_gain
 
 _spec2spec = {
@@ -52,6 +53,7 @@ def __init__(self, fft_length, eps=0, relative_floor=None, out_format="power"):
 
         self.fft_length = fft_length
         self.eps = eps
+        self.out_format = check(out_format)
 
         assert 2 <= self.fft_length
         assert 0 <= self.eps
@@ -62,11 +64,6 @@ def __init__(self, fft_length, eps=0, relative_floor=None, out_format="power"):
             assert relative_floor < 0
             self.relative_floor = 10 ** (relative_floor / 10)
 
-        if type(out_format) is int and 0 <= out_format <= len(_spec2spec.keys()):
-            out_format = list(_spec2spec.keys())[out_format]
-        assert out_format in _spec2spec.keys()
-        self.out_format = out_format
-
     def forward(self, b=None, a=None):
         """Compute spectrum.
 
@@ -125,5 +122,8 @@ def _forward(b, a, fft_length, eps, relative_floor, out_format):
         if relative_floor is not None:
             m, _ = torch.max(s, dim=-1, keepdim=True)
             s = torch.maximum(s, m * relative_floor)
-        s = _spec2spec[out_format](s)
+        try:
+            s = _spec2spec[out_format](s)
+        except KeyError:
+            raise ValueError(f"out_format {out_format} is not supported.")
         return s