wehub/tensorflow--tensorflow
1
0
Fork 0
Code Issues Pull Requests Actions 3 Deployments Models Agent Notes Packages Projects Releases Wiki Activity

chore: import upstream snapshot with attribution
cffconvert / validate (push) Has been skipped
Details
License Check / license-check (push) Failing after 2s
Details

Browse Source
This commit is contained in:
wehub-resource-sync
2026-07-13 12:14:16 +08:00
commit 8a852e4b4e
36502 changed files with 9277225 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tensorflow/python/ops/signal/BUILD
+185
View File
@@ -0,0 +1,185 @@
load("@xla//third_party/rules_python/python:py_library.bzl", "py_library")
package(
# copybara:uncomment default_applicable_licenses = ["//tensorflow:license"],
default_visibility = ["//tensorflow:internal"],
licenses = ["notice"],
)
py_library(
name = "dct_ops",
srcs = ["dct_ops.py"],
strict_deps = True,
deps = [
":fft_ops",
"//tensorflow/python/framework:dtypes",
"//tensorflow/python/framework:ops",
"//tensorflow/python/framework:smart_cond",
"//tensorflow/python/framework:tensor_shape",
"//tensorflow/python/ops:array_ops",
"//tensorflow/python/ops:math_ops",
"//tensorflow/python/util:dispatch",
"//tensorflow/python/util:tf_export",
],
)
py_library(
name = "__init__",
srcs = ["__init__.py"],
strict_deps = True,
)
py_library(
name = "fft_ops",
srcs = ["fft_ops.py"],
strict_deps = True,
deps = [
"//tensorflow/python/framework:dtypes",
"//tensorflow/python/framework:ops",
"//tensorflow/python/framework:tensor_util",
"//tensorflow/python/ops:array_ops",
"//tensorflow/python/ops:array_ops_stack",
"//tensorflow/python/ops:manip_ops",
"//tensorflow/python/ops:math_ops",
"//tensorflow/python/ops:spectral_ops_gen",
"//tensorflow/python/util:dispatch",
"//tensorflow/python/util:tf_export",
"//third_party/py/numpy",
],
)
py_library(
name = "mel_ops",
srcs = ["mel_ops.py"],
strict_deps = True,
deps = [
":shape_ops",
"//tensorflow/python/framework:dtypes",
"//tensorflow/python/framework:ops",
"//tensorflow/python/framework:tensor",
"//tensorflow/python/framework:tensor_util",
"//tensorflow/python/ops:array_ops",
"//tensorflow/python/ops:math_ops",
"//tensorflow/python/util:dispatch",
"//tensorflow/python/util:tf_export",
],
)
py_library(
name = "mfcc_ops",
srcs = ["mfcc_ops.py"],
strict_deps = True,
deps = [
":dct_ops",
"//tensorflow/python/framework:ops",
"//tensorflow/python/ops:array_ops",
"//tensorflow/python/ops:math_ops",
"//tensorflow/python/util:dispatch",
"//tensorflow/python/util:tf_export",
],
)
py_library(
name = "reconstruction_ops",
srcs = ["reconstruction_ops.py"],
strict_deps = True,
deps = [
"//tensorflow/python/framework:dtypes",
"//tensorflow/python/framework:ops",
"//tensorflow/python/framework:tensor_util",
"//tensorflow/python/ops:array_ops",
"//tensorflow/python/ops:math_ops",
"//tensorflow/python/util:dispatch",
"//tensorflow/python/util:tf_export",
],
)
py_library(
name = "shape_ops",
srcs = ["shape_ops.py"],
strict_deps = True,
deps = [
":util_ops",
"//tensorflow/python/framework:constant_op",
"//tensorflow/python/framework:ops",
"//tensorflow/python/framework:tensor_util",
"//tensorflow/python/ops:array_ops",
"//tensorflow/python/ops:math_ops",
"//tensorflow/python/util:dispatch",
"//tensorflow/python/util:tf_export",
"//third_party/py/numpy",
],
)
py_library(
name = "signal",
srcs = ["signal.py"],
strict_deps = True,
deps = [
":__init__",
":dct_ops",
":fft_ops",
":mel_ops",
":mfcc_ops",
":reconstruction_ops",
":shape_ops",
":spectral_ops",
":window_ops",
],
)
py_library(
name = "spectral_ops",
srcs = ["spectral_ops.py"],
strict_deps = True,
deps = [
":dct_ops",
":fft_ops",
":reconstruction_ops",
":shape_ops",
":window_ops",
"//tensorflow/python/framework:constant_op",
"//tensorflow/python/framework:dtypes",
"//tensorflow/python/framework:ops",
"//tensorflow/python/framework:tensor_util",
"//tensorflow/python/ops:array_ops",
"//tensorflow/python/ops:math_ops",
"//tensorflow/python/util:dispatch",
"//tensorflow/python/util:tf_export",
"//third_party/py/numpy",
],
)
py_library(
name = "util_ops",
srcs = ["util_ops.py"],
strict_deps = True,
deps = [
"//tensorflow/python/framework:ops",
"//tensorflow/python/framework:tensor_util",
"//tensorflow/python/ops:array_ops",
"//tensorflow/python/ops:math_ops",
"//tensorflow/python/ops:while_loop",
],
)
py_library(
name = "window_ops",
srcs = ["window_ops.py"],
strict_deps = True,
visibility = ["//visibility:public"],
deps = [
"//tensorflow/python/framework:constant_op",
"//tensorflow/python/framework:dtypes",
"//tensorflow/python/framework:ops",
"//tensorflow/python/framework:tensor_util",
"//tensorflow/python/ops:array_ops",
"//tensorflow/python/ops:cond",
"//tensorflow/python/ops:math_ops",
"//tensorflow/python/ops:nn_ops",
"//tensorflow/python/ops:special_math_ops",
"//tensorflow/python/util:dispatch",
"//tensorflow/python/util:tf_export",
"//third_party/py/numpy",
],
)
tensorflow/python/ops/signal/__init__.py
+37
View File
@@ -0,0 +1,37 @@
# Copyright 2023 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Signal processing operations.
See the [tf.signal](https://tensorflow.org/api_guides/python/contrib.signal)
guide.
@@frame
@@hamming_window
@@hann_window
@@inverse_stft
@@inverse_stft_window_fn
@@mfccs_from_log_mel_spectrograms
@@linear_to_mel_weight_matrix
@@overlap_and_add
@@stft
[hamming]: https://en.wikipedia.org/wiki/Window_function#Hamming_window
[hann]: https://en.wikipedia.org/wiki/Window_function#Hann_window
[mel]: https://en.wikipedia.org/wiki/Mel_scale
[mfcc]: https://en.wikipedia.org/wiki/Mel-frequency_cepstrum
[stft]: https://en.wikipedia.org/wiki/Short-time_Fourier_transform
API docstring: tensorflow.signal
"""
tensorflow/python/ops/signal/dct_ops.py
+266
View File
@@ -0,0 +1,266 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Discrete Cosine Transform ops."""
import math as _math
from tensorflow.python.framework import dtypes as _dtypes
from tensorflow.python.framework import ops as _ops
from tensorflow.python.framework import smart_cond
from tensorflow.python.framework import tensor_shape
from tensorflow.python.ops import array_ops as _array_ops
from tensorflow.python.ops import math_ops as _math_ops
from tensorflow.python.ops.signal import fft_ops
from tensorflow.python.util import dispatch
from tensorflow.python.util.tf_export import tf_export
def _validate_dct_arguments(input_tensor, dct_type, n, axis, norm):
"""Checks that DCT/IDCT arguments are compatible and well formed."""
if axis != -1:
raise NotImplementedError("axis must be -1. Got: %s" % axis)
if n is not None and n < 1:
raise ValueError("n should be a positive integer or None")
if dct_type not in (1, 2, 3, 4):
raise ValueError("Types I, II, III and IV (I)DCT are supported.")
if dct_type == 1:
if norm == "ortho":
raise ValueError("Normalization is not supported for the Type-I DCT.")
if input_tensor.shape[-1] is not None and input_tensor.shape[-1] < 2:
raise ValueError(
"Type-I DCT requires the dimension to be greater than one.")
if norm not in (None, "ortho"):
raise ValueError(
"Unknown normalization. Expected None or 'ortho', got: %s" % norm)
# TODO(rjryan): Implement `axis` parameter.
@tf_export("signal.dct", v1=["signal.dct", "spectral.dct"])
@dispatch.add_dispatch_support
def dct(input, type=2, n=None, axis=-1, norm=None, name=None): # pylint: disable=redefined-builtin
"""Computes the 1D [Discrete Cosine Transform (DCT)][dct] of `input`.
Types I, II, III and IV are supported.
Type I is implemented using a length `2N` padded `tf.signal.rfft`.
Type II is implemented using a length `2N` padded `tf.signal.rfft`, as
described here: [Type 2 DCT using 2N FFT padded (Makhoul)]
(https://dsp.stackexchange.com/a/10606).
Type III is a fairly straightforward inverse of Type II
(i.e. using a length `2N` padded `tf.signal.irfft`).
Type IV is calculated through 2N length DCT2 of padded signal and
picking the odd indices.
@compatibility(scipy)
Equivalent to [scipy.fftpack.dct]
(https://docs.scipy.org/doc/scipy-1.4.0/reference/generated/scipy.fftpack.dct.html)
for Type-I, Type-II, Type-III and Type-IV DCT.
@end_compatibility
Args:
input: A `[..., samples]` `float32`/`float64` `Tensor` containing the
signals to take the DCT of.
type: The DCT type to perform. Must be 1, 2, 3 or 4.
n: The length of the transform. If length is less than sequence length,
only the first n elements of the sequence are considered for the DCT.
If n is greater than the sequence length, zeros are padded and then
the DCT is computed as usual.
axis: For future expansion. The axis to compute the DCT along. Must be `-1`.
norm: The normalization to apply. `None` for no normalization or `'ortho'`
for orthonormal normalization.
name: An optional name for the operation.
Returns:
A `[..., samples]` `float32`/`float64` `Tensor` containing the DCT of
`input`.
Raises:
ValueError: If `type` is not `1`, `2`, `3` or `4`, `axis` is
not `-1`, `n` is not `None` or greater than 0,
or `norm` is not `None` or `'ortho'`.
ValueError: If `type` is `1` and `norm` is `ortho`.
[dct]: https://en.wikipedia.org/wiki/Discrete_cosine_transform
"""
_validate_dct_arguments(input, type, n, axis, norm)
return _dct_internal(input, type, n, axis, norm, name)
def _dct_internal(input, type=2, n=None, axis=-1, norm=None, name=None): # pylint: disable=redefined-builtin
"""Computes the 1D Discrete Cosine Transform (DCT) of `input`.
This internal version of `dct` does not perform any validation and accepts a
dynamic value for `n` in the form of a rank 0 tensor.
Args:
input: A `[..., samples]` `float32`/`float64` `Tensor` containing the
signals to take the DCT of.
type: The DCT type to perform. Must be 1, 2, 3 or 4.
n: The length of the transform. If length is less than sequence length,
only the first n elements of the sequence are considered for the DCT.
If n is greater than the sequence length, zeros are padded and then
the DCT is computed as usual. Can be an int or rank 0 tensor.
axis: For future expansion. The axis to compute the DCT along. Must be `-1`.
norm: The normalization to apply. `None` for no normalization or `'ortho'`
for orthonormal normalization.
name: An optional name for the operation.
Returns:
A `[..., samples]` `float32`/`float64` `Tensor` containing the DCT of
`input`.
"""
with _ops.name_scope(name, "dct", [input]):
input = _ops.convert_to_tensor(input)
zero = _ops.convert_to_tensor(0.0, dtype=input.dtype)
seq_len = (
tensor_shape.dimension_value(input.shape[-1]) or
_array_ops.shape(input)[-1])
if n is not None:
def truncate_input():
return input[..., 0:n]
def pad_input():
rank = len(input.shape)
padding = [[0, 0] for _ in range(rank)]
padding[rank - 1][1] = n - seq_len
padding = _ops.convert_to_tensor(padding, dtype=_dtypes.int32)
return _array_ops.pad(input, paddings=padding)
input = smart_cond.smart_cond(n <= seq_len, truncate_input, pad_input)
axis_dim = (tensor_shape.dimension_value(input.shape[-1])
or _array_ops.shape(input)[-1])
axis_dim_float = _math_ops.cast(axis_dim, input.dtype)
if type == 1:
dct1_input = _array_ops.concat([input, input[..., -2:0:-1]], axis=-1)
dct1 = _math_ops.real(fft_ops.rfft(dct1_input))
return dct1
if type == 2:
scale = 2.0 * _math_ops.exp(
_math_ops.complex(
zero, -_math_ops.range(axis_dim_float) * _math.pi * 0.5 /
axis_dim_float))
# TODO(rjryan): Benchmark performance and memory usage of the various
# approaches to computing a DCT via the RFFT.
dct2 = _math_ops.real(
fft_ops.rfft(
input, fft_length=[2 * axis_dim])[..., :axis_dim] * scale)
if norm == "ortho":
n1 = 0.5 * _math_ops.rsqrt(axis_dim_float)
n2 = n1 * _math.sqrt(2.0)
# Use tf.pad to make a vector of [n1, n2, n2, n2, ...].
weights = _array_ops.pad(
_array_ops.expand_dims(n1, 0), [[0, axis_dim - 1]],
constant_values=n2)
dct2 *= weights
return dct2
elif type == 3:
if norm == "ortho":
n1 = _math_ops.sqrt(axis_dim_float)
n2 = n1 * _math.sqrt(0.5)
# Use tf.pad to make a vector of [n1, n2, n2, n2, ...].
weights = _array_ops.pad(
_array_ops.expand_dims(n1, 0), [[0, axis_dim - 1]],
constant_values=n2)
input *= weights
else:
input *= axis_dim_float
scale = 2.0 * _math_ops.exp(
_math_ops.complex(
zero,
_math_ops.range(axis_dim_float) * _math.pi * 0.5 /
axis_dim_float))
dct3 = _math_ops.real(
fft_ops.irfft(
scale * _math_ops.complex(input, zero),
fft_length=[2 * axis_dim]))[..., :axis_dim]
return dct3
elif type == 4:
# DCT-2 of 2N length zero-padded signal, unnormalized.
dct2 = _dct_internal(input, type=2, n=2*axis_dim, axis=axis, norm=None)
# Get odd indices of DCT-2 of zero padded 2N signal to obtain
# DCT-4 of the original N length signal.
dct4 = dct2[..., 1::2]
if norm == "ortho":
dct4 *= _math.sqrt(0.5) * _math_ops.rsqrt(axis_dim_float)
return dct4
# TODO(rjryan): Implement `n` and `axis` parameters.
@tf_export("signal.idct", v1=["signal.idct", "spectral.idct"])
@dispatch.add_dispatch_support
def idct(input, type=2, n=None, axis=-1, norm=None, name=None): # pylint: disable=redefined-builtin
"""Computes the 1D [Inverse Discrete Cosine Transform (DCT)][idct] of `input`.
Currently Types I, II, III, IV are supported. Type III is the inverse of
Type II, and vice versa.
Note that you must re-normalize by 1/(2n) to obtain an inverse if `norm` is
not `'ortho'`. That is:
`signal == idct(dct(signal)) * 0.5 / signal.shape[-1]`.
When `norm='ortho'`, we have:
`signal == idct(dct(signal, norm='ortho'), norm='ortho')`.
@compatibility(scipy)
Equivalent to [scipy.fftpack.idct]
(https://docs.scipy.org/doc/scipy-1.4.0/reference/generated/scipy.fftpack.idct.html)
for Type-I, Type-II, Type-III and Type-IV DCT.
@end_compatibility
Args:
input: A `[..., samples]` `float32`/`float64` `Tensor` containing the
signals to take the IDCT of.
type: The IDCT type to perform. Must be 1, 2, 3 or 4.
n: The length of the transform. If length is less than sequence length, only
the first n elements of the sequence are considered for the IDCT. If n is
greater than the sequence length, zeros are padded and then the IDCT is
computed as usual. If `None` (the default), the sequence length is used
unchanged.
axis: For future expansion. The axis to compute the DCT along. Must be `-1`.
norm: The normalization to apply. `None` for no normalization or `'ortho'`
for orthonormal normalization.
name: An optional name for the operation.
Returns:
A `[..., samples]` `float32`/`float64` `Tensor` containing the IDCT of
`input`.
Raises:
ValueError: If `type` is not `1`, `2`, `3` or `4`, `axis` is
not `-1`, `n` is not `None` or greater than 0,
or `norm` is not `None` or `'ortho'`.
ValueError: If `type` is `1` and `norm` is `'ortho'`.
[idct]:
https://en.wikipedia.org/wiki/Discrete_cosine_transform#Inverse_transforms
"""
# The actual implementation routes through `_dct_internal`, which
# already supports an arbitrary positive `n` (truncate / zero-pad on
# the last axis); the docstring previously claimed `n` had to be
# `None`. See https://github.com/tensorflow/tensorflow/issues/102418.
_validate_dct_arguments(input, type, n, axis, norm)
inverse_type = {1: 1, 2: 3, 3: 2, 4: 4}[type]
return _dct_internal(
input, type=inverse_type, n=n, axis=axis, norm=norm, name=name)
tensorflow/python/ops/signal/fft_ops.py
+705
View File
@@ -0,0 +1,705 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Fast-Fourier Transform ops."""
import re
import numpy as np
from tensorflow.python.framework import dtypes as _dtypes
from tensorflow.python.framework import ops as _ops
from tensorflow.python.framework import tensor_util as _tensor_util
from tensorflow.python.ops import array_ops as _array_ops
from tensorflow.python.ops import array_ops_stack as _array_ops_stack
from tensorflow.python.ops import gen_spectral_ops
from tensorflow.python.ops import manip_ops
from tensorflow.python.ops import math_ops as _math_ops
from tensorflow.python.util import dispatch
from tensorflow.python.util.tf_export import tf_export
def _infer_fft_length_for_fftn(input_tensor):
return _array_ops.shape(input_tensor)[-len(input_tensor.shape) :]
def _infer_fft_length_for_irfftn(input_tensor):
fft_shape = input_tensor.get_shape()[-len(input_tensor.shape) :]
fft_length = fft_shape.as_list()
fft_length[-1] = max(0, 2 * (fft_length[-1] - 1))
return _ops.convert_to_tensor(fft_length, _dtypes.int32)
def _infer_axes_for_fftn(input_tensor):
return _ops.convert_to_tensor(
np.arange(len(input_tensor.shape)), _dtypes.int32
)
def _process_empty_axes(input_tensor, axes):
if axes is None:
axes = _infer_axes_for_fftn(input_tensor)
else:
axes = _ops.convert_to_tensor(axes, _dtypes.int32)
return axes
def _infer_fft_length_for_rfft(input_tensor, fft_rank):
"""Infers the `fft_length` argument for a `rank` RFFT from `input_tensor`."""
# A TensorShape for the inner fft_rank dimensions.
fft_shape = input_tensor.get_shape()[-fft_rank:]
# If any dim is unknown, fall back to tensor-based math.
if not fft_shape.is_fully_defined():
return _array_ops.shape(input_tensor)[-fft_rank:]
# Otherwise, return a constant.
return _ops.convert_to_tensor(fft_shape.as_list(), _dtypes.int32)
def _infer_fft_length_for_irfft(input_tensor, fft_rank):
"""Infers the `fft_length` argument for a `rank` IRFFT from `input_tensor`."""
# A TensorShape for the inner fft_rank dimensions.
fft_shape = input_tensor.get_shape()[-fft_rank:]
# If any dim is unknown, fall back to tensor-based math.
if not fft_shape.is_fully_defined():
fft_length = _array_ops_stack.unstack(
_array_ops.shape(input_tensor)[-fft_rank:])
fft_length[-1] = _math_ops.maximum(0, 2 * (fft_length[-1] - 1))
return _array_ops_stack.stack(fft_length)
# Otherwise, return a constant.
fft_length = fft_shape.as_list()
if fft_length:
fft_length[-1] = max(0, 2 * (fft_length[-1] - 1))
return _ops.convert_to_tensor(fft_length, _dtypes.int32)
def _maybe_pad_for_rfft(input_tensor, fft_rank, fft_length, is_reverse=False):
"""Pads `input_tensor` to `fft_length` on its inner-most `fft_rank` dims."""
fft_shape = _tensor_util.constant_value_as_shape(fft_length)
# Edge case: skip padding empty tensors.
if (input_tensor.shape.ndims is not None and
any(dim.value == 0 for dim in input_tensor.shape.dims)):
return input_tensor
# If we know the shapes ahead of time, we can either skip or pre-compute the
# appropriate paddings. Otherwise, fall back to computing paddings in
# TensorFlow.
if fft_shape.is_fully_defined() and input_tensor.shape.ndims is not None:
# Slice the last FFT-rank dimensions from input_tensor's shape.
input_fft_shape = input_tensor.shape[-fft_shape.ndims:] # pylint: disable=invalid-unary-operand-type
if input_fft_shape.is_fully_defined():
# In reverse, we only pad the inner-most dimension to fft_length / 2 + 1.
if is_reverse:
fft_shape = fft_shape[:-1].concatenate(
fft_shape.dims[-1].value // 2 + 1)
paddings = [[0, max(fft_dim.value - input_dim.value, 0)]
for fft_dim, input_dim in zip(
fft_shape.dims, input_fft_shape.dims)]
if any(pad > 0 for _, pad in paddings):
outer_paddings = [[0, 0]] * max((input_tensor.shape.ndims -
fft_shape.ndims), 0)
return _array_ops.pad(input_tensor, outer_paddings + paddings)
return input_tensor
# If we can't determine the paddings ahead of time, then we have to pad. If
# the paddings end up as zero, tf.pad has a special-case that does no work.
input_rank = _array_ops.rank(input_tensor)
input_fft_shape = _array_ops.shape(input_tensor)[-fft_rank:]
outer_dims = _math_ops.maximum(0, input_rank - fft_rank)
outer_paddings = _array_ops.zeros([outer_dims], fft_length.dtype)
# In reverse, we only pad the inner-most dimension to fft_length / 2 + 1.
if is_reverse:
fft_length = _array_ops.concat([fft_length[:-1],
fft_length[-1:] // 2 + 1], 0)
fft_paddings = _math_ops.maximum(0, fft_length - input_fft_shape)
paddings = _array_ops.concat([outer_paddings, fft_paddings], 0)
paddings = _array_ops_stack.stack(
[_array_ops.zeros_like(paddings), paddings], axis=1)
return _array_ops.pad(input_tensor, paddings)
def _rfft_wrapper(fft_fn, fft_rank, default_name):
"""Wrapper around gen_spectral_ops.rfft* that infers fft_length argument."""
def _rfft(input_tensor, fft_length=None, name=None):
"""Wrapper around gen_spectral_ops.rfft* that infers fft_length argument."""
with _ops.name_scope(name, default_name,
[input_tensor, fft_length]) as name:
input_tensor = _ops.convert_to_tensor(input_tensor,
preferred_dtype=_dtypes.float32)
if input_tensor.dtype not in (_dtypes.float32, _dtypes.float64):
raise ValueError(
"RFFT requires tf.float32 or tf.float64 inputs, got: %s" %
input_tensor)
real_dtype = input_tensor.dtype
if real_dtype == _dtypes.float32:
complex_dtype = _dtypes.complex64
else:
assert real_dtype == _dtypes.float64
complex_dtype = _dtypes.complex128
input_tensor.shape.with_rank_at_least(fft_rank)
if fft_length is None:
fft_length = _infer_fft_length_for_rfft(input_tensor, fft_rank)
else:
fft_length = _ops.convert_to_tensor(fft_length, _dtypes.int32)
input_tensor = _maybe_pad_for_rfft(input_tensor, fft_rank, fft_length)
fft_length_static = _tensor_util.constant_value(fft_length)
if fft_length_static is not None:
fft_length = fft_length_static
return fft_fn(input_tensor, fft_length, Tcomplex=complex_dtype, name=name)
if fft_fn.__doc__ is not None:
_rfft.__doc__ = re.sub(" Tcomplex.*?\n", "", fft_fn.__doc__)
return _rfft
def _irfft_wrapper(ifft_fn, fft_rank, default_name):
"""Wrapper around gen_spectral_ops.irfft* that infers fft_length argument."""
def _irfft(input_tensor, fft_length=None, name=None):
"""Wrapper irfft* that infers fft_length argument."""
with _ops.name_scope(name, default_name,
[input_tensor, fft_length]) as name:
input_tensor = _ops.convert_to_tensor(input_tensor,
preferred_dtype=_dtypes.complex64)
input_tensor.shape.with_rank_at_least(fft_rank)
if input_tensor.dtype not in (_dtypes.complex64, _dtypes.complex128):
raise ValueError(
"IRFFT requires tf.complex64 or tf.complex128 inputs, got: %s" %
input_tensor)
complex_dtype = input_tensor.dtype
real_dtype = complex_dtype.real_dtype
if fft_length is None:
fft_length = _infer_fft_length_for_irfft(input_tensor, fft_rank)
else:
fft_length = _ops.convert_to_tensor(fft_length, _dtypes.int32)
input_tensor = _maybe_pad_for_rfft(input_tensor, fft_rank, fft_length,
is_reverse=True)
fft_length_static = _tensor_util.constant_value(fft_length)
if fft_length_static is not None:
fft_length = fft_length_static
return ifft_fn(input_tensor, fft_length, Treal=real_dtype, name=name)
if ifft_fn.__doc__ is not None:
_irfft.__doc__ = re.sub(
"`input`",
"`input_tensor`",
re.sub(" Treal.*?\n", "", ifft_fn.__doc__),
)
return _irfft
def _fftn_wrapper(fft_n, default_name):
"""Wrapper around gen_spectral_ops.fftn."""
def _fftn(input_tensor, fft_length=None, axes=None, norm=None, name=None):
"""Wrapper around gen_spectral_ops.*fft that infers fft_length and axes arguments."""
with _ops.name_scope(
name, default_name, [input_tensor, fft_length, axes]
) as name:
axes = _process_empty_axes(input_tensor, axes)
fft_rank = axes.shape[0]
input_tensor = _ops.convert_to_tensor(
input_tensor, preferred_dtype=_dtypes.complex64
)
input_tensor.shape.with_rank_at_least(fft_rank)
if fft_length is None:
fft_length = _infer_fft_length_for_fftn(input_tensor)
else:
fft_length = _ops.convert_to_tensor(fft_length, _dtypes.int32)
input_tensor = _maybe_pad_for_rfft(input_tensor, fft_rank, fft_length)
fft_length_static = _tensor_util.constant_value(fft_length)
if fft_length_static is not None:
fft_length = fft_length_static
if norm is None:
norm = "backward"
n = 1
if norm != "backward":
for fft_length_i in fft_length:
n *= fft_length_i
if norm == "forward":
input_tensor /= n
elif norm == "ortho":
input_tensor /= np.sqrt(n) # should be sqrt(N)
return fft_n(input_tensor, fft_length, axes, name=name)
if fft_n.__doc__ is not None:
_fftn.__doc__ = re.sub(r" Tcomplex.*?\n", "", fft_n.__doc__)
return _fftn
def _ifftn_wrapper(ifft_n, default_name):
"""Wrapper around gen_spectral_ops.ifftn."""
def _ifftn(input_tensor, fft_length=None, axes=None, norm=None, name=None):
"""Wrapper around gen_spectral_ops.*fft that infers fft_length and axes arguments."""
with _ops.name_scope(
name, default_name, [input_tensor, fft_length, axes]
) as name:
axes = _process_empty_axes(input_tensor, axes)
fft_rank = axes.shape[0]
input_tensor = _ops.convert_to_tensor(
input_tensor, preferred_dtype=_dtypes.complex64
)
input_tensor.shape.with_rank_at_least(fft_rank)
if fft_length is None:
fft_length = _infer_fft_length_for_fftn(input_tensor)
else:
fft_length = _ops.convert_to_tensor(fft_length, _dtypes.int32)
input_tensor = _maybe_pad_for_rfft(input_tensor, fft_rank, fft_length)
fft_length_static = _tensor_util.constant_value(fft_length)
if fft_length_static is not None:
fft_length = fft_length_static
if norm is None:
norm = "backward"
n = 1
if norm != "backward":
for fft_length_i in fft_length:
n *= fft_length_i
if norm == "forward":
input_tensor *= n
elif norm == "ortho":
input_tensor *= np.sqrt(n) # should be sqrt(N)
return ifft_n(input_tensor, fft_length, axes, name=name)
if ifft_n.__doc__ is not None:
_ifftn.__doc__ = re.sub(r" Tcomplex.*?\n", "", ifft_n.__doc__)
return _ifftn
def _rfftn_wrapper(rfft_n, default_name):
"""Wrapper around gen_spectral_ops.rfftn."""
def _rfftn(input_tensor, fft_length=None, axes=None, norm=None, name=None):
"""Wrapper around gen_spectral_ops.*fft that infers fft_length and axes arguments."""
with _ops.name_scope(
name, default_name, [input_tensor, fft_length, axes]
) as name:
axes = _process_empty_axes(input_tensor, axes)
fft_rank = axes.shape[0]
input_tensor = _ops.convert_to_tensor(
input_tensor, preferred_dtype=_dtypes.float32
)
if input_tensor.dtype not in (_dtypes.float32, _dtypes.float64):
raise ValueError(
"RFFT requires tf.float32 or tf.float64 inputs, got: %s"
% input_tensor
)
real_dtype = input_tensor.dtype
if real_dtype == _dtypes.float32:
complex_dtype = _dtypes.complex64
else:
assert real_dtype == _dtypes.float64
complex_dtype = _dtypes.complex128
input_tensor.shape.with_rank_at_least(fft_rank)
if fft_length is None:
fft_length = _infer_fft_length_for_fftn(input_tensor)
else:
fft_length = _ops.convert_to_tensor(fft_length, _dtypes.int32)
input_tensor = _maybe_pad_for_rfft(input_tensor, fft_rank, fft_length)
fft_length_static = _tensor_util.constant_value(fft_length)
if fft_length_static is not None:
fft_length = fft_length_static
if norm is None:
norm = "backward"
n = 1
if norm != "backward":
for fft_length_i in fft_length:
n *= fft_length_i
if norm == "forward":
input_tensor /= n
elif norm == "ortho":
input_tensor /= np.sqrt(n) # should be sqrt(N)
return rfft_n(
input_tensor,
fft_length,
axes,
Tcomplex=complex_dtype,
name=name,
)
if rfft_n.__doc__ is not None:
_rfftn.__doc__ = re.sub(r" Tcomplex.*?\n", "", rfft_n.__doc__)
return _rfftn
def _irfftn_wrapper(irfft_n, default_name):
"""Wrapper around gen_spectral_ops.irfftn."""
def _irfftn(input_tensor, fft_length=None, axes=None, norm=None, name=None):
"""Wrapper irfft* that infers fft_length argument."""
with _ops.name_scope(
name, default_name, [input_tensor, fft_length]
) as name:
axes = _process_empty_axes(input_tensor, axes)
fft_rank = axes.shape[0]
input_tensor = _ops.convert_to_tensor(
input_tensor, preferred_dtype=_dtypes.complex64
)
input_tensor.shape.with_rank_at_least(fft_rank)
if input_tensor.dtype not in (_dtypes.complex64, _dtypes.complex128):
raise ValueError(
"IRFFT requires tf.complex64 or tf.complex128 inputs, got: %s"
% input_tensor
)
complex_dtype = input_tensor.dtype
real_dtype = complex_dtype.real_dtype
if fft_length is None:
fft_length = _infer_fft_length_for_irfftn(input_tensor)
else:
fft_length = _ops.convert_to_tensor(fft_length, _dtypes.int32)
input_tensor = _maybe_pad_for_rfft(
input_tensor, fft_rank, fft_length, is_reverse=True
)
fft_length_static = _tensor_util.constant_value(fft_length)
if fft_length_static is not None:
fft_length = fft_length_static
if norm is None:
norm = "backward"
n = 1
if norm != "backward":
for fft_length_i in fft_length:
n *= fft_length_i
if norm == "forward":
input_tensor *= n
elif norm == "ortho":
input_tensor *= np.sqrt(n) # should be sqrt(N)
return irfft_n(
input_tensor, fft_length, axes, Treal=real_dtype, name=name
)
if irfft_n.__doc__ is not None:
_irfftn.__doc__ = re.sub(
"`input`",
"`input_tensor`",
re.sub(r" Treal.*?\n", "", irfft_n.__doc__),
)
return _irfftn
# FFT/IFFT 1/2/3D are exported via
# third_party/tensorflow/core/api_def/python_api/
fft = gen_spectral_ops.fft
ifft = gen_spectral_ops.ifft
fft2d = gen_spectral_ops.fft2d
ifft2d = gen_spectral_ops.ifft2d
fft3d = gen_spectral_ops.fft3d
ifft3d = gen_spectral_ops.ifft3d
fftnd = _fftn_wrapper(gen_spectral_ops.fftnd, "fftnd")
tf_export("signal.fftnd")(
dispatch.add_dispatch_support(fftnd)
)
ifftnd = _ifftn_wrapper(gen_spectral_ops.ifftnd, "ifftnd")
tf_export("signal.ifftnd")(
dispatch.add_dispatch_support(ifftnd)
)
rfft = _rfft_wrapper(gen_spectral_ops.rfft, 1, "rfft")
tf_export("signal.rfft", v1=["signal.rfft", "spectral.rfft"])(
dispatch.add_dispatch_support(rfft))
irfft = _irfft_wrapper(gen_spectral_ops.irfft, 1, "irfft")
tf_export("signal.irfft", v1=["signal.irfft", "spectral.irfft"])(
dispatch.add_dispatch_support(irfft))
rfft2d = _rfft_wrapper(gen_spectral_ops.rfft2d, 2, "rfft2d")
tf_export("signal.rfft2d", v1=["signal.rfft2d", "spectral.rfft2d"])(
dispatch.add_dispatch_support(rfft2d))
irfft2d = _irfft_wrapper(gen_spectral_ops.irfft2d, 2, "irfft2d")
tf_export("signal.irfft2d", v1=["signal.irfft2d", "spectral.irfft2d"])(
dispatch.add_dispatch_support(irfft2d))
rfft3d = _rfft_wrapper(gen_spectral_ops.rfft3d, 3, "rfft3d")
tf_export("signal.rfft3d", v1=["signal.rfft3d", "spectral.rfft3d"])(
dispatch.add_dispatch_support(rfft3d))
irfft3d = _irfft_wrapper(gen_spectral_ops.irfft3d, 3, "irfft3d")
tf_export("signal.irfft3d", v1=["signal.irfft3d", "spectral.irfft3d"])(
dispatch.add_dispatch_support(irfft3d))
rfftnd = _rfftn_wrapper(gen_spectral_ops.rfftnd, "rfftnd")
tf_export("signal.rfftnd")(
dispatch.add_dispatch_support(rfftnd)
)
irfftnd = _irfftn_wrapper(gen_spectral_ops.irfftnd, "irfftnd")
tf_export("signal.irfftnd")(
dispatch.add_dispatch_support(irfftnd)
)
def _fft_size_for_grad(grad, rank):
return _math_ops.reduce_prod(_array_ops.shape(grad)[-rank:])
@_ops.RegisterGradient("FFT")
def _fft_grad(_, grad):
size = _math_ops.cast(_fft_size_for_grad(grad, 1), grad.dtype)
return ifft(grad) * size
@_ops.RegisterGradient("IFFT")
def _ifft_grad(_, grad):
rsize = _math_ops.cast(
1. / _math_ops.cast(_fft_size_for_grad(grad, 1), grad.dtype.real_dtype),
grad.dtype)
return fft(grad) * rsize
@_ops.RegisterGradient("FFT2D")
def _fft2d_grad(_, grad):
size = _math_ops.cast(_fft_size_for_grad(grad, 2), grad.dtype)
return ifft2d(grad) * size
@_ops.RegisterGradient("IFFT2D")
def _ifft2d_grad(_, grad):
rsize = _math_ops.cast(
1. / _math_ops.cast(_fft_size_for_grad(grad, 2), grad.dtype.real_dtype),
grad.dtype)
return fft2d(grad) * rsize
@_ops.RegisterGradient("FFT3D")
def _fft3d_grad(_, grad):
size = _math_ops.cast(_fft_size_for_grad(grad, 3), grad.dtype)
return ifft3d(grad) * size
@_ops.RegisterGradient("IFFT3D")
def _ifft3d_grad(_, grad):
rsize = _math_ops.cast(
1. / _math_ops.cast(_fft_size_for_grad(grad, 3), grad.dtype.real_dtype),
grad.dtype)
return fft3d(grad) * rsize
def _rfft_grad_helper(rank, irfft_fn):
"""Returns a gradient function for an RFFT of the provided rank."""
# Can't happen because we don't register a gradient for RFFT3D.
assert rank in (1, 2), "Gradient for RFFT3D is not implemented."
def _grad(op, grad):
"""A gradient function for RFFT with the provided `rank` and `irfft_fn`."""
fft_length = op.inputs[1]
complex_dtype = grad.dtype
real_dtype = complex_dtype.real_dtype
input_shape = _array_ops.shape(op.inputs[0])
is_even = _math_ops.cast(1 - (fft_length[-1] % 2), complex_dtype)
def _tile_for_broadcasting(matrix, t):
expanded = _array_ops.reshape(
matrix,
_array_ops.concat([
_array_ops.ones([_array_ops.rank(t) - 2], _dtypes.int32),
_array_ops.shape(matrix)
], 0))
return _array_ops.tile(
expanded, _array_ops.concat([_array_ops.shape(t)[:-2], [1, 1]], 0))
def _mask_matrix(length):
"""Computes t_n = exp(sqrt(-1) * pi * n^2 / line_len)."""
# TODO(rjryan): Speed up computation of twiddle factors using the
# following recurrence relation and cache them across invocations of RFFT.
#
# t_n = exp(sqrt(-1) * pi * n^2 / line_len)
# for n = 0, 1,..., line_len-1.
# For n > 2, use t_n = t_{n-1}^2 / t_{n-2} * t_1^2
a = _array_ops.tile(
_array_ops.expand_dims(_math_ops.range(length), 0), (length, 1))
b = _array_ops.transpose(a, [1, 0])
return _math_ops.exp(
-2j * np.pi * _math_ops.cast(a * b, complex_dtype) /
_math_ops.cast(length, complex_dtype))
def _ymask(length):
"""A sequence of [1+0j, -1+0j, 1+0j, -1+0j, ...] with length `length`."""
return _math_ops.cast(1 - 2 * (_math_ops.range(length) % 2),
complex_dtype)
y0 = grad[..., 0:1]
if rank == 1:
ym = grad[..., -1:]
extra_terms = y0 + is_even * ym * _ymask(input_shape[-1])
elif rank == 2:
# Create a mask matrix for y0 and ym.
base_mask = _mask_matrix(input_shape[-2])
# Tile base_mask to match y0 in shape so that we can batch-matmul the
# inner 2 dimensions.
tiled_mask = _tile_for_broadcasting(base_mask, y0)
y0_term = _math_ops.matmul(tiled_mask, _math_ops.conj(y0))
extra_terms = y0_term
ym = grad[..., -1:]
ym_term = _math_ops.matmul(tiled_mask, _math_ops.conj(ym))
inner_dim = input_shape[-1]
ym_term = _array_ops.tile(
ym_term,
_array_ops.concat([
_array_ops.ones([_array_ops.rank(grad) - 1], _dtypes.int32),
[inner_dim]
], 0)) * _ymask(inner_dim)
extra_terms += is_even * ym_term
# The gradient of RFFT is the IRFFT of the incoming gradient times a scaling
# factor, plus some additional terms to make up for the components dropped
# due to Hermitian symmetry.
input_size = _math_ops.cast(
_fft_size_for_grad(op.inputs[0], rank), real_dtype)
the_irfft = irfft_fn(grad, fft_length)
return 0.5 * (the_irfft * input_size + _math_ops.real(extra_terms)), None
return _grad
def _irfft_grad_helper(rank, rfft_fn):
"""Returns a gradient function for an IRFFT of the provided rank."""
# Can't happen because we don't register a gradient for IRFFT3D.
assert rank in (1, 2), "Gradient for IRFFT3D is not implemented."
def _grad(op, grad):
"""A gradient function for IRFFT with the provided `rank` and `rfft_fn`."""
# Generate a simple mask like [1.0, 2.0, ..., 2.0, 1.0] for even-length FFTs
# and [1.0, 2.0, ..., 2.0] for odd-length FFTs. To reduce extra ops in the
# graph we special-case the situation where the FFT length and last
# dimension of the input are known at graph construction time.
fft_length = op.inputs[1]
fft_length_static = _tensor_util.constant_value(fft_length)
if fft_length_static is not None:
fft_length = fft_length_static
real_dtype = grad.dtype
if real_dtype == _dtypes.float32:
complex_dtype = _dtypes.complex64
elif real_dtype == _dtypes.float64:
complex_dtype = _dtypes.complex128
is_odd = _math_ops.mod(fft_length[-1], 2)
input_last_dimension = _array_ops.shape(op.inputs[0])[-1]
mask = _array_ops.concat(
[[1.0], 2.0 * _array_ops.ones(
[input_last_dimension - 2 + is_odd], real_dtype),
_array_ops.ones([1 - is_odd], real_dtype)], 0)
rsize = _math_ops.reciprocal(_math_ops.cast(
_fft_size_for_grad(grad, rank), real_dtype))
# The gradient of IRFFT is the RFFT of the incoming gradient times a scaling
# factor and a mask. The mask scales the gradient for the Hermitian
# symmetric components of the RFFT by a factor of two, since these
# components are de-duplicated in the RFFT.
the_rfft = rfft_fn(grad, fft_length)
return the_rfft * _math_ops.cast(rsize * mask, complex_dtype), None
return _grad
@tf_export("signal.fftshift")
@dispatch.add_dispatch_support
def fftshift(x, axes=None, name=None):
"""Shift the zero-frequency component to the center of the spectrum.
This function swaps half-spaces for all axes listed (defaults to all).
Note that ``y[0]`` is the Nyquist component only if ``len(x)`` is even.
@compatibility(numpy)
Equivalent to numpy.fft.fftshift.
https://docs.scipy.org/doc/numpy/reference/generated/numpy.fft.fftshift.html
@end_compatibility
For example:
```python
x = tf.signal.fftshift([ 0., 1., 2., 3., 4., -5., -4., -3., -2., -1.])
x.numpy() # array([-5., -4., -3., -2., -1., 0., 1., 2., 3., 4.])
```
Args:
x: `Tensor`, input tensor.
axes: `int` or shape `tuple`, optional Axes over which to shift. Default is
None, which shifts all axes.
name: An optional name for the operation.
Returns:
A `Tensor`, The shifted tensor.
"""
with _ops.name_scope(name, "fftshift") as name:
x = _ops.convert_to_tensor(x)
if axes is None:
axes = tuple(range(x.shape.ndims))
shift = _array_ops.shape(x) // 2
elif isinstance(axes, int):
shift = _array_ops.shape(x)[axes] // 2
else:
rank = _array_ops.rank(x)
# allows negative axis
axes = _array_ops.where(_math_ops.less(axes, 0), axes + rank, axes)
shift = _array_ops.gather(_array_ops.shape(x), axes) // 2
return manip_ops.roll(x, shift, axes, name)
@tf_export("signal.ifftshift")
@dispatch.add_dispatch_support
def ifftshift(x, axes=None, name=None):
"""The inverse of fftshift.
Although identical for even-length x,
the functions differ by one sample for odd-length x.
@compatibility(numpy)
Equivalent to numpy.fft.ifftshift.
https://docs.scipy.org/doc/numpy/reference/generated/numpy.fft.ifftshift.html
@end_compatibility
For example:
```python
x = tf.signal.ifftshift([[ 0., 1., 2.],[ 3., 4., -4.],[-3., -2., -1.]])
x.numpy() # array([[ 4., -4., 3.],[-2., -1., -3.],[ 1., 2., 0.]])
```
Args:
x: `Tensor`, input tensor.
axes: `int` or shape `tuple` Axes over which to calculate. Defaults to None,
which shifts all axes.
name: An optional name for the operation.
Returns:
A `Tensor`, The shifted tensor.
"""
with _ops.name_scope(name, "ifftshift") as name:
x = _ops.convert_to_tensor(x)
if axes is None:
axes = tuple(range(x.shape.ndims))
shift = -(_array_ops.shape(x) // 2)
elif isinstance(axes, int):
shift = -(_array_ops.shape(x)[axes] // 2)
else:
rank = _array_ops.rank(x)
# allows negative axis
axes = _array_ops.where(_math_ops.less(axes, 0), axes + rank, axes)
shift = -(_array_ops.gather(_array_ops.shape(x), axes) // 2)
return manip_ops.roll(x, shift, axes, name)
_ops.RegisterGradient("RFFT")(_rfft_grad_helper(1, irfft))
_ops.RegisterGradient("IRFFT")(_irfft_grad_helper(1, rfft))
_ops.RegisterGradient("RFFT2D")(_rfft_grad_helper(2, irfft2d))
_ops.RegisterGradient("IRFFT2D")(_irfft_grad_helper(2, rfft2d))
tensorflow/python/ops/signal/mel_ops.py
+216
View File
@@ -0,0 +1,216 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""mel conversion ops."""
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor
from tensorflow.python.framework import tensor_util
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops.signal import shape_ops
from tensorflow.python.util import dispatch
from tensorflow.python.util.tf_export import tf_export
# mel spectrum constants.
_MEL_BREAK_FREQUENCY_HERTZ = 700.0
_MEL_HIGH_FREQUENCY_Q = 1127.0
def _mel_to_hertz(mel_values, name=None):
"""Converts frequencies in `mel_values` from the mel scale to linear scale.
Args:
mel_values: A `Tensor` of frequencies in the mel scale.
name: An optional name for the operation.
Returns:
A `Tensor` of the same shape and type as `mel_values` containing linear
scale frequencies in Hertz.
"""
with ops.name_scope(name, 'mel_to_hertz', [mel_values]):
mel_values = ops.convert_to_tensor(mel_values)
return _MEL_BREAK_FREQUENCY_HERTZ * (
math_ops.exp(mel_values / _MEL_HIGH_FREQUENCY_Q) - 1.0
)
def _hertz_to_mel(frequencies_hertz, name=None):
"""Converts frequencies in `frequencies_hertz` in Hertz to the mel scale.
Args:
frequencies_hertz: A `Tensor` of frequencies in Hertz.
name: An optional name for the operation.
Returns:
A `Tensor` of the same shape and type of `frequencies_hertz` containing
frequencies in the mel scale.
"""
with ops.name_scope(name, 'hertz_to_mel', [frequencies_hertz]):
frequencies_hertz = ops.convert_to_tensor(frequencies_hertz)
return _MEL_HIGH_FREQUENCY_Q * math_ops.log(
1.0 + (frequencies_hertz / _MEL_BREAK_FREQUENCY_HERTZ))
def _validate_arguments(num_mel_bins, sample_rate,
lower_edge_hertz, upper_edge_hertz, dtype):
"""Checks the inputs to linear_to_mel_weight_matrix."""
if num_mel_bins <= 0:
raise ValueError('num_mel_bins must be positive. Got: %s' % num_mel_bins)
if lower_edge_hertz < 0.0:
raise ValueError('lower_edge_hertz must be non-negative. Got: %s' %
lower_edge_hertz)
if lower_edge_hertz >= upper_edge_hertz:
raise ValueError('lower_edge_hertz %.1f >= upper_edge_hertz %.1f' %
(lower_edge_hertz, upper_edge_hertz))
if not isinstance(sample_rate, tensor.Tensor):
if sample_rate <= 0.0:
raise ValueError('sample_rate must be positive. Got: %s' % sample_rate)
if upper_edge_hertz > sample_rate / 2:
raise ValueError('upper_edge_hertz must not be larger than the Nyquist '
'frequency (sample_rate / 2). Got %s for sample_rate: %s'
% (upper_edge_hertz, sample_rate))
if not dtype.is_floating:
raise ValueError('dtype must be a floating point type. Got: %s' % dtype)
@tf_export('signal.linear_to_mel_weight_matrix')
@dispatch.add_dispatch_support
def linear_to_mel_weight_matrix(num_mel_bins=20,
num_spectrogram_bins=129,
sample_rate=8000,
lower_edge_hertz=125.0,
upper_edge_hertz=3800.0,
dtype=dtypes.float32,
name=None):
r"""Returns a matrix to warp linear scale spectrograms to the [mel scale][mel].
Returns a weight matrix that can be used to re-weight a `Tensor` containing
`num_spectrogram_bins` linearly sampled frequency information from
`[0, sample_rate / 2]` into `num_mel_bins` frequency information from
`[lower_edge_hertz, upper_edge_hertz]` on the [mel scale][mel].
This function follows the [Hidden Markov Model Toolkit
(HTK)](http://htk.eng.cam.ac.uk/) convention, defining the mel scale in
terms of a frequency in hertz according to the following formula:
$$\textrm{mel}(f) = 2595 * \textrm{log}_{10}(1 + \frac{f}{700})$$
In the returned matrix, all the triangles (filterbanks) have a peak value
of 1.0.
For example, the returned matrix `A` can be used to right-multiply a
spectrogram `S` of shape `[frames, num_spectrogram_bins]` of linear
scale spectrum values (e.g. STFT magnitudes) to generate a "mel spectrogram"
`M` of shape `[frames, num_mel_bins]`.
# `S` has shape [frames, num_spectrogram_bins]
# `M` has shape [frames, num_mel_bins]
M = tf.matmul(S, A)
The matrix can be used with `tf.tensordot` to convert an arbitrary rank
`Tensor` of linear-scale spectral bins into the mel scale.
# S has shape [..., num_spectrogram_bins].
# M has shape [..., num_mel_bins].
M = tf.tensordot(S, A, 1)
Args:
num_mel_bins: Python int. How many bands in the resulting mel spectrum.
num_spectrogram_bins: An integer `Tensor`. How many bins there are in the
source spectrogram data, which is understood to be `fft_size // 2 + 1`,
i.e. the spectrogram only contains the nonredundant FFT bins.
sample_rate: An integer or float `Tensor`. Samples per second of the input
signal used to create the spectrogram. Used to figure out the frequencies
corresponding to each spectrogram bin, which dictates how they are mapped
into the mel scale.
lower_edge_hertz: Python float. Lower bound on the frequencies to be
included in the mel spectrum. This corresponds to the lower edge of the
lowest triangular band.
upper_edge_hertz: Python float. The desired top edge of the highest
frequency band.
dtype: The `DType` of the result matrix. Must be a floating point type.
name: An optional name for the operation.
Returns:
A `Tensor` of shape `[num_spectrogram_bins, num_mel_bins]`.
Raises:
ValueError: If `num_mel_bins`/`num_spectrogram_bins`/`sample_rate` are not
positive, `lower_edge_hertz` is negative, frequency edges are incorrectly
ordered, `upper_edge_hertz` is larger than the Nyquist frequency.
[mel]: https://en.wikipedia.org/wiki/Mel_scale
"""
with ops.name_scope(name, 'linear_to_mel_weight_matrix') as name:
# Convert Tensor `sample_rate` to float, if possible.
if isinstance(sample_rate, tensor.Tensor):
maybe_const_val = tensor_util.constant_value(sample_rate)
if maybe_const_val is not None:
sample_rate = maybe_const_val
# Note: As num_spectrogram_bins is passed to `math_ops.linspace`
# and the validation is already done in linspace (both in shape function
# and in kernel), there is no need to validate num_spectrogram_bins here.
_validate_arguments(num_mel_bins, sample_rate,
lower_edge_hertz, upper_edge_hertz, dtype)
# This function can be constant folded by graph optimization since there are
# no Tensor inputs.
sample_rate = math_ops.cast(
sample_rate, dtype, name='sample_rate')
lower_edge_hertz = ops.convert_to_tensor(
lower_edge_hertz, dtype, name='lower_edge_hertz')
upper_edge_hertz = ops.convert_to_tensor(
upper_edge_hertz, dtype, name='upper_edge_hertz')
zero = ops.convert_to_tensor(0.0, dtype)
# HTK excludes the spectrogram DC bin.
bands_to_zero = 1
nyquist_hertz = sample_rate / 2.0
linear_frequencies = math_ops.linspace(
zero, nyquist_hertz, num_spectrogram_bins)[bands_to_zero:]
spectrogram_bins_mel = array_ops.expand_dims(
_hertz_to_mel(linear_frequencies), 1)
# Compute num_mel_bins triples of (lower_edge, center, upper_edge). The
# center of each band is the lower and upper edge of the adjacent bands.
# Accordingly, we divide [lower_edge_hertz, upper_edge_hertz] into
# num_mel_bins + 2 pieces.
band_edges_mel = shape_ops.frame(
math_ops.linspace(_hertz_to_mel(lower_edge_hertz),
_hertz_to_mel(upper_edge_hertz),
num_mel_bins + 2), frame_length=3, frame_step=1)
# Split the triples up and reshape them into [1, num_mel_bins] tensors.
lower_edge_mel, center_mel, upper_edge_mel = tuple(array_ops.reshape(
t, [1, num_mel_bins]) for t in array_ops.split(
band_edges_mel, 3, axis=1))
# Calculate lower and upper slopes for every spectrogram bin.
# Line segments are linear in the mel domain, not Hertz.
lower_slopes = (spectrogram_bins_mel - lower_edge_mel) / (
center_mel - lower_edge_mel)
upper_slopes = (upper_edge_mel - spectrogram_bins_mel) / (
upper_edge_mel - center_mel)
# Intersect the line segments with each other and zero.
mel_weights_matrix = math_ops.maximum(
zero, math_ops.minimum(lower_slopes, upper_slopes))
# Re-add the zeroed lower bins we sliced out above.
return array_ops.pad(
mel_weights_matrix, [[bands_to_zero, 0], [0, 0]], name=name)
tensorflow/python/ops/signal/mfcc_ops.py
+107
View File
@@ -0,0 +1,107 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Mel-Frequency Cepstral Coefficients (MFCCs) ops."""
from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops.signal import dct_ops
from tensorflow.python.util import dispatch
from tensorflow.python.util.tf_export import tf_export
@tf_export('signal.mfccs_from_log_mel_spectrograms')
@dispatch.add_dispatch_support
def mfccs_from_log_mel_spectrograms(log_mel_spectrograms, name=None):
"""Computes [MFCCs][mfcc] of `log_mel_spectrograms`.
Implemented with GPU-compatible ops and supports gradients.
[Mel-Frequency Cepstral Coefficient (MFCC)][mfcc] calculation consists of
taking the DCT-II of a log-magnitude mel-scale spectrogram. [HTK][htk]'s MFCCs
use a particular scaling of the DCT-II which is almost orthogonal
normalization. We follow this convention.
All `num_mel_bins` MFCCs are returned and it is up to the caller to select
a subset of the MFCCs based on their application. For example, it is typical
to only use the first few for speech recognition, as this results in
an approximately pitch-invariant representation of the signal.
For example:
```python
batch_size, num_samples, sample_rate = 32, 32000, 16000.0
# A Tensor of [batch_size, num_samples] mono PCM samples in the range [-1, 1].
pcm = tf.random.normal([batch_size, num_samples], dtype=tf.float32)
# A 1024-point STFT with frames of 64 ms and 75% overlap.
stfts = tf.signal.stft(pcm, frame_length=1024, frame_step=256,
fft_length=1024)
spectrograms = tf.abs(stfts)
# Warp the linear scale spectrograms into the mel-scale.
num_spectrogram_bins = stfts.shape[-1].value
lower_edge_hertz, upper_edge_hertz, num_mel_bins = 80.0, 7600.0, 80
linear_to_mel_weight_matrix = tf.signal.linear_to_mel_weight_matrix(
num_mel_bins, num_spectrogram_bins, sample_rate, lower_edge_hertz,
upper_edge_hertz)
mel_spectrograms = tf.tensordot(
spectrograms, linear_to_mel_weight_matrix, 1)
mel_spectrograms.set_shape(spectrograms.shape[:-1].concatenate(
linear_to_mel_weight_matrix.shape[-1:]))
# Compute a stabilized log to get log-magnitude mel-scale spectrograms.
log_mel_spectrograms = tf.math.log(mel_spectrograms + 1e-6)
# Compute MFCCs from log_mel_spectrograms and take the first 13.
mfccs = tf.signal.mfccs_from_log_mel_spectrograms(
log_mel_spectrograms)[..., :13]
```
Args:
log_mel_spectrograms: A `[..., num_mel_bins]` `float32`/`float64` `Tensor`
of log-magnitude mel-scale spectrograms.
name: An optional name for the operation.
Returns:
A `[..., num_mel_bins]` `float32`/`float64` `Tensor` of the MFCCs of
`log_mel_spectrograms`.
Raises:
ValueError: If `num_mel_bins` is not positive.
[mfcc]: https://en.wikipedia.org/wiki/Mel-frequency_cepstrum
[htk]: https://en.wikipedia.org/wiki/HTK_(software)
"""
with ops.name_scope(name, 'mfccs_from_log_mel_spectrograms',
[log_mel_spectrograms]):
# Compute the DCT-II of the resulting log-magnitude mel-scale spectrogram.
# The DCT used in HTK scales every basis vector by sqrt(2/N), which is the
# scaling required for an "orthogonal" DCT-II *except* in the 0th bin, where
# the true orthogonal DCT (as implemented by scipy) scales by sqrt(1/N). For
# this reason, we don't apply orthogonal normalization and scale the DCT by
# `0.5 * sqrt(2/N)` manually.
log_mel_spectrograms = ops.convert_to_tensor(log_mel_spectrograms)
if (log_mel_spectrograms.shape.ndims and
log_mel_spectrograms.shape.dims[-1].value is not None):
num_mel_bins = log_mel_spectrograms.shape.dims[-1].value
if num_mel_bins == 0:
raise ValueError('num_mel_bins must be positive. Got: %s' %
log_mel_spectrograms)
else:
num_mel_bins = array_ops.shape(log_mel_spectrograms)[-1]
dct2 = dct_ops.dct(log_mel_spectrograms, type=2)
return dct2 * math_ops.rsqrt(
math_ops.cast(num_mel_bins, dct2.dtype) * 2.0)
tensorflow/python/ops/signal/reconstruction_ops.py
+163
View File
@@ -0,0 +1,163 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Signal reconstruction via overlapped addition of frames."""
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_util
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.util import dispatch
from tensorflow.python.util.tf_export import tf_export
@tf_export("signal.overlap_and_add")
@dispatch.add_dispatch_support
def overlap_and_add(signal, frame_step, name=None):
"""Reconstructs a signal from a framed representation.
Adds potentially overlapping frames of a signal with shape
`[..., frames, frame_length]`, offsetting subsequent frames by `frame_step`.
The resulting tensor has shape `[..., output_size]` where
output_size = (frames - 1) * frame_step + frame_length
Args:
signal: A [..., frames, frame_length] `Tensor`. All dimensions may be
unknown, and rank must be at least 2.
frame_step: An integer or scalar `Tensor` denoting overlap offsets. Must be
less than or equal to `frame_length`.
name: An optional name for the operation.
Returns:
A `Tensor` with shape `[..., output_size]` containing the overlap-added
frames of `signal`'s inner-most two dimensions.
Raises:
ValueError: If `signal`'s rank is less than 2, or `frame_step` is not a
scalar integer.
"""
with ops.name_scope(name, "overlap_and_add", [signal, frame_step]):
signal = ops.convert_to_tensor(signal, name="signal")
signal.shape.with_rank_at_least(2)
frame_step = ops.convert_to_tensor(frame_step, name="frame_step")
frame_step.shape.assert_has_rank(0)
if not frame_step.dtype.is_integer:
raise ValueError("frame_step must be an integer. Got %s" %
frame_step.dtype)
frame_step_static = tensor_util.constant_value(frame_step)
frame_step_is_static = frame_step_static is not None
frame_step = frame_step_static if frame_step_is_static else frame_step
signal_shape = array_ops.shape(signal)
signal_shape_static = tensor_util.constant_value(signal_shape)
if signal_shape_static is not None:
signal_shape = signal_shape_static
# All dimensions that are not part of the overlap-and-add. Can be empty for
# rank 2 inputs.
outer_dimensions = signal_shape[:-2]
outer_rank = array_ops.size(outer_dimensions)
outer_rank_static = tensor_util.constant_value(outer_rank)
if outer_rank_static is not None:
outer_rank = outer_rank_static
def full_shape(inner_shape):
return array_ops.concat([outer_dimensions, inner_shape], 0)
frame_length = signal_shape[-1]
frames = signal_shape[-2]
# Compute output length.
output_length = frame_length + frame_step * (frames - 1)
# If frame_length is equal to frame_step, there's no overlap so just
# reshape the tensor.
if (frame_step_is_static and signal.shape.dims is not None and
frame_step == signal.shape.dims[-1].value):
output_shape = full_shape([output_length])
return array_ops.reshape(signal, output_shape, name="fast_path")
# The following code is documented using this example:
#
# frame_step = 2
# signal.shape = (3, 5)
# a b c d e
# f g h i j
# k l m n o
# Compute the number of segments, per frame.
segments = -(-frame_length // frame_step) # Divide and round up.
# Pad the frame_length dimension to a multiple of the frame step.
# Pad the frames dimension by `segments` so that signal.shape = (6, 6)
# a b c d e 0
# f g h i j 0
# k l m n o 0
# 0 0 0 0 0 0
# 0 0 0 0 0 0
# 0 0 0 0 0 0
paddings = [[0, segments], [0, segments * frame_step - frame_length]]
outer_paddings = array_ops.zeros([outer_rank, 2], dtypes.int32)
paddings = array_ops.concat([outer_paddings, paddings], 0)
signal = array_ops.pad(signal, paddings)
# Reshape so that signal.shape = (3, 6, 2)
# ab cd e0
# fg hi j0
# kl mn o0
# 00 00 00
# 00 00 00
# 00 00 00
shape = full_shape([frames + segments, segments, frame_step])
signal = array_ops.reshape(signal, shape)
# Transpose dimensions so that signal.shape = (3, 6, 2)
# ab fg kl 00 00 00
# cd hi mn 00 00 00
# e0 j0 o0 00 00 00
perm = array_ops.concat(
[math_ops.range(outer_rank), outer_rank + [1, 0, 2]], 0)
perm_static = tensor_util.constant_value(perm)
perm = perm_static if perm_static is not None else perm
signal = array_ops.transpose(signal, perm)
# Reshape so that signal.shape = (18, 2)
# ab fg kl 00 00 00 cd hi mn 00 00 00 e0 j0 o0 00 00 00
shape = full_shape([(frames + segments) * segments, frame_step])
signal = array_ops.reshape(signal, shape)
# Truncate so that signal.shape = (15, 2)
# ab fg kl 00 00 00 cd hi mn 00 00 00 e0 j0 o0
signal = signal[..., :(frames + segments - 1) * segments, :]
# Reshape so that signal.shape = (3, 5, 2)
# ab fg kl 00 00
# 00 cd hi mn 00
# 00 00 e0 j0 o0
shape = full_shape([segments, (frames + segments - 1), frame_step])
signal = array_ops.reshape(signal, shape)
# Now, reduce over the columns, to achieve the desired sum.
signal = math_ops.reduce_sum(signal, -3)
# Flatten the array.
shape = full_shape([(frames + segments - 1) * frame_step])
signal = array_ops.reshape(signal, shape)
# Truncate to final length.
signal = signal[..., :output_length]
return signal
tensorflow/python/ops/signal/shape_ops.py
+231
View File
@@ -0,0 +1,231 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""General shape ops for frames."""
import numpy as np
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_util
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops.signal import util_ops
from tensorflow.python.util import dispatch
from tensorflow.python.util.tf_export import tf_export
def _infer_frame_shape(signal, frame_length, frame_step, pad_end, axis):
"""Infers the shape of the return value of `frame`."""
frame_length = tensor_util.constant_value(frame_length)
frame_step = tensor_util.constant_value(frame_step)
axis = tensor_util.constant_value(axis)
if signal.shape.ndims is None:
return None
if axis is None:
return [None] * (signal.shape.ndims + 1)
signal_shape = signal.shape.as_list()
num_frames = None
frame_axis = signal_shape[axis]
outer_dimensions = signal_shape[:axis]
inner_dimensions = signal_shape[axis:][1:]
if signal_shape and frame_axis is not None:
if frame_step is not None and pad_end:
# Double negative is so that we round up.
num_frames = max(0, -(-frame_axis // frame_step))
elif frame_step is not None and frame_length is not None:
assert not pad_end
num_frames = max(
0, (frame_axis - frame_length + frame_step) // frame_step)
return outer_dimensions + [num_frames, frame_length] + inner_dimensions
@tf_export("signal.frame")
@dispatch.add_dispatch_support
def frame(signal, frame_length, frame_step, pad_end=False, pad_value=0, axis=-1,
name=None):
"""Expands `signal`'s `axis` dimension into frames of `frame_length`.
Slides a window of size `frame_length` over `signal`'s `axis` dimension
with a stride of `frame_step`, replacing the `axis` dimension with
`[frames, frame_length]` frames.
If `pad_end` is True, window positions that are past the end of the `axis`
dimension are padded with `pad_value` until the window moves fully past the
end of the dimension. Otherwise, only window positions that fully overlap the
`axis` dimension are produced.
For example:
>>> # A batch size 3 tensor of 9152 audio samples.
>>> audio = tf.random.normal([3, 9152])
>>>
>>> # Compute overlapping frames of length 512 with a step of 180 (frames overlap
>>> # by 332 samples). By default, only 49 frames are generated since a frame
>>> # with start position j*180 for j > 48 would overhang the end.
>>> frames = tf.signal.frame(audio, 512, 180)
>>> frames.shape.assert_is_compatible_with([3, 49, 512])
>>>
>>> # When pad_end is enabled, the final two frames are kept (padded with zeros).
>>> frames = tf.signal.frame(audio, 512, 180, pad_end=True)
>>> frames.shape.assert_is_compatible_with([3, 51, 512])
If the dimension along `axis` is N, and `pad_end=False`, the number of frames
can be computed by:
```python
num_frames = 1 + (N - frame_size) // frame_step
```
If `pad_end=True`, the number of frames can be computed by:
```python
num_frames = -(-N // frame_step) # ceiling division
```
Args:
signal: A `[..., samples, ...]` `Tensor`. The rank and dimensions
may be unknown. Rank must be at least 1.
frame_length: The frame length in samples. An integer or scalar `Tensor`.
frame_step: The frame hop size in samples. An integer or scalar `Tensor`.
pad_end: Whether to pad the end of `signal` with `pad_value`.
pad_value: An optional scalar `Tensor` to use where the input signal
does not exist when `pad_end` is True.
axis: A scalar integer `Tensor` indicating the axis to frame. Defaults to
the last axis. Supports negative values for indexing from the end.
name: An optional name for the operation.
Returns:
A `Tensor` of frames with shape `[..., num_frames, frame_length, ...]`.
Raises:
ValueError: If `frame_length`, `frame_step`, `pad_value`, or `axis` are not
scalar.
"""
with ops.name_scope(name, "frame", [signal, frame_length, frame_step,
pad_value]):
signal = ops.convert_to_tensor(signal, name="signal")
frame_length = ops.convert_to_tensor(frame_length, name="frame_length")
frame_step = ops.convert_to_tensor(frame_step, name="frame_step")
axis = ops.convert_to_tensor(axis, name="axis")
signal.shape.with_rank_at_least(1)
frame_length.shape.assert_has_rank(0)
frame_step.shape.assert_has_rank(0)
axis.shape.assert_has_rank(0)
result_shape = _infer_frame_shape(signal, frame_length, frame_step, pad_end,
axis)
def maybe_constant(val):
val_static = tensor_util.constant_value(val)
return (val_static, True) if val_static is not None else (val, False)
signal_shape, signal_shape_is_static = maybe_constant(
array_ops.shape(signal))
axis, axis_is_static = maybe_constant(axis)
if signal_shape_is_static and axis_is_static:
# Axis can be negative. Convert it to positive.
axis = range(len(signal_shape))[axis]
outer_dimensions, length_samples, inner_dimensions = np.split(
signal_shape, indices_or_sections=[axis, axis + 1])
length_samples = length_samples.item()
else:
signal_rank = array_ops.rank(signal)
# Axis can be negative. Convert it to positive.
axis = math_ops.range(signal_rank)[axis]
outer_dimensions, length_samples, inner_dimensions = array_ops.split(
signal_shape, [axis, 1, signal_rank - 1 - axis])
length_samples = array_ops.reshape(length_samples, [])
num_outer_dimensions = array_ops.size(outer_dimensions)
num_inner_dimensions = array_ops.size(inner_dimensions)
# If padding is requested, pad the input signal tensor with pad_value.
if pad_end:
pad_value = ops.convert_to_tensor(pad_value, signal.dtype)
pad_value.shape.assert_has_rank(0)
# Calculate number of frames, using double negatives to round up.
num_frames = -(-length_samples // frame_step)
# Pad the signal by up to frame_length samples based on how many samples
# are remaining starting from last_frame_position.
pad_samples = math_ops.maximum(
0, frame_length + frame_step * (num_frames - 1) - length_samples)
# Pad the inner dimension of signal by pad_samples.
paddings = array_ops.concat([
array_ops.zeros([num_outer_dimensions, 2], dtype=pad_samples.dtype),
ops.convert_to_tensor([[0, pad_samples]]),
array_ops.zeros([num_inner_dimensions, 2], dtype=pad_samples.dtype)
], 0)
signal = array_ops.pad(signal, paddings, constant_values=pad_value)
signal_shape = array_ops.shape(signal)
length_samples = signal_shape[axis]
else:
num_frames = math_ops.maximum(
constant_op.constant(0, dtype=frame_length.dtype),
1 + (length_samples - frame_length) // frame_step)
subframe_length, _ = maybe_constant(util_ops.gcd(frame_length, frame_step))
subframes_per_frame = frame_length // subframe_length
subframes_per_hop = frame_step // subframe_length
num_subframes = length_samples // subframe_length
slice_shape = array_ops.concat([outer_dimensions,
[num_subframes * subframe_length],
inner_dimensions], 0)
subframe_shape = array_ops.concat([outer_dimensions,
[num_subframes, subframe_length],
inner_dimensions], 0)
subframes = array_ops.reshape(array_ops.strided_slice(
signal, array_ops.zeros_like(signal_shape),
slice_shape), subframe_shape)
# frame_selector is a [num_frames, subframes_per_frame] tensor
# that indexes into the appropriate frame in subframes. For example:
# [[0, 0, 0, 0], [2, 2, 2, 2], [4, 4, 4, 4]]
frame_selector = array_ops.reshape(
math_ops.range(num_frames, dtype=frame_length.dtype) *
subframes_per_hop, [num_frames, 1])
# subframe_selector is a [num_frames, subframes_per_frame] tensor
# that indexes into the appropriate subframe within a frame. For example:
# [[0, 1, 2, 3], [0, 1, 2, 3], [0, 1, 2, 3]]
subframe_selector = array_ops.reshape(
math_ops.range(subframes_per_frame, dtype=frame_length.dtype),
[1, subframes_per_frame])
# Adding the 2 selector tensors together produces a [num_frames,
# subframes_per_frame] tensor of indices to use with tf.gather to select
# subframes from subframes. We then reshape the inner-most
# subframes_per_frame dimension to stitch the subframes together into
# frames. For example: [[0, 1, 2, 3], [2, 3, 4, 5], [4, 5, 6, 7]].
selector = frame_selector + subframe_selector
# Dtypes have to match.
outer_dimensions = ops.convert_to_tensor(outer_dimensions)
inner_dimensions = ops.convert_to_tensor(
inner_dimensions, dtype=outer_dimensions.dtype)
mid_dimensions = ops.convert_to_tensor([num_frames, frame_length],
dtype=outer_dimensions.dtype)
frames = array_ops.reshape(
array_ops.gather(subframes, selector, axis=axis),
array_ops.concat([outer_dimensions, mid_dimensions, inner_dimensions],
0))
if result_shape:
frames.set_shape(result_shape)
return frames
tensorflow/python/ops/signal/signal.py
+44
View File
@@ -0,0 +1,44 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Signal processing operations."""
# pylint: disable=unused-import
from tensorflow.python.ops import signal
from tensorflow.python.ops.signal.dct_ops import dct
from tensorflow.python.ops.signal.fft_ops import fft
from tensorflow.python.ops.signal.fft_ops import fft2d
from tensorflow.python.ops.signal.fft_ops import fft3d
from tensorflow.python.ops.signal.fft_ops import fftshift
from tensorflow.python.ops.signal.fft_ops import rfft
from tensorflow.python.ops.signal.fft_ops import rfft2d
from tensorflow.python.ops.signal.fft_ops import rfft3d
from tensorflow.python.ops.signal.dct_ops import idct
from tensorflow.python.ops.signal.fft_ops import ifft
from tensorflow.python.ops.signal.fft_ops import ifft2d
from tensorflow.python.ops.signal.fft_ops import ifft3d
from tensorflow.python.ops.signal.fft_ops import ifftshift
from tensorflow.python.ops.signal.fft_ops import irfft
from tensorflow.python.ops.signal.fft_ops import irfft2d
from tensorflow.python.ops.signal.fft_ops import irfft3d
from tensorflow.python.ops.signal.mel_ops import linear_to_mel_weight_matrix
from tensorflow.python.ops.signal.mfcc_ops import mfccs_from_log_mel_spectrograms
from tensorflow.python.ops.signal.reconstruction_ops import overlap_and_add
from tensorflow.python.ops.signal.shape_ops import frame
from tensorflow.python.ops.signal.spectral_ops import inverse_stft
from tensorflow.python.ops.signal.spectral_ops import inverse_stft_window_fn
from tensorflow.python.ops.signal.spectral_ops import stft
from tensorflow.python.ops.signal.window_ops import hamming_window
from tensorflow.python.ops.signal.window_ops import hann_window
# pylint: enable=unused-import
tensorflow/python/ops/signal/spectral_ops.py
+488
View File
@@ -0,0 +1,488 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Spectral operations (e.g. Short-time Fourier Transform)."""
import numpy as np
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_util
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops.signal import dct_ops
from tensorflow.python.ops.signal import fft_ops
from tensorflow.python.ops.signal import reconstruction_ops
from tensorflow.python.ops.signal import shape_ops
from tensorflow.python.ops.signal import window_ops
from tensorflow.python.util import dispatch
from tensorflow.python.util.tf_export import tf_export
@tf_export('signal.stft')
@dispatch.add_dispatch_support
def stft(signals, frame_length, frame_step, fft_length=None,
window_fn=window_ops.hann_window,
pad_end=False, name=None):
"""Computes the [Short-time Fourier Transform][stft] of `signals`.
Implemented with TPU/GPU-compatible ops and supports gradients.
Args:
signals: A `[..., samples]` `float32`/`float64` `Tensor` of real-valued
signals.
frame_length: An integer scalar `Tensor`. The window length in samples.
frame_step: An integer scalar `Tensor`. The number of samples to step.
fft_length: An integer scalar `Tensor`. The size of the FFT to apply.
If not provided, uses the smallest power of 2 enclosing `frame_length`.
window_fn: A callable that takes a window length and a `dtype` keyword
argument and returns a `[window_length]` `Tensor` of samples in the
provided datatype. If set to `None`, no windowing is used.
pad_end: Whether to pad the end of `signals` with zeros when the provided
frame length and step produces a frame that lies partially past its end.
name: An optional name for the operation.
Returns:
A `[..., frames, fft_unique_bins]` `Tensor` of `complex64`/`complex128`
STFT values where `fft_unique_bins` is `fft_length // 2 + 1` (the unique
components of the FFT).
Raises:
ValueError: If `signals` is not at least rank 1, `frame_length` is
not scalar, or `frame_step` is not scalar.
[stft]: https://en.wikipedia.org/wiki/Short-time_Fourier_transform
"""
with ops.name_scope(name, 'stft', [signals, frame_length,
frame_step]):
signals = ops.convert_to_tensor(signals, name='signals')
signals.shape.with_rank_at_least(1)
frame_length = ops.convert_to_tensor(frame_length, name='frame_length')
frame_length.shape.assert_has_rank(0)
frame_step = ops.convert_to_tensor(frame_step, name='frame_step')
frame_step.shape.assert_has_rank(0)
# Validate frame_length / fft_length statically when possible. The
# FFT backend (DUCC) aborts the process on a zero-length transform
# rather than returning an error tensor, so we have to gate the
# invalid configuration here. See #117843 for the matching crash in
# `inverse_stft`.
_validate_positive_static(frame_length, 'frame_length')
if fft_length is None:
fft_length = _enclosing_power_of_two(frame_length)
else:
fft_length = ops.convert_to_tensor(fft_length, name='fft_length')
_validate_positive_static(fft_length, 'fft_length')
framed_signals = shape_ops.frame(
signals, frame_length, frame_step, pad_end=pad_end)
# Optionally window the framed signals.
if window_fn is not None:
window = window_fn(frame_length, dtype=framed_signals.dtype)
framed_signals *= window
# fft_ops.rfft produces the (fft_length/2 + 1) unique components of the
# FFT of the real windowed signals in framed_signals.
return fft_ops.rfft(framed_signals, [fft_length])
@tf_export('signal.inverse_stft_window_fn')
@dispatch.add_dispatch_support
def inverse_stft_window_fn(frame_step,
forward_window_fn=window_ops.hann_window,
name=None):
"""Generates a window function that can be used in `inverse_stft`.
Constructs a window that is equal to the forward window with a further
pointwise amplitude correction. `inverse_stft_window_fn` is equivalent to
`forward_window_fn` in the case where it would produce an exact inverse.
See examples in `inverse_stft` documentation for usage.
Args:
frame_step: An integer scalar `Tensor`. The number of samples to step.
forward_window_fn: window_fn used in the forward transform, `stft`.
name: An optional name for the operation.
Returns:
A callable that takes a window length and a `dtype` keyword argument and
returns a `[window_length]` `Tensor` of samples in the provided datatype.
The returned window is suitable for reconstructing original waveform in
inverse_stft.
"""
def inverse_stft_window_fn_inner(frame_length, dtype):
"""Computes a window that can be used in `inverse_stft`.
Args:
frame_length: An integer scalar `Tensor`. The window length in samples.
dtype: Data type of waveform passed to `stft`.
Returns:
A window suitable for reconstructing original waveform in `inverse_stft`.
Raises:
ValueError: If `frame_length` is not scalar, `forward_window_fn` is not a
callable that takes a window length and a `dtype` keyword argument and
returns a `[window_length]` `Tensor` of samples in the provided datatype
`frame_step` is not scalar, or `frame_step` is not scalar.
"""
with ops.name_scope(name, 'inverse_stft_window_fn', [forward_window_fn]):
frame_step_ = ops.convert_to_tensor(frame_step, name='frame_step')
frame_step_.shape.assert_has_rank(0)
frame_length = ops.convert_to_tensor(frame_length, name='frame_length')
frame_length.shape.assert_has_rank(0)
# Use equation 7 from Griffin + Lim.
forward_window = forward_window_fn(frame_length, dtype=dtype)
denom = math_ops.square(forward_window)
overlaps = -(-frame_length // frame_step_) # Ceiling division. # pylint: disable=invalid-unary-operand-type
denom = array_ops.pad(denom, [(0, overlaps * frame_step_ - frame_length)])
denom = array_ops.reshape(denom, [overlaps, frame_step_])
denom = math_ops.reduce_sum(denom, 0, keepdims=True)
denom = array_ops.tile(denom, [overlaps, 1])
denom = array_ops.reshape(denom, [overlaps * frame_step_])
denom = denom[:frame_length]
return array_ops.where(
math_ops.equal(denom, 0.0),
array_ops.zeros_like(forward_window),
forward_window / denom,
)
return inverse_stft_window_fn_inner
@tf_export('signal.inverse_stft')
@dispatch.add_dispatch_support
def inverse_stft(stfts,
frame_length,
frame_step,
fft_length=None,
window_fn=window_ops.hann_window,
name=None):
"""Computes the inverse [Short-time Fourier Transform][stft] of `stfts`.
To reconstruct an original waveform, a complementary window function should
be used with `inverse_stft`. Such a window function can be constructed with
`tf.signal.inverse_stft_window_fn`.
Example:
```python
frame_length = 400
frame_step = 160
waveform = tf.random.normal(dtype=tf.float32, shape=[1000])
stft = tf.signal.stft(waveform, frame_length, frame_step)
inverse_stft = tf.signal.inverse_stft(
stft, frame_length, frame_step,
window_fn=tf.signal.inverse_stft_window_fn(frame_step))
```
If a custom `window_fn` is used with `tf.signal.stft`, it must be passed to
`tf.signal.inverse_stft_window_fn`:
```python
frame_length = 400
frame_step = 160
window_fn = tf.signal.hamming_window
waveform = tf.random.normal(dtype=tf.float32, shape=[1000])
stft = tf.signal.stft(
waveform, frame_length, frame_step, window_fn=window_fn)
inverse_stft = tf.signal.inverse_stft(
stft, frame_length, frame_step,
window_fn=tf.signal.inverse_stft_window_fn(
frame_step, forward_window_fn=window_fn))
```
Implemented with TPU/GPU-compatible ops and supports gradients.
Args:
stfts: A `complex64`/`complex128` `[..., frames, fft_unique_bins]`
`Tensor` of STFT bins representing a batch of `fft_length`-point STFTs
where `fft_unique_bins` is `fft_length // 2 + 1`
frame_length: An integer scalar `Tensor`. The window length in samples.
frame_step: An integer scalar `Tensor`. The number of samples to step.
fft_length: An integer scalar `Tensor`. The size of the FFT that produced
`stfts`. If not provided, uses the smallest power of 2 enclosing
`frame_length`.
window_fn: A callable that takes a window length and a `dtype` keyword
argument and returns a `[window_length]` `Tensor` of samples in the
provided datatype. If set to `None`, no windowing is used.
name: An optional name for the operation.
Returns:
A `[..., samples]` `Tensor` of `float32`/`float64` signals representing
the inverse STFT for each input STFT in `stfts`.
Raises:
ValueError: If `stfts` is not at least rank 2, `frame_length` is not scalar,
`frame_step` is not scalar, or `fft_length` is not scalar.
[stft]: https://en.wikipedia.org/wiki/Short-time_Fourier_transform
"""
with ops.name_scope(name, 'inverse_stft', [stfts]):
stfts = ops.convert_to_tensor(stfts, name='stfts')
stfts.shape.with_rank_at_least(2)
frame_length = ops.convert_to_tensor(frame_length, name='frame_length')
frame_length.shape.assert_has_rank(0)
frame_step = ops.convert_to_tensor(frame_step, name='frame_step')
frame_step.shape.assert_has_rank(0)
# Same rationale as in `stft` above: the FFT backend aborts the
# process on a zero-length transform, so we validate the static
# value here rather than letting it reach the kernel. Reported in
# #117843, where `frame_length=0` (or `stfts.shape[-1] == 1`) caused
# the inferred fft_length to be 0 and DUCC to abort.
_validate_positive_static(frame_length, 'frame_length')
if fft_length is None:
fft_length = _enclosing_power_of_two(frame_length)
else:
fft_length = ops.convert_to_tensor(fft_length, name='fft_length')
fft_length.shape.assert_has_rank(0)
_validate_positive_static(fft_length, 'fft_length')
real_frames = fft_ops.irfft(stfts, [fft_length])
# frame_length may be larger or smaller than fft_length, so we pad or
# truncate real_frames to frame_length.
frame_length_static = tensor_util.constant_value(frame_length)
# If we don't know the shape of real_frames's inner dimension, pad and
# truncate to frame_length.
if (frame_length_static is None or real_frames.shape.ndims is None or
real_frames.shape.as_list()[-1] is None):
real_frames = real_frames[..., :frame_length]
real_frames_rank = array_ops.rank(real_frames)
real_frames_shape = array_ops.shape(real_frames)
paddings = array_ops.concat(
[array_ops.zeros([real_frames_rank - 1, 2],
dtype=frame_length.dtype),
[[0, math_ops.maximum(0, frame_length - real_frames_shape[-1])]]], 0)
real_frames = array_ops.pad(real_frames, paddings)
# We know real_frames's last dimension and frame_length statically. If they
# are different, then pad or truncate real_frames to frame_length.
elif real_frames.shape.as_list()[-1] > frame_length_static:
real_frames = real_frames[..., :frame_length_static]
elif real_frames.shape.as_list()[-1] < frame_length_static:
pad_amount = frame_length_static - real_frames.shape.as_list()[-1]
real_frames = array_ops.pad(real_frames,
[[0, 0]] * (real_frames.shape.ndims - 1) +
[[0, pad_amount]])
# The above code pads the inner dimension of real_frames to frame_length,
# but it does so in a way that may not be shape-inference friendly.
# Restore shape information if we are able to.
if frame_length_static is not None and real_frames.shape.ndims is not None:
real_frames.set_shape([None] * (real_frames.shape.ndims - 1) +
[frame_length_static])
# Optionally window and overlap-add the inner 2 dimensions of real_frames
# into a single [samples] dimension.
if window_fn is not None:
window = window_fn(frame_length, dtype=stfts.dtype.real_dtype)
real_frames *= window
return reconstruction_ops.overlap_and_add(real_frames, frame_step)
def _enclosing_power_of_two(value):
"""Return 2**N for integer N such that 2**N >= value."""
value_static = tensor_util.constant_value(value)
if value_static is not None:
return constant_op.constant(
int(2**np.ceil(np.log(value_static) / np.log(2.0))), value.dtype)
return math_ops.cast(
math_ops.pow(
2.0,
math_ops.ceil(
math_ops.log(math_ops.cast(value, dtypes.float32)) /
math_ops.log(2.0))), value.dtype)
def _validate_positive_static(value, arg_name):
"""Raise ValueError if `value` is statically known to be non-positive.
Used to keep zero-length FFT configurations from reaching the FFT
backend (DUCC), which aborts the process on `n == 0`. See #117843.
When the value is dynamic (not constant-foldable), we skip the check
and let the kernel surface a runtime error: this matches the rest of
TF's policy for shape arguments and avoids over-eager failures in
`tf.function` traces where the value is symbolic.
"""
static = tensor_util.constant_value(value)
if static is not None and int(static) <= 0:
raise ValueError(
f'`{arg_name}` must be a positive integer, got {int(static)}. '
'A zero or negative FFT length would crash the underlying '
'FFT backend.'
)
@tf_export('signal.mdct')
@dispatch.add_dispatch_support
def mdct(signals, frame_length, window_fn=window_ops.vorbis_window,
pad_end=False, norm=None, name=None):
"""Computes the [Modified Discrete Cosine Transform][mdct] of `signals`.
Implemented with TPU/GPU-compatible ops and supports gradients.
Args:
signals: A `[..., samples]` `float32`/`float64` `Tensor` of real-valued
signals.
frame_length: An integer scalar `Tensor`. The window length in samples
which must be divisible by 4.
window_fn: A callable that takes a frame_length and a `dtype` keyword
argument and returns a `[frame_length]` `Tensor` of samples in the
provided datatype. If set to `None`, a rectangular window with a scale of
1/sqrt(2) is used. For perfect reconstruction of a signal from `mdct`
followed by `inverse_mdct`, please use `tf.signal.vorbis_window`,
`tf.signal.kaiser_bessel_derived_window` or `None`. If using another
window function, make sure that w[n]^2 + w[n + frame_length // 2]^2 = 1
and w[n] = w[frame_length - n - 1] for n = 0,...,frame_length // 2 - 1 to
achieve perfect reconstruction.
pad_end: Whether to pad the end of `signals` with zeros when the provided
frame length and step produces a frame that lies partially past its end.
norm: If it is None, unnormalized dct4 is used, if it is "ortho"
orthonormal dct4 is used.
name: An optional name for the operation.
Returns:
A `[..., frames, frame_length // 2]` `Tensor` of `float32`/`float64`
MDCT values where `frames` is roughly `samples // (frame_length // 2)`
when `pad_end=False`.
Raises:
ValueError: If `signals` is not at least rank 1, `frame_length` is
not scalar, or `frame_length` is not a multiple of `4`.
[mdct]: https://en.wikipedia.org/wiki/Modified_discrete_cosine_transform
"""
with ops.name_scope(name, 'mdct', [signals, frame_length]):
signals = ops.convert_to_tensor(signals, name='signals')
signals.shape.with_rank_at_least(1)
frame_length = ops.convert_to_tensor(frame_length, name='frame_length')
frame_length.shape.assert_has_rank(0)
# Assert that frame_length is divisible by 4.
frame_length_static = tensor_util.constant_value(frame_length)
if frame_length_static is not None:
if frame_length_static % 4 != 0:
raise ValueError('The frame length must be a multiple of 4.')
frame_step = ops.convert_to_tensor(frame_length_static // 2,
dtype=frame_length.dtype)
else:
frame_step = frame_length // 2
framed_signals = shape_ops.frame(
signals, frame_length, frame_step, pad_end=pad_end)
# Optionally window the framed signals.
if window_fn is not None:
window = window_fn(frame_length, dtype=framed_signals.dtype)
framed_signals *= window
else:
framed_signals *= 1.0 / np.sqrt(2)
split_frames = array_ops.split(framed_signals, 4, axis=-1)
frame_firsthalf = -array_ops.reverse(split_frames[2],
[-1]) - split_frames[3]
frame_secondhalf = split_frames[0] - array_ops.reverse(split_frames[1],
[-1])
frames_rearranged = array_ops.concat((frame_firsthalf, frame_secondhalf),
axis=-1)
# Below call produces the (frame_length // 2) unique components of the
# type 4 orthonormal DCT of the real windowed signals in frames_rearranged.
return dct_ops.dct(frames_rearranged, type=4, norm=norm)
@tf_export('signal.inverse_mdct')
@dispatch.add_dispatch_support
def inverse_mdct(mdcts,
window_fn=window_ops.vorbis_window,
norm=None,
name=None):
"""Computes the inverse modified DCT of `mdcts`.
To reconstruct an original waveform, the same window function should
be used with `mdct` and `inverse_mdct`.
Example usage:
>>> @tf.function
... def compare_round_trip():
... samples = 1000
... frame_length = 400
... halflen = frame_length // 2
... waveform = tf.random.normal(dtype=tf.float32, shape=[samples])
... waveform_pad = tf.pad(waveform, [[halflen, 0],])
... mdct = tf.signal.mdct(waveform_pad, frame_length, pad_end=True,
... window_fn=tf.signal.vorbis_window)
... inverse_mdct = tf.signal.inverse_mdct(mdct,
... window_fn=tf.signal.vorbis_window)
... inverse_mdct = inverse_mdct[halflen: halflen + samples]
... return waveform, inverse_mdct
>>> waveform, inverse_mdct = compare_round_trip()
>>> np.allclose(waveform.numpy(), inverse_mdct.numpy(), rtol=1e-3, atol=1e-4)
True
Implemented with TPU/GPU-compatible ops and supports gradients.
Args:
mdcts: A `float32`/`float64` `[..., frames, frame_length // 2]`
`Tensor` of MDCT bins representing a batch of `frame_length // 2`-point
MDCTs.
window_fn: A callable that takes a frame_length and a `dtype` keyword
argument and returns a `[frame_length]` `Tensor` of samples in the
provided datatype. If set to `None`, a rectangular window with a scale of
1/sqrt(2) is used. For perfect reconstruction of a signal from `mdct`
followed by `inverse_mdct`, please use `tf.signal.vorbis_window`,
`tf.signal.kaiser_bessel_derived_window` or `None`. If using another
window function, make sure that w[n]^2 + w[n + frame_length // 2]^2 = 1
and w[n] = w[frame_length - n - 1] for n = 0,...,frame_length // 2 - 1 to
achieve perfect reconstruction.
norm: If "ortho", orthonormal inverse DCT4 is performed, if it is None,
a regular dct4 followed by scaling of `1/frame_length` is performed.
name: An optional name for the operation.
Returns:
A `[..., samples]` `Tensor` of `float32`/`float64` signals representing
the inverse MDCT for each input MDCT in `mdcts` where `samples` is
`(frames - 1) * (frame_length // 2) + frame_length`.
Raises:
ValueError: If `mdcts` is not at least rank 2.
[mdct]: https://en.wikipedia.org/wiki/Modified_discrete_cosine_transform
"""
with ops.name_scope(name, 'inverse_mdct', [mdcts]):
mdcts = ops.convert_to_tensor(mdcts, name='mdcts')
mdcts.shape.with_rank_at_least(2)
half_len = math_ops.cast(mdcts.shape[-1], dtype=dtypes.int32)
if norm is None:
half_len_float = math_ops.cast(half_len, dtype=mdcts.dtype)
result_idct4 = (0.5 / half_len_float) * dct_ops.dct(mdcts, type=4)
elif norm == 'ortho':
result_idct4 = dct_ops.dct(mdcts, type=4, norm='ortho')
split_result = array_ops.split(result_idct4, 2, axis=-1)
real_frames = array_ops.concat((split_result[1],
-array_ops.reverse(split_result[1], [-1]),
-array_ops.reverse(split_result[0], [-1]),
-split_result[0]), axis=-1)
# Optionally window and overlap-add the inner 2 dimensions of real_frames
# into a single [samples] dimension.
if window_fn is not None:
window = window_fn(2 * half_len, dtype=mdcts.dtype)
real_frames *= window
else:
real_frames *= 1.0 / np.sqrt(2)
return reconstruction_ops.overlap_and_add(real_frames, half_len)
tensorflow/python/ops/signal/util_ops.py
+69
View File
@@ -0,0 +1,69 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Utility ops shared across tf.contrib.signal."""
import fractions # gcd is here for Python versions < 3
import math # Get gcd here for Python versions >= 3
import sys
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_util
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import while_loop
def gcd(a, b, name=None):
"""Returns the greatest common divisor via Euclid's algorithm.
Args:
a: The dividend. A scalar integer `Tensor`.
b: The divisor. A scalar integer `Tensor`.
name: An optional name for the operation.
Returns:
A scalar `Tensor` representing the greatest common divisor between `a` and
`b`.
Raises:
ValueError: If `a` or `b` are not scalar integers.
"""
with ops.name_scope(name, 'gcd', [a, b]):
a = ops.convert_to_tensor(a)
b = ops.convert_to_tensor(b)
a.shape.assert_has_rank(0)
b.shape.assert_has_rank(0)
if not a.dtype.is_integer:
raise ValueError('a must be an integer type. Got: %s' % a.dtype)
if not b.dtype.is_integer:
raise ValueError('b must be an integer type. Got: %s' % b.dtype)
# TPU requires static shape inference. GCD is used for subframe size
# computation, so we should prefer static computation where possible.
const_a = tensor_util.constant_value(a)
const_b = tensor_util.constant_value(b)
if const_a is not None and const_b is not None:
if sys.version_info.major < 3:
math_gcd = fractions.gcd
else:
math_gcd = math.gcd
return ops.convert_to_tensor(math_gcd(const_a, const_b))
cond = lambda _, b: math_ops.greater(b, array_ops.zeros_like(b))
body = lambda a, b: [b, math_ops.mod(a, b)]
a, b = while_loop.while_loop(cond, body, [a, b], back_prop=False)
return a
tensorflow/python/ops/signal/window_ops.py
+248
View File
@@ -0,0 +1,248 @@
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Ops for computing common window functions."""
import numpy as np
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_util
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import cond
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import nn_ops
from tensorflow.python.ops import special_math_ops
from tensorflow.python.util import dispatch
from tensorflow.python.util.tf_export import tf_export
def _check_params(window_length, dtype):
"""Check window_length and dtype params.
Args:
window_length: A scalar value or `Tensor`.
dtype: The data type to produce. Must be a floating point type.
Returns:
window_length converted to a tensor of type int32.
Raises:
ValueError: If `dtype` is not a floating point type or window_length is not
a scalar.
"""
if not dtype.is_floating:
raise ValueError('dtype must be a floating point type. Found %s' % dtype)
window_length = ops.convert_to_tensor(window_length, dtype=dtypes.int32)
window_length.shape.assert_has_rank(0)
return window_length
@tf_export('signal.kaiser_window')
@dispatch.add_dispatch_support
def kaiser_window(window_length, beta=12., dtype=dtypes.float32, name=None):
"""Generate a [Kaiser window][kaiser].
Args:
window_length: A scalar `Tensor` indicating the window length to generate.
beta: Beta parameter for Kaiser window, see reference below.
dtype: The data type to produce. Must be a floating point type.
name: An optional name for the operation.
Returns:
A `Tensor` of shape `[window_length]` of type `dtype`.
[kaiser]:
https://docs.scipy.org/doc/numpy/reference/generated/numpy.kaiser.html
"""
with ops.name_scope(name, 'kaiser_window'):
window_length = _check_params(window_length, dtype)
window_length_const = tensor_util.constant_value(window_length)
if window_length_const == 1:
return array_ops.ones([1], dtype=dtype)
# tf.range does not support float16 so we work with float32 initially.
halflen_float = (
math_ops.cast(window_length, dtype=dtypes.float32) - 1.0) / 2.0
arg = math_ops.range(-halflen_float, halflen_float + 0.1,
dtype=dtypes.float32)
# Convert everything into given dtype which can be float16.
arg = math_ops.cast(arg, dtype=dtype)
# I0 (modified Bessel function of the first kind) is an even function,
# i.e., I0(-x) = I0(x). Take abs(beta) so that negative beta values
# produce the same (correct) result as positive ones.
beta = math_ops.abs(math_ops.cast(beta, dtype=dtype))
one = math_ops.cast(1.0, dtype=dtype)
halflen_float = math_ops.cast(halflen_float, dtype=dtype)
num = beta * math_ops.sqrt(nn_ops.relu(
one - math_ops.square(arg / halflen_float)))
window = math_ops.exp(num - beta) * (
special_math_ops.bessel_i0e(num) / special_math_ops.bessel_i0e(beta))
return window
@tf_export('signal.kaiser_bessel_derived_window')
@dispatch.add_dispatch_support
def kaiser_bessel_derived_window(window_length, beta=12.,
dtype=dtypes.float32, name=None):
"""Generate a [Kaiser Bessel derived window][kbd].
Args:
window_length: A scalar `Tensor` indicating the window length to generate.
beta: Beta parameter for Kaiser window.
dtype: The data type to produce. Must be a floating point type.
name: An optional name for the operation.
Returns:
A `Tensor` of shape `[window_length]` of type `dtype`.
[kbd]:
https://en.wikipedia.org/wiki/Kaiser_window#Kaiser%E2%80%93Bessel-derived_(KBD)_window
"""
with ops.name_scope(name, 'kaiser_bessel_derived_window'):
window_length = _check_params(window_length, dtype)
halflen = window_length // 2
kaiserw = kaiser_window(halflen + 1, beta, dtype=dtype)
kaiserw_csum = math_ops.cumsum(kaiserw)
halfw = math_ops.sqrt(kaiserw_csum[:-1] / kaiserw_csum[-1])
window = array_ops.concat((halfw, halfw[::-1]), axis=0)
return window
@tf_export('signal.vorbis_window')
@dispatch.add_dispatch_support
def vorbis_window(window_length, dtype=dtypes.float32, name=None):
"""Generate a [Vorbis power complementary window][vorbis].
Args:
window_length: A scalar `Tensor` indicating the window length to generate.
dtype: The data type to produce. Must be a floating point type.
name: An optional name for the operation.
Returns:
A `Tensor` of shape `[window_length]` of type `dtype`.
[vorbis]:
https://en.wikipedia.org/wiki/Modified_discrete_cosine_transform#Window_functions
"""
with ops.name_scope(name, 'vorbis_window'):
window_length = _check_params(window_length, dtype)
arg = math_ops.cast(math_ops.range(window_length), dtype=dtype)
window = math_ops.sin(np.pi / 2.0 * math_ops.pow(math_ops.sin(
np.pi / math_ops.cast(window_length, dtype=dtype) *
(arg + 0.5)), 2.0))
return window
@tf_export('signal.hann_window')
@dispatch.add_dispatch_support
def hann_window(window_length, periodic=True, dtype=dtypes.float32, name=None):
"""Generate a [Hann window][hann].
Args:
window_length: A scalar `Tensor` indicating the window length to generate.
periodic: A bool `Tensor` indicating whether to generate a periodic or
symmetric window. Periodic windows are typically used for spectral
analysis while symmetric windows are typically used for digital
filter design.
dtype: The data type to produce. Must be a floating point type.
name: An optional name for the operation.
Returns:
A `Tensor` of shape `[window_length]` of type `dtype`.
Raises:
ValueError: If `dtype` is not a floating point type.
[hann]: https://en.wikipedia.org/wiki/Window_function#Hann_and_Hamming_windows
"""
return _raised_cosine_window(name, 'hann_window', window_length, periodic,
dtype, 0.5, 0.5)
@tf_export('signal.hamming_window')
@dispatch.add_dispatch_support
def hamming_window(window_length, periodic=True, dtype=dtypes.float32,
name=None):
"""Generate a [Hamming][hamming] window.
Args:
window_length: A scalar `Tensor` indicating the window length to generate.
periodic: A bool `Tensor` indicating whether to generate a periodic or
symmetric window. Periodic windows are typically used for spectral
analysis while symmetric windows are typically used for digital
filter design.
dtype: The data type to produce. Must be a floating point type.
name: An optional name for the operation.
Returns:
A `Tensor` of shape `[window_length]` of type `dtype`.
Raises:
ValueError: If `dtype` is not a floating point type.
[hamming]:
https://en.wikipedia.org/wiki/Window_function#Hann_and_Hamming_windows
"""
return _raised_cosine_window(name, 'hamming_window', window_length, periodic,
dtype, 0.54, 0.46)
def _raised_cosine_window(name, default_name, window_length, periodic,
dtype, a, b):
"""Helper function for computing a raised cosine window.
Args:
name: Name to use for the scope.
default_name: Default name to use for the scope.
window_length: A scalar `Tensor` or integer indicating the window length.
periodic: A bool `Tensor` indicating whether to generate a periodic or
symmetric window.
dtype: A floating point `DType`.
a: The alpha parameter to the raised cosine window.
b: The beta parameter to the raised cosine window.
Returns:
A `Tensor` of shape `[window_length]` of type `dtype`.
Raises:
ValueError: If `dtype` is not a floating point type or `window_length` is
not scalar or `periodic` is not scalar.
"""
if not dtype.is_floating:
raise ValueError('dtype must be a floating point type. Found %s' % dtype)
with ops.name_scope(name, default_name, [window_length, periodic]):
window_length = ops.convert_to_tensor(window_length, dtype=dtypes.int32,
name='window_length')
window_length.shape.assert_has_rank(0)
window_length_const = tensor_util.constant_value(window_length)
if window_length_const == 1:
return array_ops.ones([1], dtype=dtype)
periodic = math_ops.cast(
ops.convert_to_tensor(periodic, dtype=dtypes.bool, name='periodic'),
dtypes.int32)
periodic.shape.assert_has_rank(0)
even = 1 - math_ops.mod(window_length, 2)
n = math_ops.cast(window_length + periodic * even - 1, dtype=dtype)
count = math_ops.cast(math_ops.range(window_length), dtype)
cos_arg = constant_op.constant(2 * np.pi, dtype=dtype) * count / n
if window_length_const is not None:
return math_ops.cast(a - b * math_ops.cos(cos_arg), dtype=dtype)
return cond.cond(
math_ops.equal(window_length, 1),
lambda: array_ops.ones([window_length], dtype=dtype),
lambda: math_ops.cast(a - b * math_ops.cos(cos_arg), dtype=dtype))