Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/tensorflow/python/ops/signal/shape_ops.py: 16%
85 statements
« prev ^ index » next coverage.py v7.4.0, created at 2024-01-03 07:57 +0000
1# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
15"""General shape ops for frames."""
17import numpy as np
19from tensorflow.python.framework import constant_op
20from tensorflow.python.framework import ops
21from tensorflow.python.framework import tensor_util
22from tensorflow.python.ops import array_ops
23from tensorflow.python.ops import math_ops
24from tensorflow.python.ops.signal import util_ops
25from tensorflow.python.util import dispatch
26from tensorflow.python.util.tf_export import tf_export
29def _infer_frame_shape(signal, frame_length, frame_step, pad_end, axis):
30 """Infers the shape of the return value of `frame`."""
31 frame_length = tensor_util.constant_value(frame_length)
32 frame_step = tensor_util.constant_value(frame_step)
33 axis = tensor_util.constant_value(axis)
34 if signal.shape.ndims is None:
35 return None
36 if axis is None:
37 return [None] * (signal.shape.ndims + 1)
39 signal_shape = signal.shape.as_list()
40 num_frames = None
41 frame_axis = signal_shape[axis]
42 outer_dimensions = signal_shape[:axis]
43 inner_dimensions = signal_shape[axis:][1:]
44 if signal_shape and frame_axis is not None:
45 if frame_step is not None and pad_end:
46 # Double negative is so that we round up.
47 num_frames = max(0, -(-frame_axis // frame_step))
48 elif frame_step is not None and frame_length is not None:
49 assert not pad_end
50 num_frames = max(
51 0, (frame_axis - frame_length + frame_step) // frame_step)
52 return outer_dimensions + [num_frames, frame_length] + inner_dimensions
55@tf_export("signal.frame")
56@dispatch.add_dispatch_support
57def frame(signal, frame_length, frame_step, pad_end=False, pad_value=0, axis=-1,
58 name=None):
59 """Expands `signal`'s `axis` dimension into frames of `frame_length`.
61 Slides a window of size `frame_length` over `signal`'s `axis` dimension
62 with a stride of `frame_step`, replacing the `axis` dimension with
63 `[frames, frame_length]` frames.
65 If `pad_end` is True, window positions that are past the end of the `axis`
66 dimension are padded with `pad_value` until the window moves fully past the
67 end of the dimension. Otherwise, only window positions that fully overlap the
68 `axis` dimension are produced.
70 For example:
72 >>> # A batch size 3 tensor of 9152 audio samples.
73 >>> audio = tf.random.normal([3, 9152])
74 >>>
75 >>> # Compute overlapping frames of length 512 with a step of 180 (frames overlap
76 >>> # by 332 samples). By default, only 49 frames are generated since a frame
77 >>> # with start position j*180 for j > 48 would overhang the end.
78 >>> frames = tf.signal.frame(audio, 512, 180)
79 >>> frames.shape.assert_is_compatible_with([3, 49, 512])
80 >>>
81 >>> # When pad_end is enabled, the final two frames are kept (padded with zeros).
82 >>> frames = tf.signal.frame(audio, 512, 180, pad_end=True)
83 >>> frames.shape.assert_is_compatible_with([3, 51, 512])
85 If the dimension along `axis` is N, and `pad_end=False`, the number of frames
86 can be computed by:
87 ```python
88 num_frames = 1 + (N - frame_size) // frame_step
89 ```
90 If `pad_end=True`, the number of frames can be computed by:
91 ```python
92 num_frames = -(-N // frame_step) # ceiling division
93 ```
95 Args:
96 signal: A `[..., samples, ...]` `Tensor`. The rank and dimensions
97 may be unknown. Rank must be at least 1.
98 frame_length: The frame length in samples. An integer or scalar `Tensor`.
99 frame_step: The frame hop size in samples. An integer or scalar `Tensor`.
100 pad_end: Whether to pad the end of `signal` with `pad_value`.
101 pad_value: An optional scalar `Tensor` to use where the input signal
102 does not exist when `pad_end` is True.
103 axis: A scalar integer `Tensor` indicating the axis to frame. Defaults to
104 the last axis. Supports negative values for indexing from the end.
105 name: An optional name for the operation.
107 Returns:
108 A `Tensor` of frames with shape `[..., num_frames, frame_length, ...]`.
110 Raises:
111 ValueError: If `frame_length`, `frame_step`, `pad_value`, or `axis` are not
112 scalar.
113 """
114 with ops.name_scope(name, "frame", [signal, frame_length, frame_step,
115 pad_value]):
116 signal = ops.convert_to_tensor(signal, name="signal")
117 frame_length = ops.convert_to_tensor(frame_length, name="frame_length")
118 frame_step = ops.convert_to_tensor(frame_step, name="frame_step")
119 axis = ops.convert_to_tensor(axis, name="axis")
121 signal.shape.with_rank_at_least(1)
122 frame_length.shape.assert_has_rank(0)
123 frame_step.shape.assert_has_rank(0)
124 axis.shape.assert_has_rank(0)
126 result_shape = _infer_frame_shape(signal, frame_length, frame_step, pad_end,
127 axis)
129 def maybe_constant(val):
130 val_static = tensor_util.constant_value(val)
131 return (val_static, True) if val_static is not None else (val, False)
133 signal_shape, signal_shape_is_static = maybe_constant(
134 array_ops.shape(signal))
135 axis, axis_is_static = maybe_constant(axis)
137 if signal_shape_is_static and axis_is_static:
138 # Axis can be negative. Convert it to positive.
139 axis = range(len(signal_shape))[axis]
140 outer_dimensions, length_samples, inner_dimensions = np.split(
141 signal_shape, indices_or_sections=[axis, axis + 1])
142 length_samples = length_samples.item()
143 else:
144 signal_rank = array_ops.rank(signal)
145 # Axis can be negative. Convert it to positive.
146 axis = math_ops.range(signal_rank)[axis]
147 outer_dimensions, length_samples, inner_dimensions = array_ops.split(
148 signal_shape, [axis, 1, signal_rank - 1 - axis])
149 length_samples = array_ops.reshape(length_samples, [])
150 num_outer_dimensions = array_ops.size(outer_dimensions)
151 num_inner_dimensions = array_ops.size(inner_dimensions)
153 # If padding is requested, pad the input signal tensor with pad_value.
154 if pad_end:
155 pad_value = ops.convert_to_tensor(pad_value, signal.dtype)
156 pad_value.shape.assert_has_rank(0)
158 # Calculate number of frames, using double negatives to round up.
159 num_frames = -(-length_samples // frame_step)
161 # Pad the signal by up to frame_length samples based on how many samples
162 # are remaining starting from last_frame_position.
163 pad_samples = math_ops.maximum(
164 0, frame_length + frame_step * (num_frames - 1) - length_samples)
166 # Pad the inner dimension of signal by pad_samples.
167 paddings = array_ops.concat([
168 array_ops.zeros([num_outer_dimensions, 2], dtype=pad_samples.dtype),
169 ops.convert_to_tensor([[0, pad_samples]]),
170 array_ops.zeros([num_inner_dimensions, 2], dtype=pad_samples.dtype)
171 ], 0)
172 signal = array_ops.pad(signal, paddings, constant_values=pad_value)
174 signal_shape = array_ops.shape(signal)
175 length_samples = signal_shape[axis]
176 else:
177 num_frames = math_ops.maximum(
178 constant_op.constant(0, dtype=frame_length.dtype),
179 1 + (length_samples - frame_length) // frame_step)
181 subframe_length, _ = maybe_constant(util_ops.gcd(frame_length, frame_step))
182 subframes_per_frame = frame_length // subframe_length
183 subframes_per_hop = frame_step // subframe_length
184 num_subframes = length_samples // subframe_length
186 slice_shape = array_ops.concat([outer_dimensions,
187 [num_subframes * subframe_length],
188 inner_dimensions], 0)
189 subframe_shape = array_ops.concat([outer_dimensions,
190 [num_subframes, subframe_length],
191 inner_dimensions], 0)
192 subframes = array_ops.reshape(array_ops.strided_slice(
193 signal, array_ops.zeros_like(signal_shape),
194 slice_shape), subframe_shape)
196 # frame_selector is a [num_frames, subframes_per_frame] tensor
197 # that indexes into the appropriate frame in subframes. For example:
198 # [[0, 0, 0, 0], [2, 2, 2, 2], [4, 4, 4, 4]]
199 frame_selector = array_ops.reshape(
200 math_ops.range(num_frames, dtype=frame_length.dtype) *
201 subframes_per_hop, [num_frames, 1])
203 # subframe_selector is a [num_frames, subframes_per_frame] tensor
204 # that indexes into the appropriate subframe within a frame. For example:
205 # [[0, 1, 2, 3], [0, 1, 2, 3], [0, 1, 2, 3]]
206 subframe_selector = array_ops.reshape(
207 math_ops.range(subframes_per_frame, dtype=frame_length.dtype),
208 [1, subframes_per_frame])
210 # Adding the 2 selector tensors together produces a [num_frames,
211 # subframes_per_frame] tensor of indices to use with tf.gather to select
212 # subframes from subframes. We then reshape the inner-most
213 # subframes_per_frame dimension to stitch the subframes together into
214 # frames. For example: [[0, 1, 2, 3], [2, 3, 4, 5], [4, 5, 6, 7]].
215 selector = frame_selector + subframe_selector
217 # Dtypes have to match.
218 outer_dimensions = ops.convert_to_tensor(outer_dimensions)
219 inner_dimensions = ops.convert_to_tensor(
220 inner_dimensions, dtype=outer_dimensions.dtype)
221 mid_dimensions = ops.convert_to_tensor([num_frames, frame_length],
222 dtype=outer_dimensions.dtype)
224 frames = array_ops.reshape(
225 array_ops.gather(subframes, selector, axis=axis),
226 array_ops.concat([outer_dimensions, mid_dimensions, inner_dimensions],
227 0))
229 if result_shape:
230 frames.set_shape(result_shape)
231 return frames