File size: 15,467 Bytes
b386992
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
# Copyright (c) 2022, NVIDIA CORPORATION.  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.

import importlib

import numpy as np
import pytest
import torch
from einops import rearrange

from nemo.collections.audio.modules.transforms import AudioToSpectrogram, SpectrogramToAudio


class TestAudioSpectrogram:
    @pytest.mark.unit
    @pytest.mark.parametrize('fft_length', [64, 512])
    @pytest.mark.parametrize('num_channels', [1, 3])
    def test_audio_to_spec(self, fft_length: int, num_channels: int):
        """Test output length for audio to spectrogram.

        Create signals of arbitrary length and check output
        length is matching the actual transform length.
        """
        hop_lengths = [fft_length // 2, fft_length // 3, fft_length // 4]
        batch_size = 4
        num_examples = 20
        random_seed = 42
        atol = 1e-6

        _rng = np.random.default_rng(seed=random_seed)

        for n in range(num_examples):

            # Generate time-domain examples with different length
            input_length = _rng.integers(low=fft_length, high=100 * fft_length, size=batch_size)  # in samples
            x = _rng.normal(size=(batch_size, num_channels, np.max(input_length)))
            x = torch.tensor(x)
            for b in range(batch_size):
                x[b, :, input_length[b] :] = 0

            for hop_length in hop_lengths:
                # Prepare transform
                audio2spec = AudioToSpectrogram(fft_length=fft_length, hop_length=hop_length)

                # Transform the whole batch
                batch_spec, batch_spec_len = audio2spec(input=x, input_length=torch.tensor(input_length))

                for b in range(batch_size):

                    # Transform just the current example
                    b_spec, b_spec_len = audio2spec(input=x[b : b + 1, :, : input_length[b]])
                    actual_len = b_spec.size(-1)

                    # Check lengths
                    assert (
                        actual_len == b_spec_len
                    ), f'Output length not matching for example ({n}, {b}) with length {input_length[n]} (hop_length={hop_length}): true {actual_len} vs calculated {b_spec_len}.'

                    assert (
                        actual_len == batch_spec_len[b]
                    ), f'Output length not matching for example ({n}, {b}) with length {input_length[n]} (hop_length={hop_length}): true {actual_len} vs calculated batch len {batch_spec_len[b]}.'

                    # Make sure transforming a batch is the same as transforming individual examples
                    assert torch.allclose(
                        batch_spec[b, ..., :actual_len], b_spec, atol=atol
                    ), f'Spectrograms not matching for example ({n}, {b}) with length {input_length[b]} (hop_length={hop_length})'

    @pytest.mark.unit
    @pytest.mark.parametrize('fft_length', [64, 512])
    @pytest.mark.parametrize('num_channels', [1, 3])
    def test_spec_to_audio(self, fft_length: int, num_channels: int):
        """Test output length for spectrogram to audio.

        Create signals of arbitrary length and check output
        length is matching the actual transform length.
        """
        hop_lengths = [fft_length // 2, fft_length // 3, fft_length // 4]
        batch_size = 4
        num_examples = 20
        random_seed = 42
        atol = 1e-6

        _rng = np.random.default_rng(seed=random_seed)

        for n in range(num_examples):

            # Generate spectrogram examples with different lengths
            input_length = _rng.integers(low=10, high=100, size=batch_size)  # in frames
            input_shape = (batch_size, num_channels, fft_length // 2 + 1, np.max(input_length))
            spec = _rng.normal(size=input_shape) + 1j * _rng.normal(size=input_shape)
            spec = torch.tensor(spec)
            spec[..., 0, :] = spec[..., 0, :].real
            spec[..., -1, :] = spec[..., -1, :].real
            for b in range(batch_size):
                spec[b, ..., input_length[b] :] = 0

            for hop_length in hop_lengths:
                # Prepare transform
                spec2audio = SpectrogramToAudio(fft_length=fft_length, hop_length=hop_length)

                # Transform the whole batch
                batch_x, batch_x_len = spec2audio(input=spec, input_length=torch.tensor(input_length))

                for b in range(batch_size):

                    # Transform just the current example
                    b_x, b_x_len = spec2audio(input=spec[b : b + 1, ..., : input_length[b]])

                    actual_len = b_x.size(-1)

                    # Check lengths
                    assert (
                        b_x_len == actual_len
                    ), f'Output length not matching for example ({n}, {b}) with {input_length[b]} frames (hop_length={hop_length}): true {actual_len} vs calculated {b_x_len}.'

                    assert (
                        batch_x_len[b] == actual_len
                    ), f'Output length not matching for example ({n}, {b}) with {input_length[b]} frames (hop_length={hop_length}): true {actual_len} vs calculated batch {batch_x_len[b]}.'

                    # Make sure transforming a batch is the same as transforming individual examples
                    if input_length[b] < spec.size(-1):
                        # Discard the last bit of the signal which differs due to number of frames in batch (with zero padded frames) vs individual (only valid frames).
                        # The reason for this difference is normalization with `window_sumsquare` of the inverse STFT. More specifically,
                        # batched and non-batched transform are using on a different number of frames.
                        tail_length = max(fft_length // 2 - hop_length, 0)
                    else:
                        tail_length = 0
                    valid_len = actual_len - tail_length
                    batch_x_valid = batch_x[b, :, :valid_len]
                    b_x_valid = b_x[..., :valid_len]
                    assert torch.allclose(
                        batch_x_valid, b_x_valid, atol=atol
                    ), f'Signals not matching for example ({n}, {b}) with length {input_length[b]} (hop_length={hop_length}): max abs diff {torch.max(torch.abs(batch_x_valid-b_x_valid))} at {torch.argmax(torch.abs(batch_x_valid-b_x_valid))}'

    @pytest.mark.unit
    @pytest.mark.parametrize('fft_length', [128, 1024])
    @pytest.mark.parametrize('num_channels', [1, 4])
    @pytest.mark.parametrize('magnitude_power', [0.5, 1, 2])
    @pytest.mark.parametrize('scale', [0.1, 1.0])
    def test_audio_to_spectrogram_reconstruction(
        self, fft_length: int, num_channels: int, magnitude_power: float, scale: float
    ):
        """Test analysis and synthesis transform result in a perfect reconstruction."""
        batch_size = 4
        num_samples = fft_length * 50
        num_examples = 25
        random_seed = 42
        atol = 1e-6

        _rng = np.random.default_rng(seed=random_seed)

        hop_lengths = [fft_length // 2, fft_length // 4]

        for hop_length in hop_lengths:
            audio2spec = AudioToSpectrogram(
                fft_length=fft_length, hop_length=hop_length, magnitude_power=magnitude_power, scale=scale
            )
            spec2audio = SpectrogramToAudio(
                fft_length=fft_length, hop_length=hop_length, magnitude_power=magnitude_power, scale=scale
            )

            for n in range(num_examples):
                x = _rng.normal(size=(batch_size, num_channels, num_samples))

                x_spec, x_spec_length = audio2spec(input=torch.Tensor(x))
                x_hat, x_hat_length = spec2audio(input=x_spec, input_length=x_spec_length)

                assert np.allclose(
                    x_hat.cpu().detach().numpy(), x, atol=atol
                ), f'Reconstructed not matching for example {n} (hop length {hop_length})'

    @pytest.mark.unit
    @pytest.mark.parametrize('fft_length', [128, 512])
    @pytest.mark.parametrize('num_channels', [1, 4])
    @pytest.mark.parametrize('magnitude_power', [0.5, 1])
    @pytest.mark.parametrize('scale', [0.1, 1.0])
    def test_match_reference_implementation(
        self, fft_length: int, num_channels: int, magnitude_power: float, scale: float
    ):
        """Test analysis and synthesis transforms match reference implementation."""
        batch_size = 4
        num_samples = fft_length * 50
        num_examples = 8
        random_seed = 42
        atol = 1e-6

        _rng = np.random.default_rng(seed=random_seed)

        hop_lengths = [fft_length // 2, fft_length // 4]

        for hop_length in hop_lengths:
            audio2spec = AudioToSpectrogram(
                fft_length=fft_length, hop_length=hop_length, magnitude_power=magnitude_power, scale=scale
            )
            spec2audio = SpectrogramToAudio(
                fft_length=fft_length, hop_length=hop_length, magnitude_power=magnitude_power, scale=scale
            )

            # Reference implementations
            ref_window = torch.hann_window(fft_length)

            def audio2spec_ref(x):
                # Transform each channel and batch example separately
                x_spec = []
                for b in range(batch_size):
                    for c in range(num_channels):
                        x_spec_bc = torch.stft(
                            input=x[b, c, :],
                            n_fft=fft_length,
                            hop_length=hop_length,
                            win_length=fft_length,
                            window=ref_window,
                            center=True,
                            pad_mode='constant',
                            normalized=False,
                            onesided=True,
                            return_complex=True,
                        )
                        x_spec.append(x_spec_bc)
                x_spec = torch.stack(x_spec, dim=0)
                x_spec = rearrange(x_spec, '(B C) F N -> B C F N', B=batch_size, C=num_channels)
                # magnitude compression and scaling
                x_spec = (
                    torch.pow(x_spec.abs(), magnitude_power) * torch.exp(1j * x_spec.angle())
                    if magnitude_power != 1
                    else x_spec
                )
                x_spec = x_spec * scale if scale != 1 else x_spec
                return x_spec

            def spec2audio_ref(x_spec):
                # scaling and magnitude compression
                x_spec = x_spec / scale if scale != 1 else x_spec
                x_spec = (
                    torch.pow(x_spec.abs(), 1 / magnitude_power) * torch.exp(1j * x_spec.angle())
                    if magnitude_power != 1
                    else x_spec
                )
                # Transform each channel and batch example separately
                x = []
                for b in range(batch_size):
                    for c in range(num_channels):
                        x_bc = torch.istft(
                            input=x_spec[b, c, ...],
                            n_fft=fft_length,
                            hop_length=hop_length,
                            win_length=fft_length,
                            window=ref_window,
                            center=True,
                            normalized=False,
                            onesided=True,
                            return_complex=False,
                        )
                        x.append(x_bc)
                x = torch.stack(x, dim=0)
                x = rearrange(x, '(B C) T -> B C T', B=batch_size, C=num_channels)
                return x

            for n in range(num_examples):
                x = _rng.normal(size=(batch_size, num_channels, num_samples))

                # Test analysis
                x_spec, x_spec_length = audio2spec(input=torch.Tensor(x))
                x_spec_ref = audio2spec_ref(torch.Tensor(x))

                assert torch.allclose(
                    x_spec, x_spec_ref, atol=atol
                ), f'Analysis not matching for example {n} (hop length {hop_length})'

                # Test synthesis
                x_hat, _ = spec2audio(input=x_spec, input_length=x_spec_length)
                x_hat_ref = spec2audio_ref(x_spec_ref)

                assert torch.allclose(
                    x_hat, x_hat_ref, atol=atol
                ), f'Synthesis not matching for example {n} (hop length {hop_length})'

    @pytest.mark.unit
    @pytest.mark.parametrize('fft_length', [13, 63])
    def test_invalid_length(self, fft_length: int):
        """Test initializing transforms with invalid length."""

        # Only even fft lengths are supported
        with pytest.raises(ValueError):
            AudioToSpectrogram(fft_length=fft_length, hop_length=fft_length // 2)
        with pytest.raises(ValueError):
            SpectrogramToAudio(fft_length=fft_length, hop_length=fft_length // 2)

    @pytest.mark.unit
    @pytest.mark.parametrize('fft_length', [32])
    def test_invalid_compression(self, fft_length: int):
        """Test initializing transforms with invalid compression."""
        # Compression must be positive
        with pytest.raises(ValueError):
            AudioToSpectrogram(fft_length=fft_length, hop_length=fft_length // 2, magnitude_power=0.0)
        with pytest.raises(ValueError):
            SpectrogramToAudio(fft_length=fft_length, hop_length=fft_length // 2, magnitude_power=0.0)
        with pytest.raises(ValueError):
            AudioToSpectrogram(fft_length=fft_length, hop_length=fft_length // 2, magnitude_power=-1.0)
        with pytest.raises(ValueError):
            SpectrogramToAudio(fft_length=fft_length, hop_length=fft_length // 2, magnitude_power=-1.0)

        # Scaling must be positive
        with pytest.raises(ValueError):
            AudioToSpectrogram(fft_length=fft_length, hop_length=fft_length // 2, scale=0.0)
        with pytest.raises(ValueError):
            SpectrogramToAudio(fft_length=fft_length, hop_length=fft_length // 2, scale=0.0)
        with pytest.raises(ValueError):
            AudioToSpectrogram(fft_length=fft_length, hop_length=fft_length // 2, scale=-1.0)
        with pytest.raises(ValueError):
            SpectrogramToAudio(fft_length=fft_length, hop_length=fft_length // 2, scale=-1.0)

    @pytest.mark.unit
    @pytest.mark.parametrize('fft_length', [32])
    def test_invalid_spec_to_audio_input(self, fft_length: int):
        """Test invalid input for spec to audio transform."""
        s2a = SpectrogramToAudio(fft_length=fft_length, hop_length=fft_length // 2)
        # Input must be complex
        s2a(input=torch.randn(1, 1, fft_length // 2 + 1, 100, dtype=torch.cfloat))

        # Input must be complex
        with pytest.raises(ValueError):
            s2a(input=torch.randn(1, 1, fft_length // 2 + 1, 100))