File size: 10,986 Bytes
5238467 |
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 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import pytest
import torch
from audiocraft.modules.codebooks_patterns import (
DelayedPatternProvider,
ParallelPatternProvider,
Pattern,
UnrolledPatternProvider,
)
class TestParallelPatternProvider:
@pytest.mark.parametrize("n_q", [1, 4, 32])
@pytest.mark.parametrize("timesteps", [0, 1, 16, 100])
def test_get_pattern(self, n_q: int, timesteps: int):
provider = ParallelPatternProvider(n_q)
pattern = provider.get_pattern(timesteps)
# + 1 to account for 1st step
assert len(pattern.layout) == timesteps + 1
@pytest.mark.parametrize("n_q", [1, 4, 32])
@pytest.mark.parametrize("timesteps", [8, 16, 100])
def test_pattern_content(self, n_q: int, timesteps: int):
provider = ParallelPatternProvider(n_q)
pattern = provider.get_pattern(timesteps)
for s, v in enumerate(pattern.layout):
for i, code in enumerate(v):
assert i == code.q
assert code.t == s - 1 # account for the 1st empty step
@pytest.mark.parametrize("n_q", [1, 4, 32])
@pytest.mark.parametrize("timesteps", [8, 16, 100])
def test_pattern_max_delay(self, n_q: int, timesteps: int):
provider = ParallelPatternProvider(n_q)
pattern = provider.get_pattern(timesteps)
assert pattern.max_delay == 0
assert len(pattern.valid_layout) == len(pattern.layout) - pattern.max_delay
class TestDelayedPatternProvider:
@pytest.mark.parametrize("n_q", [1, 4, 32])
@pytest.mark.parametrize("timesteps", [0, 1, 16, 100])
def test_get_pattern(self, n_q: int, timesteps: int):
delays = [
list(range(n_q)),
[0] + [1] * (n_q - 1),
[0] + [4] * (n_q - 1),
]
for delay in delays:
provider = DelayedPatternProvider(n_q, delay)
pattern = provider.get_pattern(timesteps)
# + 1 to account for 1st step
assert len(pattern.layout) == timesteps + max(delay) + 1
@pytest.mark.parametrize("n_q", [1, 4, 32])
@pytest.mark.parametrize("timesteps", [8, 16, 100])
def test_pattern_content(self, n_q: int, timesteps: int):
provider = DelayedPatternProvider(n_q)
pattern = provider.get_pattern(timesteps)
for s, v in enumerate(pattern.layout):
for i, code in enumerate(v):
assert i == code.q
assert code.t == max(0, s - code.q - 1)
@pytest.mark.parametrize("timesteps", [8, 16, 100])
@pytest.mark.parametrize("delay", [[0, 1, 2, 3], [0, 1, 1, 1], [0, 3, 3, 3], [0, 3]])
def test_pattern_max_delay(self, timesteps: int, delay: list):
provider = DelayedPatternProvider(len(delay), delay)
pattern = provider.get_pattern(timesteps)
assert pattern.max_delay == max(delay)
assert len(pattern.valid_layout) == len(pattern.layout) - pattern.max_delay
class TestUnrolledPatternProvider:
@pytest.mark.parametrize("timesteps", [0, 1, 16])
@pytest.mark.parametrize("flattening", [[0, 1, 2], [0, 1, 1]])
@pytest.mark.parametrize("delays", [[0, 0, 0], [0, 5, 5]])
def test_get_pattern(self, timesteps: int, flattening: list, delays: list):
n_q = len(flattening)
max_delay = max(delays)
provider = UnrolledPatternProvider(n_q, flattening, delays)
pattern = provider.get_pattern(timesteps)
assert len(pattern.layout) == provider.num_virtual_steps(timesteps) + max_delay
@pytest.mark.parametrize("timesteps", [0, 1, 16])
@pytest.mark.parametrize("flattening", [[0, 1, 2], [0, 1, 1]])
@pytest.mark.parametrize("delays", [[0, 0, 0], [0, 5, 5]])
def test_pattern_max_delay(self, timesteps: int, flattening: list, delays: list):
n_q = len(flattening)
max_delay = max(delays)
provider = UnrolledPatternProvider(n_q, flattening, delays)
pattern = provider.get_pattern(timesteps)
assert pattern.max_delay == max_delay
class TestPattern:
def ref_build_pattern_sequence(self, z: torch.Tensor, pattern: Pattern, special_token: int):
"""Reference method to build the sequence from the pattern without using fancy scatter."""
bs, n_q, T = z.shape
z = z.cpu().numpy()
assert n_q == pattern.n_q
assert T <= pattern.timesteps
inp = torch.full((bs, n_q, len(pattern.layout)), special_token, dtype=torch.long).numpy()
inp[:] = special_token
for s, v in enumerate(pattern.layout):
for (t, q) in v:
if t < T:
inp[:, q, s] = z[:, q, t]
return torch.from_numpy(inp)
def ref_revert_pattern_sequence(self, z: torch.Tensor, pattern: Pattern, special_token: int):
"""Reference method to revert the sequence from the pattern without using fancy scatter."""
z = z.cpu().numpy()
bs, n_q, S = z.shape
assert pattern.n_q == n_q
inp = torch.full((bs, pattern.n_q, pattern.timesteps), special_token, dtype=torch.long).numpy()
inp[:] = special_token
for s, v in enumerate(pattern.layout):
for (t, q) in v:
if t < pattern.timesteps:
inp[:, q, t] = z[:, q, s]
return torch.from_numpy(inp)
def ref_revert_pattern_logits(self, z: torch.Tensor, pattern: Pattern, special_token: float):
"""Reference method to revert the logits from the pattern without using fancy scatter."""
z = z.cpu().numpy()
bs, card, n_q, S = z.shape
assert pattern.n_q == n_q
ref_layout = pattern.layout
inp = torch.full((bs, card, pattern.n_q, pattern.timesteps), special_token, dtype=torch.float).numpy()
inp[:] = special_token
for s, v in enumerate(ref_layout[1:]):
if s < S:
for (t, q) in v:
if t < pattern.timesteps:
inp[:, :, q, t] = z[:, :, q, s]
return torch.from_numpy(inp)
def _get_pattern_providers(self, n_q: int):
pattern_provider_1 = ParallelPatternProvider(n_q)
pattern_provider_2 = DelayedPatternProvider(n_q, list(range(n_q)))
pattern_provider_3 = DelayedPatternProvider(n_q, [0] + [1] * (n_q - 1))
pattern_provider_4 = UnrolledPatternProvider(
n_q, flattening=list(range(n_q)), delays=[0] * n_q
)
pattern_provider_5 = UnrolledPatternProvider(
n_q, flattening=[0] + [1] * (n_q - 1), delays=[0] * n_q
)
pattern_provider_6 = UnrolledPatternProvider(
n_q, flattening=[0] + [1] * (n_q - 1), delays=[0] + [5] * (n_q - 1)
)
return [
pattern_provider_1,
pattern_provider_2,
pattern_provider_3,
pattern_provider_4,
pattern_provider_5,
pattern_provider_6,
]
@pytest.mark.parametrize("n_q", [1, 4, 32])
@pytest.mark.parametrize("timesteps", [16, 72])
def test_build_pattern_sequence(self, n_q: int, timesteps: int):
bs = 2
card = 256
special_token = card
pattern_providers = self._get_pattern_providers(n_q)
for pattern_provider in pattern_providers:
pattern = pattern_provider.get_pattern(timesteps)
# we can correctly build the sequence from the pattern
z = torch.randint(0, card, (bs, n_q, timesteps))
ref_res = self.ref_build_pattern_sequence(z, pattern, special_token)
res, indexes, mask = pattern.build_pattern_sequence(z, special_token)
assert (res == ref_res).float().mean() == 1.0
# expected assertion fails on the number of timesteps
invalid_timesteps = [timesteps + 1]
if pattern.num_sequence_steps != pattern.timesteps:
invalid_timesteps.append(pattern.num_sequence_steps)
for i_timesteps in invalid_timesteps:
z2 = torch.randint(0, card, (bs, n_q, i_timesteps))
with pytest.raises(AssertionError):
pattern.build_pattern_sequence(z2, special_token)
# expected assertion fails on the number of codebooks
invalid_qs = [0, n_q - 1, n_q + 1]
for i_q in invalid_qs:
z3 = torch.randint(0, card, (bs, i_q, timesteps))
with pytest.raises(AssertionError):
pattern.build_pattern_sequence(z3, special_token)
@pytest.mark.parametrize("n_q", [1, 4, 32])
@pytest.mark.parametrize("timesteps", [16, 72])
def test_revert_pattern_sequence(self, n_q: int, timesteps: int):
bs = 2
card = 256
special_token = card
pattern_providers = self._get_pattern_providers(n_q)
for pattern_provider in pattern_providers:
pattern = pattern_provider.get_pattern(timesteps)
# this works assuming previous tests are successful
z = torch.randint(0, card, (bs, n_q, timesteps))
s = self.ref_build_pattern_sequence(z, pattern, special_token)
ref_out = self.ref_revert_pattern_sequence(s, pattern, special_token)
# ensure our reference script retrieve the original sequence
assert z.shape == ref_out.shape
assert (z == ref_out).float().mean() == 1.0
# now we can test the scatter version
out, indexes, mask = pattern.revert_pattern_sequence(s, special_token)
assert out.shape == ref_out.shape
assert (out == ref_out).float().mean() == 1.0
@pytest.mark.parametrize("n_q", [1, 4, 32])
@pytest.mark.parametrize("timesteps", [16, 72])
@pytest.mark.parametrize("card", [1, 2, 256, 1024])
def test_revert_pattern_logits(self, n_q: int, timesteps: int, card: int):
bs = 2
special_token = card
logits_special_token = float('nan')
pattern_providers = self._get_pattern_providers(n_q)
for pattern_provider in pattern_providers:
pattern = pattern_provider.get_pattern(timesteps)
# this works assuming previous tests are successful
z = torch.randint(0, card, (bs, n_q, timesteps))
s = self.ref_build_pattern_sequence(z, pattern, special_token)
logits = torch.randn((bs, card, n_q, s.shape[-1]))
ref_out = self.ref_revert_pattern_logits(logits, pattern, logits_special_token)
# ensure our reference script retrieve the original sequence
assert ref_out.shape == torch.Size([bs, card, n_q, timesteps])
# now we can test the scatter version
out, indexes, mask = pattern.revert_pattern_logits(logits, logits_special_token)
assert out.shape == ref_out.shape
assert (out == ref_out).float().mean() == 1.0
|