facebookresearch--audiocraft
Alexandre Défossez (autodoc)
671 行
38 KiB
HTML
671 行
38 KiB
HTML
<!doctype html>
|
||
<html lang="en">
|
||
<head>
|
||
<meta charset="utf-8">
|
||
<meta name="viewport" content="width=device-width, initial-scale=1, minimum-scale=1">
|
||
<meta name="generator" content="pdoc3 0.11.5">
|
||
<title>audiocraft.models.lm_magnet API documentation</title>
|
||
<meta name="description" content="">
|
||
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/10up-sanitize.css/13.0.0/sanitize.min.css" integrity="sha512-y1dtMcuvtTMJc1yPgEqF0ZjQbhnc/bFhyvIyVNb9Zk5mIGtqVaAB1Ttl28su8AvFMOY0EwRbAe+HCLqj6W7/KA==" crossorigin>
|
||
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/10up-sanitize.css/13.0.0/typography.min.css" integrity="sha512-Y1DYSb995BAfxobCkKepB1BqJJTPrOp3zPL74AWFugHHmmdcvO+C48WLrUOlhGMc0QG7AE3f7gmvvcrmX2fDoA==" crossorigin>
|
||
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/11.9.0/styles/default.min.css" crossorigin>
|
||
<style>:root{--highlight-color:#fe9}.flex{display:flex !important}body{line-height:1.5em}#content{padding:20px}#sidebar{padding:1.5em;overflow:hidden}#sidebar > *:last-child{margin-bottom:2cm}.http-server-breadcrumbs{font-size:130%;margin:0 0 15px 0}#footer{font-size:.75em;padding:5px 30px;border-top:1px solid #ddd;text-align:right}#footer p{margin:0 0 0 1em;display:inline-block}#footer p:last-child{margin-right:30px}h1,h2,h3,h4,h5{font-weight:300}h1{font-size:2.5em;line-height:1.1em}h2{font-size:1.75em;margin:2em 0 .50em 0}h3{font-size:1.4em;margin:1.6em 0 .7em 0}h4{margin:0;font-size:105%}h1:target,h2:target,h3:target,h4:target,h5:target,h6:target{background:var(--highlight-color);padding:.2em 0}a{color:#058;text-decoration:none;transition:color .2s ease-in-out}a:visited{color:#503}a:hover{color:#b62}.title code{font-weight:bold}h2[id^="header-"]{margin-top:2em}.ident{color:#900;font-weight:bold}pre code{font-size:.8em;line-height:1.4em;padding:1em;display:block}code{background:#f3f3f3;font-family:"DejaVu Sans Mono",monospace;padding:1px 4px;overflow-wrap:break-word}h1 code{background:transparent}pre{border-top:1px solid #ccc;border-bottom:1px solid #ccc;margin:1em 0}#http-server-module-list{display:flex;flex-flow:column}#http-server-module-list div{display:flex}#http-server-module-list dt{min-width:10%}#http-server-module-list p{margin-top:0}.toc ul,#index{list-style-type:none;margin:0;padding:0}#index code{background:transparent}#index h3{border-bottom:1px solid #ddd}#index ul{padding:0}#index h4{margin-top:.6em;font-weight:bold}@media (min-width:200ex){#index .two-column{column-count:2}}@media (min-width:300ex){#index .two-column{column-count:3}}dl{margin-bottom:2em}dl dl:last-child{margin-bottom:4em}dd{margin:0 0 1em 3em}#header-classes + dl > dd{margin-bottom:3em}dd dd{margin-left:2em}dd p{margin:10px 0}.name{background:#eee;font-size:.85em;padding:5px 10px;display:inline-block;min-width:40%}.name:hover{background:#e0e0e0}dt:target .name{background:var(--highlight-color)}.name > span:first-child{white-space:nowrap}.name.class > span:nth-child(2){margin-left:.4em}.inherited{color:#999;border-left:5px solid #eee;padding-left:1em}.inheritance em{font-style:normal;font-weight:bold}.desc h2{font-weight:400;font-size:1.25em}.desc h3{font-size:1em}.desc dt code{background:inherit}.source > summary,.git-link-div{color:#666;text-align:right;font-weight:400;font-size:.8em;text-transform:uppercase}.source summary > *{white-space:nowrap;cursor:pointer}.git-link{color:inherit;margin-left:1em}.source pre{max-height:500px;overflow:auto;margin:0}.source pre code{font-size:12px;overflow:visible;min-width:max-content}.hlist{list-style:none}.hlist li{display:inline}.hlist li:after{content:',\2002'}.hlist li:last-child:after{content:none}.hlist .hlist{display:inline;padding-left:1em}img{max-width:100%}td{padding:0 .5em}.admonition{padding:.1em 1em;margin:1em 0}.admonition-title{font-weight:bold}.admonition.note,.admonition.info,.admonition.important{background:#aef}.admonition.todo,.admonition.versionadded,.admonition.tip,.admonition.hint{background:#dfd}.admonition.warning,.admonition.versionchanged,.admonition.deprecated{background:#fd4}.admonition.error,.admonition.danger,.admonition.caution{background:lightpink}</style>
|
||
<style media="screen and (min-width: 700px)">@media screen and (min-width:700px){#sidebar{width:30%;height:100vh;overflow:auto;position:sticky;top:0}#content{width:70%;max-width:100ch;padding:3em 4em;border-left:1px solid #ddd}pre code{font-size:1em}.name{font-size:1em}main{display:flex;flex-direction:row-reverse;justify-content:flex-end}.toc ul ul,#index ul ul{padding-left:1em}.toc > ul > li{margin-top:.5em}}</style>
|
||
<style media="print">@media print{#sidebar h1{page-break-before:always}.source{display:none}}@media print{*{background:transparent !important;color:#000 !important;box-shadow:none !important;text-shadow:none !important}a[href]:after{content:" (" attr(href) ")";font-size:90%}a[href][title]:after{content:none}abbr[title]:after{content:" (" attr(title) ")"}.ir a:after,a[href^="javascript:"]:after,a[href^="#"]:after{content:""}pre,blockquote{border:1px solid #999;page-break-inside:avoid}thead{display:table-header-group}tr,img{page-break-inside:avoid}img{max-width:100% !important}@page{margin:0.5cm}p,h2,h3{orphans:3;widows:3}h1,h2,h3,h4,h5,h6{page-break-after:avoid}}</style>
|
||
<script defer src="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/11.9.0/highlight.min.js" integrity="sha512-D9gUyxqja7hBtkWpPWGt9wfbfaMGVt9gnyCvYa+jojwwPHLCzUm5i8rpk7vD7wNee9bA35eYIjobYPaQuKS1MQ==" crossorigin></script>
|
||
<script>window.addEventListener('DOMContentLoaded', () => {
|
||
hljs.configure({languages: ['bash', 'css', 'diff', 'graphql', 'ini', 'javascript', 'json', 'plaintext', 'python', 'python-repl', 'rust', 'shell', 'sql', 'typescript', 'xml', 'yaml']});
|
||
hljs.highlightAll();
|
||
/* Collapse source docstrings */
|
||
setTimeout(() => {
|
||
[...document.querySelectorAll('.hljs.language-python > .hljs-string')]
|
||
.filter(el => el.innerHTML.length > 200 && ['"""', "'''"].includes(el.innerHTML.substring(0, 3)))
|
||
.forEach(el => {
|
||
let d = document.createElement('details');
|
||
d.classList.add('hljs-string');
|
||
d.innerHTML = '<summary>"""</summary>' + el.innerHTML.substring(3);
|
||
el.replaceWith(d);
|
||
});
|
||
}, 100);
|
||
})</script>
|
||
</head>
|
||
<body>
|
||
<main>
|
||
<article id="content">
|
||
<header>
|
||
<h1 class="title">Module <code>audiocraft.models.lm_magnet</code></h1>
|
||
</header>
|
||
<section id="section-intro">
|
||
</section>
|
||
<section>
|
||
</section>
|
||
<section>
|
||
</section>
|
||
<section>
|
||
</section>
|
||
<section>
|
||
<h2 class="section-title" id="header-classes">Classes</h2>
|
||
<dl>
|
||
<dt id="audiocraft.models.lm_magnet.MagnetLMModel"><code class="flex name class">
|
||
<span>class <span class="ident">MagnetLMModel</span></span>
|
||
<span>(</span><span>subcodes_context: int = 5,<br>compression_model_framerate: int = 50,<br>segment_duration: int = 10,<br>span_len: int = 3,<br>**kwargs)</span>
|
||
</code></dt>
|
||
<dd>
|
||
<details class="source">
|
||
<summary>
|
||
<span>Expand source code</span>
|
||
</summary>
|
||
<pre><code class="python">class MagnetLMModel(LMModel):
|
||
"""Transformer-based, non-autoregressive model, operates on multiple streams of audio tokens (MAGNeT).
|
||
Args:
|
||
subcodes_context (int): The number of timesteps attended in the self-attention blocks of codebooks > 0.
|
||
When set to -1, attention is unrestricted and all timesteps are attended. Defaults to 5.
|
||
compression_model_framerate (int): frame rate of the audio tokenizer.
|
||
segment_duration (int): Sample length in seconds.
|
||
span_len (int): Determines the length of masking spans. This is the minimal length of consecutive masked tokens,
|
||
for both training and inference. Defaults to 3.
|
||
**kwargs: Additional parameters for the LMModel.
|
||
"""
|
||
def __init__(self, subcodes_context: int = 5, compression_model_framerate: int = 50,
|
||
segment_duration: int = 10, span_len: int = 3, **kwargs):
|
||
super().__init__(**kwargs)
|
||
self.causal = kwargs['causal']
|
||
self.subcodes_context = subcodes_context
|
||
self.span_len = span_len
|
||
self._build_attn_masks(compression_model_framerate=compression_model_framerate,
|
||
segment_duration=segment_duration,
|
||
num_heads=kwargs['num_heads'],
|
||
device=kwargs['device'], dtype=kwargs['dtype'])
|
||
|
||
def restricted_context_attn_mask(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
|
||
"""Creates a restricted attention mask (local attention map) where the context
|
||
is determined by self.subcodes_context.
|
||
Args:
|
||
seq_len (int): token sequence length.
|
||
device (torch.device): device of the output tensor.
|
||
dtype (torch.dtype): data type of the output tensor.
|
||
Returns:
|
||
torch.Tensor: The restricted attention mask.
|
||
"""
|
||
# Return a context restricted non-causal att mask
|
||
queries_pos = torch.arange(seq_len, device=device).view(-1, 1)
|
||
keys_pos = torch.arange(seq_len, device=device).view(1, -1)
|
||
|
||
delta = queries_pos - keys_pos
|
||
valid = torch.abs(delta) <= self.subcodes_context
|
||
return torch.where(
|
||
valid,
|
||
torch.zeros([], device=device, dtype=dtype),
|
||
torch.full([], float('-inf'), device=device, dtype=dtype))
|
||
|
||
def _stage_attn_mask(self, stage: int, seq_len: int, num_heads: int,
|
||
device: torch.device, dtype: torch.dtype) -> tp.Optional[torch.Tensor]:
|
||
"""Creates a restricted attention mask given the stage (codebook index).
|
||
Args:
|
||
stage (int): The codebook index. Takes values in [0, n_q].
|
||
seq_len (int): Token sequence length.
|
||
num_heads (int): Num transformer attention heads.
|
||
device (torch.device): device of the output tensor.
|
||
dtype (torch.dtype): data type of the output tensor.
|
||
Returns:
|
||
torch.Tensor: Either a restricted attention mask or None if stage attention is unrestricted.
|
||
"""
|
||
sa_mask = None
|
||
|
||
if stage > 0 and self.subcodes_context > -1:
|
||
# parallel - non-causal - with restricted subcodes context
|
||
sa_mask = self.restricted_context_attn_mask(seq_len, device=device, dtype=dtype)
|
||
|
||
if sa_mask is not None:
|
||
# Repeat for each attention head
|
||
sa_mask = sa_mask.repeat((1, num_heads, 1, 1))
|
||
|
||
# align8 to enable memory efficient attention
|
||
MEMORY_EFFICIENT_ATTN_ALIGN_FACTOR = 8
|
||
seq_len_aligned = \
|
||
int(np.ceil(seq_len / MEMORY_EFFICIENT_ATTN_ALIGN_FACTOR)) * MEMORY_EFFICIENT_ATTN_ALIGN_FACTOR
|
||
|
||
sa_mask_aligned = torch.zeros((1, num_heads, seq_len_aligned, seq_len_aligned), device=device, dtype=dtype)
|
||
sa_mask_aligned[..., :seq_len, :seq_len] = sa_mask
|
||
sa_mask = sa_mask_aligned
|
||
|
||
return sa_mask
|
||
|
||
def _build_attn_masks(self, compression_model_framerate: int, segment_duration: int, num_heads: int,
|
||
device: torch.device, dtype: torch.dtype):
|
||
"""Construct attention mask per stage. For each of the RVQ codebook levels in the [0, n_q] range,
|
||
either a local attention map or None would be stored as an entry in the self.attn_mask_per_stage list.
|
||
Args:
|
||
compression_model_framerate (int): The frame rate of the tokenizer.
|
||
segment_duration (int): Sample length in seconds.
|
||
num_heads (int): Num transformer attention heads.
|
||
device (torch.device): device of the output tensor.
|
||
dtype (torch.dtype): data type of the output tensor.
|
||
"""
|
||
seq_len = compression_model_framerate * segment_duration
|
||
self.attn_mask_per_stage = [self._stage_attn_mask(stage, seq_len, num_heads,
|
||
device, dtype) for stage in range(self.n_q)]
|
||
|
||
@torch.no_grad()
|
||
def generate(self,
|
||
prompt: tp.Optional[torch.Tensor] = None,
|
||
conditions: tp.List[ConditioningAttributes] = [],
|
||
num_samples: tp.Optional[int] = None,
|
||
max_gen_len: int = 256,
|
||
use_sampling: bool = True,
|
||
temp: float = 1.0,
|
||
top_k: int = 250,
|
||
top_p: float = 0.0,
|
||
cfg_coef: tp.Optional[float] = None,
|
||
cfg_coef_beta: tp.Optional[float] = None,
|
||
two_step_cfg: tp.Optional[bool] = None,
|
||
remove_prompts: bool = False,
|
||
check: bool = False,
|
||
callback: tp.Optional[tp.Callable[[int, int], None]] = None,
|
||
**kwargs) -> torch.Tensor:
|
||
|
||
assert cfg_coef is None, "Unsupported in MAGNeT. Use max_cfg_coef,min_cfg_coef instead."
|
||
assert two_step_cfg is None, "MAGNeT currently doesn't support two step classifier-free-guidance."
|
||
assert remove_prompts is False, "MAGNeT currently doesn't support the remove_prompts arg."
|
||
assert check is False, "MAGNeT currently doesn't support the check arg."
|
||
assert cfg_coef_beta is None, "MAGNeT currently doesn't support the cfg_coef_beta arg."
|
||
# Call the MAGNeT-specific generation method
|
||
return self._generate_magnet(prompt=prompt,
|
||
conditions=conditions,
|
||
num_samples=num_samples,
|
||
max_gen_len=max_gen_len,
|
||
use_sampling=use_sampling,
|
||
temp=temp,
|
||
top_k=top_k,
|
||
top_p=top_p,
|
||
callback=callback, **kwargs)
|
||
|
||
@torch.no_grad()
|
||
def _generate_magnet(self,
|
||
prompt: tp.Optional[torch.Tensor] = None,
|
||
conditions: tp.List[ConditioningAttributes] = [],
|
||
num_samples: tp.Optional[int] = None,
|
||
max_gen_len: int = 256,
|
||
use_sampling: bool = True,
|
||
temp: float = 3.0,
|
||
top_k: int = 0,
|
||
top_p: float = 0.9,
|
||
callback: tp.Optional[tp.Callable[[int, int], None]] = None,
|
||
max_cfg_coef: float = 10.0,
|
||
min_cfg_coef: float = 1.0,
|
||
decoding_steps: tp.List[int] = [20, 10, 10, 10],
|
||
anneal_temp: bool = True,
|
||
span_scoring='max',
|
||
span_arrangement='nonoverlap') -> torch.Tensor:
|
||
"""Generate audio tokens given textual conditions, and optionally given audio prompts,
|
||
by running MAGNeT's iterative decoding algorithm for each of the n_q RVQ levels.
|
||
Args:
|
||
prompt (torch.Tensor): Prompt tokens of shape [B, K, T].
|
||
conditions (list of ConditioningAttributes): List of conditions.
|
||
num_samples (int): Number of samples to generate when no prompt and no conditions are given.
|
||
max_gen_len (int): Maximum generation length.
|
||
use_sampling (bool): Whether to use a sampling strategy or not.
|
||
temp (float): Initial sampling temperature.
|
||
top_k (int): k for "top-k" sampling.
|
||
top_p (float): p for "top-p" sampling.
|
||
callback (Callback): Callback function to report generation progress.
|
||
max_clsfg_coef (float): Initial coefficient used for classifier free guidance.
|
||
min_clsfg_coef (float): Final coefficient used for classifier free guidance.
|
||
decoding_steps (list of n_q ints): The number of iterative decoding steps,
|
||
for each of the n_q RVQ codebooks.
|
||
anneal_temp (bool): When set to True, softmax temperature will be linearly decayed to zero, at each stage.
|
||
span_scoring (str): Use the maximum probability of each span ('max')
|
||
or the product of probabilities ('prod').
|
||
span_arrangement (str): Use either non-overlapping spans ('nonoverlap') or overlapping spans ('stride1').
|
||
in the masking scheme.
|
||
Returns:
|
||
torch.Tensor: Generated tokens.
|
||
"""
|
||
assert not self.training, "generation shouldn't be used in training mode."
|
||
first_param = next(iter(self.parameters()))
|
||
device = first_param.device
|
||
|
||
# Checking all input shapes are consistent.
|
||
possible_num_samples = []
|
||
if num_samples is not None:
|
||
possible_num_samples.append(num_samples)
|
||
elif prompt is not None:
|
||
possible_num_samples.append(prompt.shape[0])
|
||
elif conditions:
|
||
possible_num_samples.append(len(conditions))
|
||
else:
|
||
possible_num_samples.append(1)
|
||
assert [x == possible_num_samples[0] for x in possible_num_samples], "Inconsistent inputs shapes"
|
||
num_samples = possible_num_samples[0]
|
||
|
||
# below we create set of conditions: one conditional and one unconditional
|
||
# to do that we merge the regular condition together with the null condition
|
||
# we then do 1 forward pass instead of 2.
|
||
cfg_conditions: tp.Optional[ConditionTensors]
|
||
if conditions:
|
||
null_conditions = ClassifierFreeGuidanceDropout(p=1.0)(conditions)
|
||
conditions = conditions + null_conditions
|
||
tokenized = self.condition_provider.tokenize(conditions)
|
||
cfg_conditions = self.condition_provider(tokenized)
|
||
else:
|
||
cfg_conditions = {}
|
||
|
||
if prompt is None:
|
||
assert num_samples > 0
|
||
prompt = torch.zeros((num_samples, self.num_codebooks, 0), dtype=torch.long, device=device)
|
||
|
||
B, K, prompt_length = prompt.shape
|
||
start_offset = prompt_length
|
||
assert start_offset < max_gen_len
|
||
|
||
mask_id = self.special_token_id
|
||
|
||
# we generate codes with a fixed sequence length
|
||
shape = (B, K, max_gen_len)
|
||
|
||
gen_codes = torch.full(shape, mask_id, dtype=torch.long, device=device)
|
||
# filling the gen_codes with the prompt if needed
|
||
gen_codes[..., :start_offset] = prompt
|
||
# create the gen_sequence with proper interleaving from the pattern: [B, K, S]
|
||
gen_sequence = gen_codes
|
||
|
||
curr_step = 0
|
||
for stage, n_steps in zip(range(self.n_q), decoding_steps):
|
||
gen_sequence, curr_step = self._generate_stage(gen_sequence,
|
||
cfg_conditions,
|
||
stage=stage,
|
||
device=device,
|
||
prompt_length=prompt_length,
|
||
prompt=prompt,
|
||
temp=temp,
|
||
max_cfg_coef=max_cfg_coef,
|
||
min_cfg_coef=min_cfg_coef,
|
||
top_k=top_k,
|
||
top_p=top_p,
|
||
timesteps=n_steps,
|
||
anneal_temp=anneal_temp,
|
||
span_scoring=span_scoring,
|
||
use_sampling=use_sampling,
|
||
span_arrangement=span_arrangement,
|
||
curr_step=curr_step,
|
||
total_steps=sum(decoding_steps),
|
||
callback=callback)
|
||
|
||
return gen_sequence
|
||
|
||
@torch.no_grad()
|
||
def _generate_stage(self,
|
||
gen_sequence: torch.Tensor,
|
||
condition_tensors: tp.Optional[ConditionTensors],
|
||
stage: int,
|
||
device: torch.device,
|
||
prompt_length: int = 0,
|
||
prompt: tp.Optional[torch.Tensor] = None,
|
||
use_sampling: bool = True,
|
||
temp: float = 3.0,
|
||
max_cfg_coef: float = 10.0,
|
||
min_cfg_coef: float = 1.0,
|
||
top_k: int = 0,
|
||
top_p: float = 0.0,
|
||
timesteps: int = 10,
|
||
anneal_temp: bool = True,
|
||
span_scoring: str = 'max',
|
||
span_arrangement: str = 'nonoverlap',
|
||
curr_step: int = 0,
|
||
total_steps: int = 0,
|
||
callback: tp.Optional[tp.Callable[[int, int], None]] = None) -> tp.Tuple[torch.Tensor, int]:
|
||
"""Generate audio tokens of a single RVQ level (stage), given the previously generated stages,
|
||
and the textual conditions.
|
||
Args:
|
||
gen_sequence (torch.Tensor): Previously generated tokens.
|
||
condition_tensors (tp.Optional[ConditionTensors]): pre-computed conditioning tensors.
|
||
stage (int): RVQ level to generate.
|
||
device (torch.device): device of the output tensor.
|
||
prompt_length (int): Temporal length of the audio prompt.
|
||
prompt (torch.Tensor): Prompt tokens of shape [B, K, T].
|
||
use_sampling (bool): Whether to use a sampling strategy or not.
|
||
temp (float): Initial sampling temperature.
|
||
max_clsfg_coef (float): Initial coefficient used for classifier free guidance.
|
||
min_clsfg_coef (float): Final coefficient used for classifier free guidance.
|
||
top_k (int): k for "top-k" sampling.
|
||
top_p (float): p for "top-p" sampling.
|
||
timesteps (int): Number of iterative decoding steps.
|
||
anneal_temp (bool): When set to True, softmax temperature will be linearly decayed to zero, at each stage.
|
||
span_scoring (str): Use the maximum probability of each span ('max')
|
||
or the product of probabilities ('prod').
|
||
span_arrangement (str): Use either non-overlapping spans ('nonoverlap') or overlapping spans ('stride1').
|
||
in the masking scheme.
|
||
curr_step (int): Global iterative decoding step counter.
|
||
total_steps (int): Total decoding steps.
|
||
callback (Callback): Callback function to report generation progress.
|
||
Returns:
|
||
tuple(torch.Tensor, int): Generated tokens and the current decoding step counter.
|
||
"""
|
||
B, K, T = gen_sequence.shape
|
||
shape = (B, 1, T) # generating a single codebook per stage
|
||
|
||
mask_id = self.special_token_id
|
||
stage_gen_seq = torch.full(shape, mask_id, dtype=torch.long, device=device)
|
||
|
||
assert span_arrangement == 'nonoverlap' or span_arrangement == 'stride1'
|
||
chunk_masking = self.span_len > 1 and span_arrangement == 'nonoverlap'
|
||
|
||
DONT_REMASK_ME_SCORE = -1e4
|
||
|
||
model = self if self._fsdp is None else self._fsdp
|
||
|
||
if chunk_masking:
|
||
# span-wise scores
|
||
n_chunks = T // self.span_len
|
||
if T % self.span_len != 0:
|
||
# trim sequence ending to achieve a multiple of span_len
|
||
T = self.span_len * n_chunks
|
||
gen_sequence = gen_sequence[..., :T]
|
||
stage_gen_seq = stage_gen_seq[..., :T]
|
||
|
||
chunked_shape = (B, 1, n_chunks)
|
||
n_prompt_chunks = prompt_length // self.span_len
|
||
scores = torch.zeros(chunked_shape, dtype=torch.float32, device=device)
|
||
scores[..., :n_prompt_chunks] = DONT_REMASK_ME_SCORE
|
||
num_chunks_to_gen = n_chunks - n_prompt_chunks
|
||
else:
|
||
# token-wise scores
|
||
scores = torch.zeros(shape, dtype=torch.float32, device=device)
|
||
scores[..., :prompt_length] = DONT_REMASK_ME_SCORE
|
||
gen_T = T - prompt_length
|
||
|
||
# run MAGNeT iterative decoding for "timesteps" iterations
|
||
for timestep, steps_left in zip(torch.linspace(0, 1, timesteps, device=device), reversed(range(timesteps))):
|
||
|
||
mask_p = torch.cos(timestep * math.pi * 0.5)
|
||
|
||
if chunk_masking:
|
||
num_masked = max(int((mask_p * num_chunks_to_gen).item()), 1)
|
||
else:
|
||
num_masked = max(int((mask_p * gen_T).item()), 1)
|
||
|
||
# masking
|
||
run_lps_masking = (span_arrangement == 'stride1') and self.span_len > 1
|
||
if run_lps_masking:
|
||
# masking of the k least probable overlapping (stride 1) spans
|
||
mask = torch.concat((
|
||
[self._least_probable_span_masking(scores[[i], :, :], num_masked).to(device)
|
||
for i in range(B)]), dim=0)
|
||
stage_gen_seq[mask] = mask_id
|
||
else:
|
||
# masking of the k least probable non-overlapping spans
|
||
masked = scores.topk(num_masked, dim=-1).indices
|
||
if chunk_masking:
|
||
chunks_mask = torch.full(chunked_shape, False, dtype=torch.bool, device=device)
|
||
chunks_mask = chunks_mask.scatter(2, masked, True)
|
||
mask = torch.repeat_interleave(chunks_mask, self.span_len, dim=-1)
|
||
stage_gen_seq[mask] = mask_id
|
||
else:
|
||
stage_gen_seq = stage_gen_seq.scatter(2, masked, mask_id)
|
||
|
||
if prompt is not None:
|
||
stage_gen_seq[..., :prompt_length] = prompt[:, stage, :].unsqueeze(1)
|
||
|
||
gen_sequence[:, [stage], :] = stage_gen_seq
|
||
if condition_tensors:
|
||
# duplicate input for classifier free guidance
|
||
sequence = torch.cat([gen_sequence, gen_sequence], dim=0)
|
||
|
||
all_logits = model(sequence, [], condition_tensors, stage=stage)
|
||
|
||
if condition_tensors:
|
||
# classifier free guidance with annealing
|
||
cond_logits, uncond_logits = all_logits.split(B, dim=0) # [B, K, T, card]
|
||
clsfg_coef = float(mask_p) * max_cfg_coef + (1 - float(mask_p)) * min_cfg_coef
|
||
logits = uncond_logits + (cond_logits - uncond_logits) * clsfg_coef
|
||
else:
|
||
logits = all_logits
|
||
|
||
# temperature annealing - linear
|
||
t = temp * (steps_left / timesteps) if anneal_temp else temp
|
||
|
||
# sampling
|
||
logits = logits[:, stage, :, :].unsqueeze(1)
|
||
probs = torch.softmax(logits / max(t, 1e-2), dim=-1)
|
||
if use_sampling:
|
||
if top_p > 0.0:
|
||
sampled_tokens = utils.sample_top_p(probs, p=top_p)
|
||
elif top_k > 0:
|
||
sampled_tokens = utils.sample_top_k(probs, k=top_k)
|
||
else:
|
||
sampled_tokens = utils.multinomial(probs, num_samples=1)
|
||
else:
|
||
sampled_tokens = torch.argmax(logits, dim=-1, keepdim=True)
|
||
|
||
# place mask_id token in each of the masked positions
|
||
mask = stage_gen_seq == mask_id
|
||
stage_gen_seq = torch.where(mask, sampled_tokens[..., 0], stage_gen_seq)
|
||
gen_sequence[:, [stage], :] = stage_gen_seq
|
||
|
||
# get probs of sampled tokens
|
||
sampled_probs = torch.gather(probs, 3, sampled_tokens)[..., 0]
|
||
|
||
# span scoring
|
||
if chunk_masking:
|
||
if span_scoring == 'max':
|
||
# max in linear space
|
||
scores = 1 - torch.max(sampled_probs.reshape((B, 1, n_chunks, -1)), dim=-1)[0]
|
||
elif span_scoring == 'prod':
|
||
# prod in log space
|
||
scores = torch.sum(-torch.log(sampled_probs).reshape((B, 1, n_chunks, -1)), dim=-1)
|
||
else:
|
||
raise NotImplementedError
|
||
else:
|
||
# prod in log space for lps masking (stride1)
|
||
scores = -torch.log(sampled_probs)
|
||
|
||
# Fix unmasked tokens by placing inf probs (-inf scores)
|
||
if chunk_masking:
|
||
scores = scores.masked_fill(~chunks_mask, DONT_REMASK_ME_SCORE)
|
||
else:
|
||
scores = scores.masked_fill(~mask, DONT_REMASK_ME_SCORE)
|
||
|
||
if callback is not None:
|
||
curr_step += 1
|
||
callback(curr_step, total_steps)
|
||
|
||
return gen_sequence, curr_step
|
||
|
||
def _construct_spans_mask(self, span_starts: torch.Tensor, T: int, device: torch.device) -> torch.Tensor:
|
||
"""Build a [1x1xT] boolean mask consists of overlapping spans of True values, where
|
||
span_starts defines the initial index of each span, and the span length is
|
||
defined by self.span_len.
|
||
Args:
|
||
span_starts (torch.Tensor): Boolean mask determines the temporal location of each span start.
|
||
T (int): Sequence length.
|
||
device (torch.device): device of the output tensor.
|
||
Returns:
|
||
torch.Tensor: Spans mask of shape [1x1xT]
|
||
"""
|
||
mask = torch.full((1, 1, T), False, device=device)
|
||
mask[:, :, span_starts] = True
|
||
shifted_mask = mask.clone()
|
||
for _ in range(self.span_len - 1):
|
||
shifted_mask = torch.concat((torch.full((1, 1, 1), False, device=device), shifted_mask[:, :, :-1]), dim=-1)
|
||
mask = torch.logical_or(mask, shifted_mask)
|
||
return mask
|
||
|
||
def _least_probable_span_masking(self, scores: torch.Tensor, num_masked_trg: int) -> torch.Tensor:
|
||
"""Construct a [1x1xT] boolean mask, consists of the u least probable spans,
|
||
where the token probability is determined by -scores, and the total
|
||
number of masked tokens is as closest as possible to num_masked_trg.
|
||
Find u using binary search.
|
||
Args:
|
||
scores (torch.Tensor): Per token score [-log(prob)]
|
||
num_masked_trg: int: The desired amount of tokens to be masked.
|
||
Returns:
|
||
torch.Tensor: Spans mask of shape [1x1xT]
|
||
"""
|
||
T = scores.shape[-1]
|
||
device = scores.device
|
||
scores_unfolded = scores.unfold(2, self.span_len, 1)
|
||
# Span score is the product of probs (sum in log space)
|
||
span_scores = scores_unfolded.sum(dim=-1)
|
||
spans_by_scores = torch.argsort(span_scores[0, 0], descending=True)
|
||
|
||
num_masked_trg = max(num_masked_trg, self.span_len)
|
||
|
||
# Binary search for u - the number least probable overlapping masked spans s.t.
|
||
# the total masking rate is the closest to num_masked_trg / T.
|
||
min_u = num_masked_trg // self.span_len
|
||
max_u = num_masked_trg - self.span_len + 1
|
||
mid = round(0.5 * (min_u + max_u))
|
||
|
||
if mid == min_u or mid == max_u:
|
||
return self._construct_spans_mask(spans_by_scores[:mid], T, device)
|
||
|
||
while mid > min_u and mid < max_u:
|
||
mask = self._construct_spans_mask(spans_by_scores[:mid], T, device)
|
||
n_masked = mask.sum()
|
||
if n_masked > num_masked_trg:
|
||
max_u = mid
|
||
mid = round(0.5 * (min_u + max_u))
|
||
else:
|
||
min_u = mid
|
||
mid = round(0.5 * (min_u + max_u))
|
||
|
||
return mask</code></pre>
|
||
</details>
|
||
<div class="desc"><p>Transformer-based, non-autoregressive model, operates on multiple streams of audio tokens (MAGNeT).</p>
|
||
<h2 id="args">Args</h2>
|
||
<dl>
|
||
<dt><strong><code>subcodes_context</code></strong> : <code>int</code></dt>
|
||
<dd>The number of timesteps attended in the self-attention blocks of codebooks > 0.
|
||
When set to -1, attention is unrestricted and all timesteps are attended. Defaults to 5.</dd>
|
||
<dt><strong><code>compression_model_framerate</code></strong> : <code>int</code></dt>
|
||
<dd>frame rate of the audio tokenizer.</dd>
|
||
<dt><strong><code>segment_duration</code></strong> : <code>int</code></dt>
|
||
<dd>Sample length in seconds.</dd>
|
||
<dt><strong><code>span_len</code></strong> : <code>int</code></dt>
|
||
<dd>Determines the length of masking spans. This is the minimal length of consecutive masked tokens,
|
||
for both training and inference. Defaults to 3.</dd>
|
||
<dt><strong><code>**kwargs</code></strong></dt>
|
||
<dd>Additional parameters for the LMModel.</dd>
|
||
</dl>
|
||
<p>Initializes internal Module state, shared by both nn.Module and ScriptModule.</p></div>
|
||
<h3>Ancestors</h3>
|
||
<ul class="hlist">
|
||
<li><a title="audiocraft.models.lm.LMModel" href="lm.html#audiocraft.models.lm.LMModel">LMModel</a></li>
|
||
<li><a title="audiocraft.modules.streaming.StreamingModule" href="../modules/streaming.html#audiocraft.modules.streaming.StreamingModule">StreamingModule</a></li>
|
||
<li>torch.nn.modules.module.Module</li>
|
||
</ul>
|
||
<h3>Class variables</h3>
|
||
<dl>
|
||
<dt id="audiocraft.models.lm_magnet.MagnetLMModel.call_super_init"><code class="name">var <span class="ident">call_super_init</span> : bool</code></dt>
|
||
<dd>
|
||
<div class="desc"></div>
|
||
</dd>
|
||
<dt id="audiocraft.models.lm_magnet.MagnetLMModel.dump_patches"><code class="name">var <span class="ident">dump_patches</span> : bool</code></dt>
|
||
<dd>
|
||
<div class="desc"></div>
|
||
</dd>
|
||
<dt id="audiocraft.models.lm_magnet.MagnetLMModel.training"><code class="name">var <span class="ident">training</span> : bool</code></dt>
|
||
<dd>
|
||
<div class="desc"></div>
|
||
</dd>
|
||
</dl>
|
||
<h3>Methods</h3>
|
||
<dl>
|
||
<dt id="audiocraft.models.lm_magnet.MagnetLMModel.restricted_context_attn_mask"><code class="name flex">
|
||
<span>def <span class="ident">restricted_context_attn_mask</span></span>(<span>self, seq_len: int, device: torch.device, dtype: torch.dtype) ‑> torch.Tensor</span>
|
||
</code></dt>
|
||
<dd>
|
||
<details class="source">
|
||
<summary>
|
||
<span>Expand source code</span>
|
||
</summary>
|
||
<pre><code class="python">def restricted_context_attn_mask(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
|
||
"""Creates a restricted attention mask (local attention map) where the context
|
||
is determined by self.subcodes_context.
|
||
Args:
|
||
seq_len (int): token sequence length.
|
||
device (torch.device): device of the output tensor.
|
||
dtype (torch.dtype): data type of the output tensor.
|
||
Returns:
|
||
torch.Tensor: The restricted attention mask.
|
||
"""
|
||
# Return a context restricted non-causal att mask
|
||
queries_pos = torch.arange(seq_len, device=device).view(-1, 1)
|
||
keys_pos = torch.arange(seq_len, device=device).view(1, -1)
|
||
|
||
delta = queries_pos - keys_pos
|
||
valid = torch.abs(delta) <= self.subcodes_context
|
||
return torch.where(
|
||
valid,
|
||
torch.zeros([], device=device, dtype=dtype),
|
||
torch.full([], float('-inf'), device=device, dtype=dtype))</code></pre>
|
||
</details>
|
||
<div class="desc"><p>Creates a restricted attention mask (local attention map) where the context
|
||
is determined by self.subcodes_context.</p>
|
||
<h2 id="args">Args</h2>
|
||
<dl>
|
||
<dt><strong><code>seq_len</code></strong> : <code>int</code></dt>
|
||
<dd>token sequence length.</dd>
|
||
<dt><strong><code>device</code></strong> : <code>torch.device</code></dt>
|
||
<dd>device of the output tensor.</dd>
|
||
<dt><strong><code>dtype</code></strong> : <code>torch.dtype</code></dt>
|
||
<dd>data type of the output tensor.</dd>
|
||
</dl>
|
||
<h2 id="returns">Returns</h2>
|
||
<dl>
|
||
<dt><code>torch.Tensor</code></dt>
|
||
<dd>The restricted attention mask.</dd>
|
||
</dl></div>
|
||
</dd>
|
||
</dl>
|
||
<h3>Inherited members</h3>
|
||
<ul class="hlist">
|
||
<li><code><b><a title="audiocraft.models.lm.LMModel" href="lm.html#audiocraft.models.lm.LMModel">LMModel</a></b></code>:
|
||
<ul class="hlist">
|
||
<li><code><a title="audiocraft.models.lm.LMModel.compute_predictions" href="lm.html#audiocraft.models.lm.LMModel.compute_predictions">compute_predictions</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm.LMModel.flush" href="../modules/streaming.html#audiocraft.modules.streaming.StreamingModule.flush">flush</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm.LMModel.forward" href="lm.html#audiocraft.models.lm.LMModel.forward">forward</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm.LMModel.generate" href="lm.html#audiocraft.models.lm.LMModel.generate">generate</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm.LMModel.get_streaming_state" href="../modules/streaming.html#audiocraft.modules.streaming.StreamingModule.get_streaming_state">get_streaming_state</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm.LMModel.reset_streaming" href="../modules/streaming.html#audiocraft.modules.streaming.StreamingModule.reset_streaming">reset_streaming</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm.LMModel.set_streaming_state" href="../modules/streaming.html#audiocraft.modules.streaming.StreamingModule.set_streaming_state">set_streaming_state</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm.LMModel.streaming" href="../modules/streaming.html#audiocraft.modules.streaming.StreamingModule.streaming">streaming</a></code></li>
|
||
</ul>
|
||
</li>
|
||
</ul>
|
||
</dd>
|
||
</dl>
|
||
</section>
|
||
</article>
|
||
<nav id="sidebar">
|
||
<div class="toc">
|
||
<ul></ul>
|
||
</div>
|
||
<ul id="index">
|
||
<li><h3>Super-module</h3>
|
||
<ul>
|
||
<li><code><a title="audiocraft.models" href="index.html">audiocraft.models</a></code></li>
|
||
</ul>
|
||
</li>
|
||
<li><h3><a href="#header-classes">Classes</a></h3>
|
||
<ul>
|
||
<li>
|
||
<h4><code><a title="audiocraft.models.lm_magnet.MagnetLMModel" href="#audiocraft.models.lm_magnet.MagnetLMModel">MagnetLMModel</a></code></h4>
|
||
<ul class="">
|
||
<li><code><a title="audiocraft.models.lm_magnet.MagnetLMModel.call_super_init" href="#audiocraft.models.lm_magnet.MagnetLMModel.call_super_init">call_super_init</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm_magnet.MagnetLMModel.dump_patches" href="#audiocraft.models.lm_magnet.MagnetLMModel.dump_patches">dump_patches</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm_magnet.MagnetLMModel.restricted_context_attn_mask" href="#audiocraft.models.lm_magnet.MagnetLMModel.restricted_context_attn_mask">restricted_context_attn_mask</a></code></li>
|
||
<li><code><a title="audiocraft.models.lm_magnet.MagnetLMModel.training" href="#audiocraft.models.lm_magnet.MagnetLMModel.training">training</a></code></li>
|
||
</ul>
|
||
</li>
|
||
</ul>
|
||
</li>
|
||
</ul>
|
||
</nav>
|
||
</main>
|
||
<footer id="footer">
|
||
<p>Generated by <a href="https://pdoc3.github.io/pdoc" title="pdoc: Python API documentation generator"><cite>pdoc</cite> 0.11.5</a>.</p>
|
||
</footer>
|
||
</body>
|
||
</html>
|