mirrors / facebookresearch--audiocraft

项目文件夹

facebookresearch--audiocraft
文件
gh-docs
T
Alexandre Défossez (autodoc) aeffa9f270 api_docs
2025-03-13 16:07:03 +00:00

471 行
33 KiB
HTML
原始文件 永久链接 Blame 文件历史

此文件含有不可见的 Unicode 字符
此文件含有人类无法区分的不可见的 Unicode 字符，但可以由计算机进行不同的处理。 如果您是想特意这样的，可以安全地忽略该警告。 使用 Escape 按钮显示他们。
<!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.solvers.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.solvers.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.solvers.magnet.AudioMagnetSolver"><code class="flex name class">
<span>class <span class="ident">AudioMagnetSolver</span></span>
<span>(</span><span>cfg: omegaconf.dictconfig.DictConfig)</span>
</code></dt>
<dd>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">class AudioMagnetSolver(MagnetSolver):
&#34;&#34;&#34;Solver for audio-MAGNeT. A MAGNeT model for sound generation.
More information can be found in the MAGNeT model card.
&#34;&#34;&#34;
DATASET_TYPE: builders.DatasetType = builders.DatasetType.SOUND</code></pre>
</details>
<div class="desc"><p>Solver for audio-MAGNeT. A MAGNeT model for sound generation.</p>
<p>More information can be found in the MAGNeT model card.</p></div>
<h3>Ancestors</h3>
<ul class="hlist">
<li><a title="audiocraft.solvers.magnet.MagnetSolver" href="#audiocraft.solvers.magnet.MagnetSolver">MagnetSolver</a></li>
<li><a title="audiocraft.solvers.musicgen.MusicGenSolver" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver">MusicGenSolver</a></li>
<li><a title="audiocraft.solvers.base.StandardSolver" href="base.html#audiocraft.solvers.base.StandardSolver">StandardSolver</a></li>
<li>abc.ABC</li>
<li>flashy.solver.BaseSolver</li>
</ul>
<h3>Class variables</h3>
<dl>
<dt id="audiocraft.solvers.magnet.AudioMagnetSolver.DATASET_TYPE"><code class="name">var <span class="ident">DATASET_TYPE</span> : <a title="audiocraft.solvers.builders.DatasetType" href="builders.html#audiocraft.solvers.builders.DatasetType">DatasetType</a></code></dt>
<dd>
<div class="desc"></div>
</dd>
</dl>
<h3>Inherited members</h3>
<ul class="hlist">
<li><code><b><a title="audiocraft.solvers.magnet.MagnetSolver" href="#audiocraft.solvers.magnet.MagnetSolver">MagnetSolver</a></b></code>:
<ul class="hlist">
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.autocast" href="base.html#audiocraft.solvers.base.StandardSolver.autocast">autocast</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.best_metric_name" href="base.html#audiocraft.solvers.base.StandardSolver.best_metric_name">best_metric_name</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.build_dataloaders" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.build_dataloaders">build_dataloaders</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.build_model" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.build_model">build_model</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.commit" href="base.html#audiocraft.solvers.base.StandardSolver.commit">commit</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.common_train_valid" href="base.html#audiocraft.solvers.base.StandardSolver.common_train_valid">common_train_valid</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.evaluate" href="base.html#audiocraft.solvers.base.StandardSolver.evaluate">evaluate</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.evaluate_audio_generation" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.evaluate_audio_generation">evaluate_audio_generation</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.generate" href="base.html#audiocraft.solvers.base.StandardSolver.generate">generate</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.generate_audio" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.generate_audio">generate_audio</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.get_eval_solver_from_sig" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.get_eval_solver_from_sig">get_eval_solver_from_sig</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.initialize_ema" href="base.html#audiocraft.solvers.base.StandardSolver.initialize_ema">initialize_ema</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.load_checkpoints" href="base.html#audiocraft.solvers.base.StandardSolver.load_checkpoints">load_checkpoints</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.log_model_summary" href="base.html#audiocraft.solvers.base.StandardSolver.log_model_summary">log_model_summary</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.register_best_state" href="base.html#audiocraft.solvers.base.StandardSolver.register_best_state">register_best_state</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.register_ema" href="base.html#audiocraft.solvers.base.StandardSolver.register_ema">register_ema</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.restore" href="base.html#audiocraft.solvers.base.StandardSolver.restore">restore</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.run" href="base.html#audiocraft.solvers.base.StandardSolver.run">run</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.run_epoch" href="base.html#audiocraft.solvers.base.StandardSolver.run_epoch">run_epoch</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.run_generate_step" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.run_generate_step">run_generate_step</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.run_one_stage" href="base.html#audiocraft.solvers.base.StandardSolver.run_one_stage">run_one_stage</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.run_step" href="base.html#audiocraft.solvers.base.StandardSolver.run_step">run_step</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.save_checkpoints" href="base.html#audiocraft.solvers.base.StandardSolver.save_checkpoints">save_checkpoints</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.should_run_stage" href="base.html#audiocraft.solvers.base.StandardSolver.should_run_stage">should_run_stage</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.should_stop_training" href="base.html#audiocraft.solvers.base.StandardSolver.should_stop_training">should_stop_training</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.show" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.show">show</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.train" href="base.html#audiocraft.solvers.base.StandardSolver.train">train</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.update_best_state_from_stage" href="base.html#audiocraft.solvers.base.StandardSolver.update_best_state_from_stage">update_best_state_from_stage</a></code></li>
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.valid" href="base.html#audiocraft.solvers.base.StandardSolver.valid">valid</a></code></li>
</ul>
</li>
</ul>
</dd>
<dt id="audiocraft.solvers.magnet.MagnetSolver"><code class="flex name class">
<span>class <span class="ident">MagnetSolver</span></span>
<span>(</span><span>cfg: omegaconf.dictconfig.DictConfig)</span>
</code></dt>
<dd>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">class MagnetSolver(musicgen.MusicGenSolver):
&#34;&#34;&#34;Solver for MAGNeT - Masked Audio Generation using
a single Non-autoregressive Transformer https://arxiv.org/abs/2401.04577.
&#34;&#34;&#34;
def __init__(self, cfg: DictConfig):
super().__init__(cfg)
# initialize generation parameters by config
self.generation_params = {
&#39;use_sampling&#39;: self.cfg.generate.lm.use_sampling,
&#39;temp&#39;: self.cfg.generate.lm.temp,
&#39;top_k&#39;: self.cfg.generate.lm.top_k,
&#39;top_p&#39;: self.cfg.generate.lm.top_p,
&#39;max_cfg_coef&#39;: self.cfg.generate.lm.max_cfg_coef,
&#39;min_cfg_coef&#39;: self.cfg.generate.lm.min_cfg_coef,
&#39;decoding_steps&#39;: list(self.cfg.generate.lm.decoding_steps),
&#39;anneal_temp&#39;: self.cfg.generate.lm.anneal_temp,
&#39;span_scoring&#39;: self.cfg.generate.lm.span_scoring,
&#39;span_arrangement&#39;: self.cfg.generate.lm.span_arrangement
}
sequence_len = int(cfg.dataset.segment_duration * self.compression_model.frame_rate)
self.mean_maskrate_to_u = torch.tensor(self._calc_mean_maskrate_to_u_LUT(sequence_len), device=self.device)
self.ce_per_codebook = [torch.log(torch.tensor(self.compression_model.cardinality, device=self.device))
for _ in range(cfg.transformer_lm.n_q)]
def build_model(self) -&gt; None:
self.cfg.transformer_lm.segment_duration = self.cfg.dataset.segment_duration
self.cfg.transformer_lm.span_len = self.cfg.masking.span_len
assert self.cfg.efficient_attention_backend == &#34;xformers&#34;, &#34;MAGNeT v1 models support only xformers backend.&#34;
super().build_model()
def _calc_mean_maskrate_to_u_LUT(self, T: int):
&#34;&#34;&#34; Create a Look Up Table (LUT) transforming a discrete masking percentage m in 0,1,...,100 to u,
the number of overlapping spans of length L to place s.t. the masking rate is approximately m/float(100).
It first creates the inverse transformation, of the masking rate as function of u,
using the expression choose(T - L, u) / choose(T, u), where L is the atomic span length used
during masking. See https://arxiv.org/abs/2401.04577,
appendix C, for the mean mask rate derivation.
We leverage the fact that:
choose(T - L, u) / choose(T, u) = Prod_{j = 0}^{u - 1}((T - L - j)/(T - j))
in the provided implementation, in order to avoid overflow.
Args:
T (float): Sequence length.
Returns:
(List) A LUT transforming m in 0,1,...,100 to u,
s.t. the masking rate of the span-L mask is approximately m/float(100).
&#34;&#34;&#34;
L = self.cfg.masking.span_len
u2mean = [0.0] # mean mask rate is 0.0 for u = 0
v = (T - L) / float(T)
for u in range(1, T):
u2mean.append(1 - v)
v *= (T - L - u) / (T - u) # Overflow-safe implementation of choose(T - L, u) / choose(T, u).
mean2u = []
for maskperc in range(101):
maskrate = maskperc / float(100)
u = int(np.searchsorted(u2mean, maskrate))
mean2u.append(u)
return mean2u
def _non_spans_mask(self, mask_probs: torch.Tensor, B: int, T: int, device: torch.device) -&gt; torch.Tensor:
&#34;&#34;&#34; Construct a boolean mask of shape [B, T, 1], with masking rates defined by mask_probs.
The masked tokens are singletons, placed uniformly at random.
Args:
mask_probs (torch.Tensor): The desired masking rate per sample, of shape [B,]
B (int): Batch size.
T (int): Sequence length.
device (torch.device): device of the output tensor
Returns:
(torch.Tensor): A mask of shape [B, T]
&#34;&#34;&#34;
num_token_masked = (T * mask_probs).round().clamp(min=1)
batch_randperm = torch.rand((B, T), device=device).argsort(dim=-1)
return batch_randperm &lt; rearrange(num_token_masked, &#39;b -&gt; b 1&#39;)
def _spans_mask(self, mask_probs: torch.Tensor, B: int, T: int, device: torch.device) -&gt; torch.Tensor:
&#34;&#34;&#34; Construct a spans mask with masking rates defined by mask_probs,
where the atomic span length ( &gt; 1 ) is defined by cfg.masking.span_len.
Args:
mask_probs (torch.Tensor): The desired masking rate per sample, of shape [B,]
B (int): Batch size.
T (int): Sequence length.
device (torch.device): device of the output tensor
Returns:
(torch.Tensor): A spans mask of shape [B, T]
&#34;&#34;&#34;
rounded_probs = torch.round(100 * mask_probs).long()
k = self.mean_maskrate_to_u[rounded_probs].clamp(min=1) # k is the number of span starts
# sample random span starts
batch_randperm = torch.rand((B, T), device=device).argsort(dim=-1)
mask = batch_randperm &lt; rearrange(k, &#39;b -&gt; b 1&#39;)
B, T = mask.shape
shifted_mask = mask.clone()
for _ in range(self.cfg.masking.span_len - 1):
shifted_mask = torch.concat((torch.full((B, 1), False, device=device), shifted_mask[:, :-1]), dim=1)
mask = torch.logical_or(mask, shifted_mask)
return mask
def _get_mask(self, mask_probs: torch.Tensor, B: int, T: int, device: torch.device) -&gt; torch.Tensor:
&#34;&#34;&#34; Construct a boolean mask with masking rates defined by mask_probs, and atomic
span length defined by cfg.masking.span_len.
Args:
mask_probs (torch.Tensor): The desired masking rate per sample, of shape [B,]
B (int): Batch size.
T (int): Sequence length.
device (torch.device): device of the output tensor
Returns:
(torch.Tensor): A boolean tensor of shape [B, T]
&#34;&#34;&#34;
if self.cfg.masking.span_len &lt;= 1:
return self._non_spans_mask(mask_probs, B, T, device)
return self._spans_mask(mask_probs, B, T, device)
def _compute_cross_entropy_magnet(self, logits: torch.Tensor,
targets: torch.Tensor, mask: torch.Tensor, stage: torch.Tensor) -&gt; torch.Tensor:
&#34;&#34;&#34; Compute cross entropy between multi-codebook targets and model&#39;s logits.
The cross entropy is computed only on a specific codebook, defined by the stage argument.
Valid timesteps for each codebook are pulled from the mask, where invalid
timesteps are set to 0.
Args:
logits (torch.Tensor): Model&#39;s logits of shape [B, K, T, card].
targets (torch.Tensor): Target codes, of shape [B, K, T].
mask (torch.Tensor): Mask for valid target codes, of shape [B, K, T].
stage (torch.Tensor): The codebook (idx) that is being optimized, as a scalar tensor.
Returns:
ce (torch.Tensor): Cross entropy of the codebook that is being optimized.
&#34;&#34;&#34;
assert logits.shape[:-1] == targets.shape
assert mask.shape == targets.shape
ce = torch.zeros([], device=targets.device)
logits_k = logits[:, stage, ...].contiguous().view(-1, logits.size(-1)) # [B x T, card]
targets_k = targets[:, stage, ...].contiguous().view(-1) # [B x T]
mask_k = mask[:, stage, ...].contiguous().view(-1) # [B x T]
IGNORE_IDX = -1
targets_k[~mask_k] = IGNORE_IDX
q_ce = F.cross_entropy(logits_k, targets_k, ignore_index=IGNORE_IDX)
ce += q_ce
return ce
def run_step(self, idx: int, batch: tp.Tuple[torch.Tensor, tp.List[SegmentWithAttributes]], metrics: dict) -&gt; dict:
&#34;&#34;&#34;Perform one training or valid step on a given batch.&#34;&#34;&#34;
check_synchronization_points = idx == 1 and self.device == &#39;cuda&#39;
condition_tensors, audio_tokens, padding_mask = self._prepare_tokens_and_attributes(
batch, check_synchronization_points)
self.deadlock_detect.update(&#39;tokens_and_conditions&#39;)
if check_synchronization_points:
torch.cuda.set_sync_debug_mode(&#39;warn&#39;)
B, K, T = audio_tokens.shape
device = self.device
# Choose the stage (codebook idx) for update, uniformly at random.
stage_ = random.randint(0, K - 1)
stage = torch.full((1, ), stage_, device=device)
# masking
rand_time = torch.zeros((B,), device=device).float().uniform_(0, 1)
rand_mask_probs = torch.cos(rand_time * math.pi * 0.5)
# stage mask
stage_mask = self._get_mask(rand_mask_probs, B, T, device) # [B, T]
stage_mask = stage_mask.unsqueeze(1) # [B, 1, T]
# Keep all preceding codebooks.
mask = torch.full((B, K, T), False, device=device)
mask[:, stage, :] = stage_mask
# Mask all codebooks larger than stage_
mask_id = self.model.special_token_id
mask[:, (stage_+1):, :] = torch.full((B, K - stage_ - 1, T), True, device=device)
input_tokens = torch.where(mask, mask_id, audio_tokens)
# Take loss only on the chosen stage, and only on the masked tokens.
loss_mask = torch.full((B, K, T), False, device=device)
loss_mask[:, stage, :] = stage_mask
with self.autocast:
model_output = self.model.compute_predictions(input_tokens, [], condition_tensors, stage=stage_)
logits = model_output.logits
loss_mask &amp;= padding_mask
ce = self._compute_cross_entropy_magnet(logits, audio_tokens, loss_mask, stage)
loss = ce
self.deadlock_detect.update(&#39;loss&#39;)
if check_synchronization_points:
torch.cuda.set_sync_debug_mode(&#39;default&#39;)
if self.is_training:
metrics[&#39;lr&#39;] = self.optimizer.param_groups[0][&#39;lr&#39;]
if self.scaler is not None:
loss = self.scaler.scale(loss)
self.deadlock_detect.update(&#39;scale&#39;)
if self.cfg.fsdp.use:
loss.backward()
flashy.distrib.average_tensors(self.model.buffers())
elif self.cfg.optim.eager_sync:
with flashy.distrib.eager_sync_model(self.model):
loss.backward()
else:
# this should always be slower but can be useful
# for weird use cases like multiple backwards.
loss.backward()
flashy.distrib.sync_model(self.model)
self.deadlock_detect.update(&#39;backward&#39;)
if self.scaler is not None:
self.scaler.unscale_(self.optimizer)
if self.cfg.optim.max_norm:
if self.cfg.fsdp.use:
metrics[&#39;grad_norm&#39;] = self.model.clip_grad_norm_(self.cfg.optim.max_norm) # type: ignore
else:
metrics[&#39;grad_norm&#39;] = torch.nn.utils.clip_grad_norm_(
self.model.parameters(), self.cfg.optim.max_norm
)
if self.scaler is None:
self.optimizer.step()
else:
self.scaler.step(self.optimizer)
self.scaler.update()
if self.lr_scheduler:
self.lr_scheduler.step()
self.optimizer.zero_grad()
self.deadlock_detect.update(&#39;optim&#39;)
if self.scaler is not None:
scale = self.scaler.get_scale()
metrics[&#39;grad_scale&#39;] = scale
if not loss.isfinite().all():
raise RuntimeError(&#34;Model probably diverged.&#34;)
metrics[&#39;ce&#39;] = ce
metrics[&#39;ppl&#39;] = torch.exp(ce)
return metrics</code></pre>
</details>
<div class="desc"><p>Solver for MAGNeT - Masked Audio Generation using
a single Non-autoregressive Transformer <a href="https://arxiv.org/abs/2401.04577.">https://arxiv.org/abs/2401.04577.</a></p></div>
<h3>Ancestors</h3>
<ul class="hlist">
<li><a title="audiocraft.solvers.musicgen.MusicGenSolver" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver">MusicGenSolver</a></li>
<li><a title="audiocraft.solvers.base.StandardSolver" href="base.html#audiocraft.solvers.base.StandardSolver">StandardSolver</a></li>
<li>abc.ABC</li>
<li>flashy.solver.BaseSolver</li>
</ul>
<h3>Subclasses</h3>
<ul class="hlist">
<li><a title="audiocraft.solvers.magnet.AudioMagnetSolver" href="#audiocraft.solvers.magnet.AudioMagnetSolver">AudioMagnetSolver</a></li>
</ul>
<h3>Class variables</h3>
<dl>
<dt id="audiocraft.solvers.magnet.MagnetSolver.DATASET_TYPE"><code class="name">var <span class="ident">DATASET_TYPE</span> : <a title="audiocraft.solvers.builders.DatasetType" href="builders.html#audiocraft.solvers.builders.DatasetType">DatasetType</a></code></dt>
<dd>
<div class="desc"></div>
</dd>
</dl>
<h3>Inherited members</h3>
<ul class="hlist">
<li><code><b><a title="audiocraft.solvers.musicgen.MusicGenSolver" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver">MusicGenSolver</a></b></code>:
<ul class="hlist">
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.autocast" href="base.html#audiocraft.solvers.base.StandardSolver.autocast">autocast</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.best_metric_name" href="base.html#audiocraft.solvers.base.StandardSolver.best_metric_name">best_metric_name</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.build_dataloaders" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.build_dataloaders">build_dataloaders</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.build_model" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.build_model">build_model</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.commit" href="base.html#audiocraft.solvers.base.StandardSolver.commit">commit</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.common_train_valid" href="base.html#audiocraft.solvers.base.StandardSolver.common_train_valid">common_train_valid</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.evaluate" href="base.html#audiocraft.solvers.base.StandardSolver.evaluate">evaluate</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.evaluate_audio_generation" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.evaluate_audio_generation">evaluate_audio_generation</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.generate" href="base.html#audiocraft.solvers.base.StandardSolver.generate">generate</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.generate_audio" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.generate_audio">generate_audio</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.get_eval_solver_from_sig" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.get_eval_solver_from_sig">get_eval_solver_from_sig</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.initialize_ema" href="base.html#audiocraft.solvers.base.StandardSolver.initialize_ema">initialize_ema</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.load_checkpoints" href="base.html#audiocraft.solvers.base.StandardSolver.load_checkpoints">load_checkpoints</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.log_model_summary" href="base.html#audiocraft.solvers.base.StandardSolver.log_model_summary">log_model_summary</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.register_best_state" href="base.html#audiocraft.solvers.base.StandardSolver.register_best_state">register_best_state</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.register_ema" href="base.html#audiocraft.solvers.base.StandardSolver.register_ema">register_ema</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.restore" href="base.html#audiocraft.solvers.base.StandardSolver.restore">restore</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.run" href="base.html#audiocraft.solvers.base.StandardSolver.run">run</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.run_epoch" href="base.html#audiocraft.solvers.base.StandardSolver.run_epoch">run_epoch</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.run_generate_step" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.run_generate_step">run_generate_step</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.run_one_stage" href="base.html#audiocraft.solvers.base.StandardSolver.run_one_stage">run_one_stage</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.run_step" href="base.html#audiocraft.solvers.base.StandardSolver.run_step">run_step</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.save_checkpoints" href="base.html#audiocraft.solvers.base.StandardSolver.save_checkpoints">save_checkpoints</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.should_run_stage" href="base.html#audiocraft.solvers.base.StandardSolver.should_run_stage">should_run_stage</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.should_stop_training" href="base.html#audiocraft.solvers.base.StandardSolver.should_stop_training">should_stop_training</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.show" href="musicgen.html#audiocraft.solvers.musicgen.MusicGenSolver.show">show</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.train" href="base.html#audiocraft.solvers.base.StandardSolver.train">train</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.update_best_state_from_stage" href="base.html#audiocraft.solvers.base.StandardSolver.update_best_state_from_stage">update_best_state_from_stage</a></code></li>
<li><code><a title="audiocraft.solvers.musicgen.MusicGenSolver.valid" href="base.html#audiocraft.solvers.base.StandardSolver.valid">valid</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.solvers" href="index.html">audiocraft.solvers</a></code></li>
</ul>
</li>
<li><h3><a href="#header-classes">Classes</a></h3>
<ul>
<li>
<h4><code><a title="audiocraft.solvers.magnet.AudioMagnetSolver" href="#audiocraft.solvers.magnet.AudioMagnetSolver">AudioMagnetSolver</a></code></h4>
<ul class="">
<li><code><a title="audiocraft.solvers.magnet.AudioMagnetSolver.DATASET_TYPE" href="#audiocraft.solvers.magnet.AudioMagnetSolver.DATASET_TYPE">DATASET_TYPE</a></code></li>
</ul>
</li>
<li>
<h4><code><a title="audiocraft.solvers.magnet.MagnetSolver" href="#audiocraft.solvers.magnet.MagnetSolver">MagnetSolver</a></code></h4>
<ul class="">
<li><code><a title="audiocraft.solvers.magnet.MagnetSolver.DATASET_TYPE" href="#audiocraft.solvers.magnet.MagnetSolver.DATASET_TYPE">DATASET_TYPE</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>
在新工单中引用 查看 Git Blame 复制永久链接