facebookresearch--audiocraft
Alexandre Défossez (autodoc)
584 行
26 KiB
HTML
584 行
26 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.optim.dadam 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.optim.dadam</code></h1>
|
||
</header>
|
||
<section id="section-intro">
|
||
</section>
|
||
<section>
|
||
</section>
|
||
<section>
|
||
</section>
|
||
<section>
|
||
<h2 class="section-title" id="header-functions">Functions</h2>
|
||
<dl>
|
||
<dt id="audiocraft.optim.dadam.to_real"><code class="name flex">
|
||
<span>def <span class="ident">to_real</span></span>(<span>x)</span>
|
||
</code></dt>
|
||
<dd>
|
||
<details class="source">
|
||
<summary>
|
||
<span>Expand source code</span>
|
||
</summary>
|
||
<pre><code class="python">def to_real(x):
|
||
if torch.is_complex(x):
|
||
return x.real
|
||
else:
|
||
return x</code></pre>
|
||
</details>
|
||
<div class="desc"></div>
|
||
</dd>
|
||
</dl>
|
||
</section>
|
||
<section>
|
||
<h2 class="section-title" id="header-classes">Classes</h2>
|
||
<dl>
|
||
<dt id="audiocraft.optim.dadam.DAdaptAdam"><code class="flex name class">
|
||
<span>class <span class="ident">DAdaptAdam</span></span>
|
||
<span>(</span><span>params,<br>lr=1.0,<br>betas=(0.9, 0.999),<br>eps=1e-08,<br>weight_decay=0,<br>log_every=0,<br>decouple=True,<br>d0=1e-06,<br>growth_rate=inf)</span>
|
||
</code></dt>
|
||
<dd>
|
||
<details class="source">
|
||
<summary>
|
||
<span>Expand source code</span>
|
||
</summary>
|
||
<pre><code class="python">class DAdaptAdam(torch.optim.Optimizer):
|
||
"""Adam with D-Adaptation automatic step-sizes.
|
||
Leave LR set to 1 unless you encounter instability.
|
||
|
||
Args:
|
||
params (iterable):
|
||
Iterable of parameters to optimize or dicts defining parameter groups.
|
||
lr (float):
|
||
Learning rate adjustment parameter. Increases or decreases the D-adapted learning rate.
|
||
betas (tuple[float, float], optional): coefficients used for computing
|
||
running averages of gradient and its square (default: (0.9, 0.999))
|
||
momentum (float):
|
||
Momentum value in the range [0,1) (default: 0.9).
|
||
eps (float):
|
||
Term added to the denominator outside of the root operation to improve numerical stability. (default: 1e-8).
|
||
weight_decay (float):
|
||
Weight decay, i.e. a L2 penalty (default: 0).
|
||
log_every (int):
|
||
Log using print every k steps, default 0 (no logging).
|
||
decouple (boolean):
|
||
Use AdamW style decoupled weight decay
|
||
d0 (float):
|
||
Initial D estimate for D-adaptation (default 1e-6). Rarely needs changing.
|
||
growth_rate (float):
|
||
prevent the D estimate from growing faster than this multiplicative rate.
|
||
Default is inf, for unrestricted. Values like 1.02 give a kind of learning
|
||
rate warmup effect.
|
||
fsdp_in_use (bool):
|
||
If you're using sharded parameters, this should be set to True. The optimizer
|
||
will attempt to auto-detect this, but if you're using an implementation other
|
||
than PyTorch's builtin version, the auto-detection won't work.
|
||
"""
|
||
def __init__(self, params, lr=1.0,
|
||
betas=(0.9, 0.999),
|
||
eps=1e-8,
|
||
weight_decay=0,
|
||
log_every=0,
|
||
decouple=True,
|
||
d0=1e-6,
|
||
growth_rate=float('inf')):
|
||
if not 0.0 < d0:
|
||
raise ValueError("Invalid d0 value: {}".format(d0))
|
||
if not 0.0 < lr:
|
||
raise ValueError("Invalid learning rate: {}".format(lr))
|
||
if not 0.0 < eps:
|
||
raise ValueError("Invalid epsilon value: {}".format(eps))
|
||
if not 0.0 <= betas[0] < 1.0:
|
||
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
|
||
if not 0.0 <= betas[1] < 1.0:
|
||
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
|
||
|
||
if decouple:
|
||
logger.info("Using decoupled weight decay")
|
||
|
||
from .fsdp import is_fsdp_used
|
||
fsdp_in_use = is_fsdp_used()
|
||
defaults = dict(lr=lr, betas=betas, eps=eps,
|
||
weight_decay=weight_decay,
|
||
d=d0,
|
||
k=0,
|
||
gsq_weighted=0.0,
|
||
log_every=log_every,
|
||
decouple=decouple,
|
||
growth_rate=growth_rate,
|
||
fsdp_in_use=fsdp_in_use)
|
||
|
||
super().__init__(params, defaults)
|
||
|
||
@property
|
||
def supports_memory_efficient_fp16(self):
|
||
return False
|
||
|
||
@property
|
||
def supports_flat_params(self):
|
||
return True
|
||
|
||
def step(self, closure=None):
|
||
"""Performs a single optimization step.
|
||
|
||
Args:
|
||
closure (callable, optional): A closure that reevaluates the model
|
||
and returns the loss.
|
||
"""
|
||
loss = None
|
||
if closure is not None:
|
||
loss = closure()
|
||
|
||
g_sq = 0.0
|
||
sksq_weighted = 0.0
|
||
sk_l1 = 0.0
|
||
|
||
lr = max(group['lr'] for group in self.param_groups)
|
||
|
||
group = self.param_groups[0]
|
||
gsq_weighted = group['gsq_weighted']
|
||
d = group['d']
|
||
dlr = d*lr
|
||
|
||
growth_rate = group['growth_rate']
|
||
decouple = group['decouple']
|
||
fsdp_in_use = group['fsdp_in_use']
|
||
log_every = group['log_every']
|
||
|
||
beta1, beta2 = group['betas']
|
||
|
||
for group in self.param_groups:
|
||
group_lr = group['lr']
|
||
decay = group['weight_decay']
|
||
k = group['k']
|
||
eps = group['eps']
|
||
|
||
if group_lr not in [lr, 0.0]:
|
||
raise RuntimeError("Setting different lr values in different parameter "
|
||
"groups is only supported for values of 0")
|
||
|
||
for p in group['params']:
|
||
if p.grad is None:
|
||
continue
|
||
if hasattr(p, "_fsdp_flattened"):
|
||
fsdp_in_use = True
|
||
grad = p.grad.data
|
||
|
||
# Apply weight decay (coupled variant)
|
||
if decay != 0 and not decouple:
|
||
grad.add_(p.data, alpha=decay)
|
||
|
||
state = self.state[p]
|
||
|
||
# State initialization
|
||
if 'step' not in state:
|
||
state['step'] = 0
|
||
state['s'] = torch.zeros_like(p.data, memory_format=torch.preserve_format).detach()
|
||
# Exponential moving average of gradient values
|
||
state['exp_avg'] = torch.zeros_like(p.data, memory_format=torch.preserve_format).detach()
|
||
# Exponential moving average of squared gradient values
|
||
state['exp_avg_sq'] = torch.zeros_like(
|
||
to_real(p.data), memory_format=torch.preserve_format).detach()
|
||
|
||
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
|
||
|
||
grad_grad = to_real(grad * grad.conj())
|
||
|
||
# Adam EMA updates
|
||
if group_lr > 0:
|
||
exp_avg.mul_(beta1).add_(grad, alpha=dlr*(1-beta1))
|
||
exp_avg_sq.mul_(beta2).add_(grad_grad, alpha=1-beta2)
|
||
|
||
denom = exp_avg_sq.sqrt().add_(eps)
|
||
|
||
g_sq += grad_grad.div_(denom).sum().item()
|
||
|
||
s = state['s']
|
||
s.mul_(beta2).add_(grad, alpha=dlr*(1-beta2))
|
||
sksq_weighted += to_real(s * s.conj()).div_(denom).sum().item()
|
||
sk_l1 += s.abs().sum().item()
|
||
|
||
######
|
||
|
||
gsq_weighted = beta2*gsq_weighted + g_sq*(dlr**2)*(1-beta2)
|
||
d_hat = d
|
||
|
||
# if we have not done any progres, return
|
||
# if we have any gradients available, will have sk_l1 > 0 (unless \|g\|=0)
|
||
if sk_l1 == 0:
|
||
return loss
|
||
|
||
if lr > 0.0:
|
||
if fsdp_in_use:
|
||
dist_tensor = torch.zeros(3, device='cuda')
|
||
dist_tensor[0] = sksq_weighted
|
||
dist_tensor[1] = gsq_weighted
|
||
dist_tensor[2] = sk_l1
|
||
dist.all_reduce(dist_tensor, op=dist.ReduceOp.SUM)
|
||
global_sksq_weighted = dist_tensor[0]
|
||
global_gsq_weighted = dist_tensor[1]
|
||
global_sk_l1 = dist_tensor[2]
|
||
else:
|
||
global_sksq_weighted = sksq_weighted
|
||
global_gsq_weighted = gsq_weighted
|
||
global_sk_l1 = sk_l1
|
||
|
||
d_hat = (global_sksq_weighted/(1-beta2) - global_gsq_weighted)/global_sk_l1
|
||
d = max(d, min(d_hat, d*growth_rate))
|
||
|
||
if log_every > 0 and k % log_every == 0:
|
||
logger.info(
|
||
f"(k={k}) dlr: {dlr:1.1e} d_hat: {d_hat:1.1e}, d: {d:1.8}. "
|
||
f"sksq_weighted={global_sksq_weighted:1.1e} gsq_weighted={global_gsq_weighted:1.1e} "
|
||
f"sk_l1={global_sk_l1:1.1e}{' (FSDP)' if fsdp_in_use else ''}")
|
||
|
||
for group in self.param_groups:
|
||
group['gsq_weighted'] = gsq_weighted
|
||
group['d'] = d
|
||
|
||
group_lr = group['lr']
|
||
decay = group['weight_decay']
|
||
k = group['k']
|
||
eps = group['eps']
|
||
|
||
for p in group['params']:
|
||
if p.grad is None:
|
||
continue
|
||
grad = p.grad.data
|
||
|
||
state = self.state[p]
|
||
|
||
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
|
||
|
||
state['step'] += 1
|
||
|
||
denom = exp_avg_sq.sqrt().add_(eps)
|
||
denom = denom.type(p.type())
|
||
|
||
# Apply weight decay (decoupled variant)
|
||
if decay != 0 and decouple and group_lr > 0:
|
||
p.data.add_(p.data, alpha=-decay * dlr)
|
||
|
||
# Take step
|
||
p.data.addcdiv_(exp_avg, denom, value=-1)
|
||
|
||
group['k'] = k + 1
|
||
|
||
return loss</code></pre>
|
||
</details>
|
||
<div class="desc"><p>Adam with D-Adaptation automatic step-sizes.
|
||
Leave LR set to 1 unless you encounter instability.</p>
|
||
<h2 id="args">Args</h2>
|
||
<dl>
|
||
<dt>params (iterable):</dt>
|
||
<dt>Iterable of parameters to optimize or dicts defining parameter groups.</dt>
|
||
<dt>lr (float):</dt>
|
||
<dt>Learning rate adjustment parameter. Increases or decreases the D-adapted learning rate.</dt>
|
||
<dt><strong><code>betas</code></strong> : <code>tuple[float, float]</code>, optional</dt>
|
||
<dd>coefficients used for computing
|
||
running averages of gradient and its square (default: (0.9, 0.999))</dd>
|
||
</dl>
|
||
<p>momentum (float):
|
||
Momentum value in
|
||
the range [0,1) (default: 0.9).
|
||
eps (float):
|
||
Term added to the denominator outside of the root operation to improve numerical stability. (default: 1e-8).
|
||
weight_decay (float):
|
||
Weight decay, i.e. a L2 penalty (default: 0).
|
||
log_every (int):
|
||
Log using print every k steps, default 0 (no logging).
|
||
decouple (boolean):
|
||
Use AdamW style decoupled weight decay
|
||
d0 (float):
|
||
Initial D estimate for D-adaptation (default 1e-6). Rarely needs changing.
|
||
growth_rate (float):
|
||
prevent the D estimate from growing faster than this multiplicative rate.
|
||
Default is inf, for unrestricted. Values like 1.02 give a kind of learning
|
||
rate warmup effect.
|
||
fsdp_in_use (bool):
|
||
If you're using sharded parameters, this should be set to True. The optimizer
|
||
will attempt to auto-detect this, but if you're using an implementation other
|
||
than PyTorch's builtin version, the auto-detection won't work.</p></div>
|
||
<h3>Ancestors</h3>
|
||
<ul class="hlist">
|
||
<li>torch.optim.optimizer.Optimizer</li>
|
||
</ul>
|
||
<h3>Class variables</h3>
|
||
<dl>
|
||
<dt id="audiocraft.optim.dadam.DAdaptAdam.OptimizerPostHook"><code class="name">var <span class="ident">OptimizerPostHook</span> : typing_extensions.TypeAlias</code></dt>
|
||
<dd>
|
||
<div class="desc"></div>
|
||
</dd>
|
||
<dt id="audiocraft.optim.dadam.DAdaptAdam.OptimizerPreHook"><code class="name">var <span class="ident">OptimizerPreHook</span> : typing_extensions.TypeAlias</code></dt>
|
||
<dd>
|
||
<div class="desc"></div>
|
||
</dd>
|
||
</dl>
|
||
<h3>Instance variables</h3>
|
||
<dl>
|
||
<dt id="audiocraft.optim.dadam.DAdaptAdam.supports_flat_params"><code class="name">prop <span class="ident">supports_flat_params</span></code></dt>
|
||
<dd>
|
||
<details class="source">
|
||
<summary>
|
||
<span>Expand source code</span>
|
||
</summary>
|
||
<pre><code class="python">@property
|
||
def supports_flat_params(self):
|
||
return True</code></pre>
|
||
</details>
|
||
<div class="desc"></div>
|
||
</dd>
|
||
<dt id="audiocraft.optim.dadam.DAdaptAdam.supports_memory_efficient_fp16"><code class="name">prop <span class="ident">supports_memory_efficient_fp16</span></code></dt>
|
||
<dd>
|
||
<details class="source">
|
||
<summary>
|
||
<span>Expand source code</span>
|
||
</summary>
|
||
<pre><code class="python">@property
|
||
def supports_memory_efficient_fp16(self):
|
||
return False</code></pre>
|
||
</details>
|
||
<div class="desc"></div>
|
||
</dd>
|
||
</dl>
|
||
<h3>Methods</h3>
|
||
<dl>
|
||
<dt id="audiocraft.optim.dadam.DAdaptAdam.step"><code class="name flex">
|
||
<span>def <span class="ident">step</span></span>(<span>self, closure=None)</span>
|
||
</code></dt>
|
||
<dd>
|
||
<details class="source">
|
||
<summary>
|
||
<span>Expand source code</span>
|
||
</summary>
|
||
<pre><code class="python">def step(self, closure=None):
|
||
"""Performs a single optimization step.
|
||
|
||
Args:
|
||
closure (callable, optional): A closure that reevaluates the model
|
||
and returns the loss.
|
||
"""
|
||
loss = None
|
||
if closure is not None:
|
||
loss = closure()
|
||
|
||
g_sq = 0.0
|
||
sksq_weighted = 0.0
|
||
sk_l1 = 0.0
|
||
|
||
lr = max(group['lr'] for group in self.param_groups)
|
||
|
||
group = self.param_groups[0]
|
||
gsq_weighted = group['gsq_weighted']
|
||
d = group['d']
|
||
dlr = d*lr
|
||
|
||
growth_rate = group['growth_rate']
|
||
decouple = group['decouple']
|
||
fsdp_in_use = group['fsdp_in_use']
|
||
log_every = group['log_every']
|
||
|
||
beta1, beta2 = group['betas']
|
||
|
||
for group in self.param_groups:
|
||
group_lr = group['lr']
|
||
decay = group['weight_decay']
|
||
k = group['k']
|
||
eps = group['eps']
|
||
|
||
if group_lr not in [lr, 0.0]:
|
||
raise RuntimeError("Setting different lr values in different parameter "
|
||
"groups is only supported for values of 0")
|
||
|
||
for p in group['params']:
|
||
if p.grad is None:
|
||
continue
|
||
if hasattr(p, "_fsdp_flattened"):
|
||
fsdp_in_use = True
|
||
grad = p.grad.data
|
||
|
||
# Apply weight decay (coupled variant)
|
||
if decay != 0 and not decouple:
|
||
grad.add_(p.data, alpha=decay)
|
||
|
||
state = self.state[p]
|
||
|
||
# State initialization
|
||
if 'step' not in state:
|
||
state['step'] = 0
|
||
state['s'] = torch.zeros_like(p.data, memory_format=torch.preserve_format).detach()
|
||
# Exponential moving average of gradient values
|
||
state['exp_avg'] = torch.zeros_like(p.data, memory_format=torch.preserve_format).detach()
|
||
# Exponential moving average of squared gradient values
|
||
state['exp_avg_sq'] = torch.zeros_like(
|
||
to_real(p.data), memory_format=torch.preserve_format).detach()
|
||
|
||
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
|
||
|
||
grad_grad = to_real(grad * grad.conj())
|
||
|
||
# Adam EMA updates
|
||
if group_lr > 0:
|
||
exp_avg.mul_(beta1).add_(grad, alpha=dlr*(1-beta1))
|
||
exp_avg_sq.mul_(beta2).add_(grad_grad, alpha=1-beta2)
|
||
|
||
denom = exp_avg_sq.sqrt().add_(eps)
|
||
|
||
g_sq += grad_grad.div_(denom).sum().item()
|
||
|
||
s = state['s']
|
||
s.mul_(beta2).add_(grad, alpha=dlr*(1-beta2))
|
||
sksq_weighted += to_real(s * s.conj()).div_(denom).sum().item()
|
||
sk_l1 += s.abs().sum().item()
|
||
|
||
######
|
||
|
||
gsq_weighted = beta2*gsq_weighted + g_sq*(dlr**2)*(1-beta2)
|
||
d_hat = d
|
||
|
||
# if we have not done any progres, return
|
||
# if we have any gradients available, will have sk_l1 > 0 (unless \|g\|=0)
|
||
if sk_l1 == 0:
|
||
return loss
|
||
|
||
if lr > 0.0:
|
||
if fsdp_in_use:
|
||
dist_tensor = torch.zeros(3, device='cuda')
|
||
dist_tensor[0] = sksq_weighted
|
||
dist_tensor[1] = gsq_weighted
|
||
dist_tensor[2] = sk_l1
|
||
dist.all_reduce(dist_tensor, op=dist.ReduceOp.SUM)
|
||
global_sksq_weighted = dist_tensor[0]
|
||
global_gsq_weighted = dist_tensor[1]
|
||
global_sk_l1 = dist_tensor[2]
|
||
else:
|
||
global_sksq_weighted = sksq_weighted
|
||
global_gsq_weighted = gsq_weighted
|
||
global_sk_l1 = sk_l1
|
||
|
||
d_hat = (global_sksq_weighted/(1-beta2) - global_gsq_weighted)/global_sk_l1
|
||
d = max(d, min(d_hat, d*growth_rate))
|
||
|
||
if log_every > 0 and k % log_every == 0:
|
||
logger.info(
|
||
f"(k={k}) dlr: {dlr:1.1e} d_hat: {d_hat:1.1e}, d: {d:1.8}. "
|
||
f"sksq_weighted={global_sksq_weighted:1.1e} gsq_weighted={global_gsq_weighted:1.1e} "
|
||
f"sk_l1={global_sk_l1:1.1e}{' (FSDP)' if fsdp_in_use else ''}")
|
||
|
||
for group in self.param_groups:
|
||
group['gsq_weighted'] = gsq_weighted
|
||
group['d'] = d
|
||
|
||
group_lr = group['lr']
|
||
decay = group['weight_decay']
|
||
k = group['k']
|
||
eps = group['eps']
|
||
|
||
for p in group['params']:
|
||
if p.grad is None:
|
||
continue
|
||
grad = p.grad.data
|
||
|
||
state = self.state[p]
|
||
|
||
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
|
||
|
||
state['step'] += 1
|
||
|
||
denom = exp_avg_sq.sqrt().add_(eps)
|
||
denom = denom.type(p.type())
|
||
|
||
# Apply weight decay (decoupled variant)
|
||
if decay != 0 and decouple and group_lr > 0:
|
||
p.data.add_(p.data, alpha=-decay * dlr)
|
||
|
||
# Take step
|
||
p.data.addcdiv_(exp_avg, denom, value=-1)
|
||
|
||
group['k'] = k + 1
|
||
|
||
return loss</code></pre>
|
||
</details>
|
||
<div class="desc"><p>Performs a single optimization step.</p>
|
||
<h2 id="args">Args</h2>
|
||
<dl>
|
||
<dt><strong><code>closure</code></strong> : <code>callable</code>, optional</dt>
|
||
<dd>A closure that reevaluates the model
|
||
and returns the loss.</dd>
|
||
</dl></div>
|
||
</dd>
|
||
</dl>
|
||
</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.optim" href="index.html">audiocraft.optim</a></code></li>
|
||
</ul>
|
||
</li>
|
||
<li><h3><a href="#header-functions">Functions</a></h3>
|
||
<ul class="">
|
||
<li><code><a title="audiocraft.optim.dadam.to_real" href="#audiocraft.optim.dadam.to_real">to_real</a></code></li>
|
||
</ul>
|
||
</li>
|
||
<li><h3><a href="#header-classes">Classes</a></h3>
|
||
<ul>
|
||
<li>
|
||
<h4><code><a title="audiocraft.optim.dadam.DAdaptAdam" href="#audiocraft.optim.dadam.DAdaptAdam">DAdaptAdam</a></code></h4>
|
||
<ul class="">
|
||
<li><code><a title="audiocraft.optim.dadam.DAdaptAdam.OptimizerPostHook" href="#audiocraft.optim.dadam.DAdaptAdam.OptimizerPostHook">OptimizerPostHook</a></code></li>
|
||
<li><code><a title="audiocraft.optim.dadam.DAdaptAdam.OptimizerPreHook" href="#audiocraft.optim.dadam.DAdaptAdam.OptimizerPreHook">OptimizerPreHook</a></code></li>
|
||
<li><code><a title="audiocraft.optim.dadam.DAdaptAdam.step" href="#audiocraft.optim.dadam.DAdaptAdam.step">step</a></code></li>
|
||
<li><code><a title="audiocraft.optim.dadam.DAdaptAdam.supports_flat_params" href="#audiocraft.optim.dadam.DAdaptAdam.supports_flat_params">supports_flat_params</a></code></li>
|
||
<li><code><a title="audiocraft.optim.dadam.DAdaptAdam.supports_memory_efficient_fp16" href="#audiocraft.optim.dadam.DAdaptAdam.supports_memory_efficient_fp16">supports_memory_efficient_fp16</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>
|