course/pr_1213/ro/_app/immutable/nodes/48.b0f2bcaf.js

import{s as Tt,n as ht,o as ft}from"../chunks/scheduler.893fe8c9.js";import{S as Ct,i as Zt,e as i,s,c as r,h as Bt,a as p,d as t,b as n,f as wt,g as c,j as d,k as Nl,l as vt,m as a,n as M,t as o,o as m,p as u}from"../chunks/index.6ee278c6.js";import{C as It,H as Ie,E as Gt}from"../chunks/MermaidChart.svelte_svelte_type_style_lang.0be48b8a.js";import{Y as Ut}from"../chunks/Youtube.59d04f41.js";import{C as y}from"../chunks/CodeBlock.4d58b8dc.js";import{C as kt}from"../chunks/CourseFloatingBanner.b902b8d5.js";function Xt(El){let b,Ge,Be,ke,w,Xe,U,$e,T,We,h,ge,f,Ql="Acum vom vedea cum să obținem aceleași rezultate ca în secțiunea anterioară, dar fără să folosim clasa <code>Trainer</code>. Din nou, presupunem că ați parcurs deja procesarea datelor din secțiunea 2. Iată un scurt rezumat al tot ceea ce veți avea nevoie:",Re,C,Ye,Z,_e,B,xl="Înainte de a scrie efectiv bucla de antrenament, va trebui să definim câteva obiecte. Primele sunt încărcătoarele de date (dataloaders) pe care le vom folosi pentru a itera pe batch-uri. Dar, înainte de a putea defini acele dataloaders, trebuie să aplicăm un pic de postprocesare dataset-urilor noastre <code>tokenized_datasets</code>, pentru a ne ocupa de câteva lucruri pe care <code>Trainer</code> le făcea automat pentru noi. Mai exact, trebuie să:",Ae,v,Fl="<li>Eliminăm coloanele care corespund valorilor pe care modelul nu le așteaptă (cum ar fi coloanele <code>sentence1</code> și <code>sentence2</code>).</li> <li>Redenumim coloana <code>label</code> în <code>labels</code> (pentru că modelul se așteaptă ca argumentul să se numească <code>labels</code>).</li> <li>Setăm formatul dataset-urilor astfel încât să returneze tensori PyTorch în loc de liste.</li>",Ve,I,Hl="<code>tokenized_datasets</code> are câte o metodă pentru fiecare dintre acești pași:",ze,G,Ne,k,Sl="Putem apoi să verificăm că rezultatul are doar coloanele pe care modelul le va accepta:",Ee,X,Qe,$,ql="Acum, după ce am terminat acest pas, putem defini foarte ușor dataloader-urile noastre:",xe,W,Fe,g,Ll="Pentru a verifica rapid că nu există nicio eroare în procesarea datelor, putem inspecta un batch astfel:",He,R,Se,Y,qe,_,Kl="Observați că formele reale ar putea fi ușor diferite pentru voi, pentru că am setat <code>shuffle=True</code> în dataloader-ul nostru de antrenament și pentru că împachetăm (padding) la lungimea maximă în interiorul batch-ului.",Le,A,Pl="Acum că am terminat complet procesarea datelor (un obiectiv satisfăcător, dar uneori greu de atins pentru orice practician ML), să trecem la model. Îl instanțiem exact ca în secțiunea anterioară:",Ke,V,Pe,z,Dl="Pentru a ne asigura că totul va decurge fără probleme în timpul antrenamentului, trecem batch-ul nostru prin model:",De,N,Oe,E,el,Q,Ol="Toate modelele 🤗 Transformers vor returna pierderea (loss) când <code>labels</code> sunt furnizate, și, de asemenea, obținem logits (două pentru fiecare intrare în batch, deci un tensor de mărimea 8 x 2).",ll,x,et='Suntem aproape gata să scriem bucla de antrenament! Ne mai lipsesc două lucruri: un optimizer și un scheduler pentru rata de învățare. Pentru că încercăm să reproducem ceea ce făcea <code>Trainer</code>, vom folosi aceleași valori implicite. Optimizer-ul folosit de <code>Trainer</code> este <code>AdamW</code>, care este același cu Adam, dar cu o abordare particulară pentru regularizarea weight decay (vedeți lucrarea <a href="https://arxiv.org/abs/1711.05101" rel="nofollow">“Decoupled Weight Decay Regularization”</a> de Ilya Loshchilov și Frank Hutter):',tl,F,al,H,lt="În final, scheduler-ul pentru rata de învățare folosit implicit este doar o descreștere liniară de la valoarea maximă (5e-5) la 0. Pentru a-l defini corect, trebuie să știm numărul de pași de antrenament pe care îi vom face, care este numărul de epoci dorit înmulțit cu numărul de batch-uri de antrenament (care este lungimea dataloader-ului nostru de antrenament). <code>Trainer</code> folosește trei epoci implicit, așa că vom urma acest exemplu:",sl,S,nl,q,il,L,rl,K,tt="Încă un lucru: vom dori să folosim GPU-ul dacă avem acces la unul (pe un CPU, antrenamentul poate dura câteva ore în loc de câteva minute). Pentru asta, definim un <code>device</code> pe care vom pune modelul și batch-urile noastre:",pl,P,cl,D,Ml,O,at="Acum suntem gata de antrenament! Pentru a ne face o idee despre momentul în care se va termina antrenamentul, adăugăm o bară de progres peste numărul de pași de antrenament, folosind biblioteca <code>tqdm</code>:",ol,ee,ml,le,st="Observați că partea principală a buclei de antrenament arată foarte asemănător cu cea din introducere. Nu am cerut niciun raport, așa că această buclă de antrenament nu ne va spune nimic despre performanța modelului. Pentru a avea feedback, trebuie să adăugăm o buclă de evaluare.",ul,te,dl,ae,nt="Ca și înainte, vom folosi o metrică oferită de biblioteca 🤗 Evaluate. Am văzut deja metoda <code>metric.compute()</code>, dar metricle pot de fapt să acumuleze batch-uri pentru noi în timp ce parcurgem bucla de predicție, cu metoda <code>add_batch()</code>. Odată ce am acumulat toate batch-urile, putem obține rezultatul final cu <code>metric.compute()</code>. Iată cum să implementăm toate acestea într-o buclă de evaluare:",yl,se,bl,ne,jl,ie,it="Din nou, rezultatele voastre vor fi ușor diferite din cauza aleatorietății în inițializarea layer-ului final (model head) și a amestecării datelor, dar ar trebui să fie în aceeași zonă valorică.",Jl,j,rt="<p>✏️ <strong>Încercați!</strong> Modificați bucla de antrenament anterioară pentru a vă rafina modelul pe dataset-ul SST-2.</p>",wl,re,Ul,pe,Tl,ce,pt='Bucla de antrenament pe care am definit-o anterior funcționează bine pe un singur CPU sau GPU. Dar, folosind biblioteca <a href="https://github.com/huggingface/accelerate" rel="nofollow">🤗 Accelerate</a>, cu doar câteva ajustări putem activa antrenarea distribuită pe mai multe GPU-uri sau TPU-uri. Pornind de la crearea dataloader-urilor de antrenament și validare, iată cum arată bucla noastră manuală de antrenament:',hl,Me,fl,oe,ct="Iar aici sunt modificările:",Cl,me,Zl,ue,Mt="Prima linie de adăugat este linia de import. A doua linie instanțiază un obiect <code>Accelerator</code> care va examina mediul și va inițializa setarea distribuită corespunzătoare. 🤗 Accelerate se ocupă de plasarea pe device pentru voi, așa că puteți elimina liniile care pun modelul pe device (sau, dacă preferați, le puteți schimba să folosească <code>accelerator.device</code> în loc de <code>device</code>).",Bl,de,ot="Apoi, partea principală a muncii este făcută în linia care trimite dataloaders, modelul și optimizer-ul la <code>accelerator.prepare()</code>. Aceasta va împacheta acele obiecte în containerul potrivit pentru a vă asigura că antrenarea distribuită funcționează corespunzător. Restul modificărilor constau în eliminarea liniei care mută batch-ul pe <code>device</code> (din nou, dacă doriți să o păstrați puteți doar să o schimbați să folosească <code>accelerator.device</code>) și înlocuirea <code>loss.backward()</code> cu <code>accelerator.backward(loss)</code>.",vl,J,mt="<p>⚠️ Pentru a beneficia de creșterea vitezei oferită de Cloud TPU-uri, vă recomandăm să împachetați mostrele la o lungime fixă folosind argumentele <code>padding=&quot;max_length&quot;</code> și <code>max_length</code> ale tokenizer-ului.</p>",Il,ye,ut="Dacă vreți să copiați și să lipiți pentru a vă juca, iată cum arată bucla completă de antrenament cu 🤗 Accelerate:",Gl,be,kl,je,dt="Plasând acest cod într-un fișier <code>train.py</code> îl face rulabil pe orice tip de configurare distribuită. Pentru a-l încerca în configurarea voastră distribuită, rulați comanda:",Xl,Je,$l,we,yt="care vă va cere să răspundeți la câteva întrebări și vă va crea un fișier de configurare folosit de comanda:",Wl,Ue,gl,Te,bt="care va porni antrenarea distribuită.",Rl,he,jt="Dacă vreți să încercați asta într-un Notebook (de exemplu, pentru a-l testa cu TPU-uri pe Colab), doar lipiți codul într-o funcție <code>training_function()</code> și rulați într-o celulă finală:",Yl,fe,_l,Ce,Jt='Puteți găsi mai multe exemple în <a href="https://github.com/huggingface/accelerate/tree/main/examples" rel="nofollow">repo-ul 🤗 Accelerate</a>.',Al,Ze,Vl,ve,zl;return w=new It({props:{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"}}),U=new Ie({props:{title:"O instruire completă",local:"o-instruire-completă",headingTag:"h1"}}),T=new kt({props:{chapter:3,classNames:"absolute z-10 right-0 top-0",notebooks:[{label:"Google Colab",value:"https://colab.research.google.com/github/huggingface/notebooks/blob/master/course/en/chapter3/section4.ipynb"},{label:"Aws Studio",value:"https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/master/course/en/chapter3/section4.ipynb"}]}}),h=new Ut({props:{id:"Dh9CL8fyG80"}}),C=new y({props:{code:"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",highlighted:`<span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> load_dataset
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer, DataCollatorWithPadding

raw_datasets = load_dataset(<span class="hljs-string">&quot;glue&quot;</span>, <span class="hljs-string">&quot;mrpc&quot;</span>)
checkpoint = <span class="hljs-string">&quot;bert-base-uncased&quot;</span>
tokenizer = AutoTokenizer.from_pretrained(checkpoint)


<span class="hljs-keyword">def</span> <span class="hljs-title function_">tokenize_function</span>(<span class="hljs-params">example</span>):
    <span class="hljs-keyword">return</span> tokenizer(example[<span class="hljs-string">&quot;sentence1&quot;</span>], example[<span class="hljs-string">&quot;sentence2&quot;</span>], truncation=<span class="hljs-literal">True</span>)


tokenized_datasets = raw_datasets.<span class="hljs-built_in">map</span>(tokenize_function, batched=<span class="hljs-literal">True</span>)
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)`,wrap:!1}}),Z=new Ie({props:{title:"Pregătirea pentru antrenament",local:"pregătirea-pentru-antrenament",headingTag:"h3"}}),G=new y({props:{code:"dG9rZW5pemVkX2RhdGFzZXRzJTIwJTNEJTIwdG9rZW5pemVkX2RhdGFzZXRzLnJlbW92ZV9jb2x1bW5zKCU1QiUyMnNlbnRlbmNlMSUyMiUyQyUyMCUyMnNlbnRlbmNlMiUyMiUyQyUyMCUyMmlkeCUyMiU1RCklMEF0b2tlbml6ZWRfZGF0YXNldHMlMjAlM0QlMjB0b2tlbml6ZWRfZGF0YXNldHMucmVuYW1lX2NvbHVtbiglMjJsYWJlbCUyMiUyQyUyMCUyMmxhYmVscyUyMiklMEF0b2tlbml6ZWRfZGF0YXNldHMuc2V0X2Zvcm1hdCglMjJ0b3JjaCUyMiklMEF0b2tlbml6ZWRfZGF0YXNldHMlNUIlMjJ0cmFpbiUyMiU1RC5jb2x1bW5fbmFtZXM=",highlighted:`tokenized_datasets = tokenized_datasets.remove_columns([<span class="hljs-string">&quot;sentence1&quot;</span>, <span class="hljs-string">&quot;sentence2&quot;</span>, <span class="hljs-string">&quot;idx&quot;</span>])
tokenized_datasets = tokenized_datasets.rename_column(<span class="hljs-string">&quot;label&quot;</span>, <span class="hljs-string">&quot;labels&quot;</span>)
tokenized_datasets.set_format(<span class="hljs-string">&quot;torch&quot;</span>)
tokenized_datasets[<span class="hljs-string">&quot;train&quot;</span>].column_names`,wrap:!1}}),X=new y({props:{code:"JTVCJTIyYXR0ZW50aW9uX21hc2slMjIlMkMlMjAlMjJpbnB1dF9pZHMlMjIlMkMlMjAlMjJsYWJlbHMlMjIlMkMlMjAlMjJ0b2tlbl90eXBlX2lkcyUyMiU1RA==",highlighted:'[<span class="hljs-string">&quot;attention_mask&quot;</span>, <span class="hljs-string">&quot;input_ids&quot;</span>, <span class="hljs-string">&quot;labels&quot;</span>, <span class="hljs-string">&quot;token_type_ids&quot;</span>]',wrap:!1}}),W=new y({props:{code:"ZnJvbSUyMHRvcmNoLnV0aWxzLmRhdGElMjBpbXBvcnQlMjBEYXRhTG9hZGVyJTBBJTBBdHJhaW5fZGF0YWxvYWRlciUyMCUzRCUyMERhdGFMb2FkZXIoJTBBJTIwJTIwJTIwJTIwdG9rZW5pemVkX2RhdGFzZXRzJTVCJTIydHJhaW4lMjIlNUQlMkMlMjBzaHVmZmxlJTNEVHJ1ZSUyQyUyMGJhdGNoX3NpemUlM0Q4JTJDJTIwY29sbGF0ZV9mbiUzRGRhdGFfY29sbGF0b3IlMEEpJTBBZXZhbF9kYXRhbG9hZGVyJTIwJTNEJTIwRGF0YUxvYWRlciglMEElMjAlMjAlMjAlMjB0b2tlbml6ZWRfZGF0YXNldHMlNUIlMjJ2YWxpZGF0aW9uJTIyJTVEJTJDJTIwYmF0Y2hfc2l6ZSUzRDglMkMlMjBjb2xsYXRlX2ZuJTNEZGF0YV9jb2xsYXRvciUwQSk=",highlighted:`<span class="hljs-keyword">from</span> torch.utils.data <span class="hljs-keyword">import</span> DataLoader

train_dataloader = DataLoader(
    tokenized_datasets[<span class="hljs-string">&quot;train&quot;</span>], shuffle=<span class="hljs-literal">True</span>, batch_size=<span class="hljs-number">8</span>, collate_fn=data_collator
)
eval_dataloader = DataLoader(
    tokenized_datasets[<span class="hljs-string">&quot;validation&quot;</span>], batch_size=<span class="hljs-number">8</span>, collate_fn=data_collator
)`,wrap:!1}}),R=new y({props:{code:"Zm9yJTIwYmF0Y2glMjBpbiUyMHRyYWluX2RhdGFsb2FkZXIlM0ElMEElMjAlMjAlMjAlMjBicmVhayUwQSU3QmslM0ElMjB2LnNoYXBlJTIwZm9yJTIwayUyQyUyMHYlMjBpbiUyMGJhdGNoLml0ZW1zKCklN0Q=",highlighted:`<span class="hljs-keyword">for</span> batch <span class="hljs-keyword">in</span> train_dataloader:
    <span class="hljs-keyword">break</span>
{k: v.shape <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> batch.items()}`,wrap:!1}}),Y=new y({props:{code:"JTdCJ2F0dGVudGlvbl9tYXNrJyUzQSUyMHRvcmNoLlNpemUoJTVCOCUyQyUyMDY1JTVEKSUyQyUwQSUyMCdpbnB1dF9pZHMnJTNBJTIwdG9yY2guU2l6ZSglNUI4JTJDJTIwNjUlNUQpJTJDJTBBJTIwJ2xhYmVscyclM0ElMjB0b3JjaC5TaXplKCU1QjglNUQpJTJDJTBBJTIwJ3Rva2VuX3R5cGVfaWRzJyUzQSUyMHRvcmNoLlNpemUoJTVCOCUyQyUyMDY1JTVEKSU3RA==",highlighted:`{<span class="hljs-string">&#x27;attention_mask&#x27;</span>: torch.Size([<span class="hljs-number">8</span>, <span class="hljs-number">65</span>]),
 <span class="hljs-string">&#x27;input_ids&#x27;</span>: torch.Size([<span class="hljs-number">8</span>, <span class="hljs-number">65</span>]),
 <span class="hljs-string">&#x27;labels&#x27;</span>: torch.Size([<span class="hljs-number">8</span>]),
 <span class="hljs-string">&#x27;token_type_ids&#x27;</span>: torch.Size([<span class="hljs-number">8</span>, <span class="hljs-number">65</span>])}`,wrap:!1}}),V=new y({props:{code:"ZnJvbSUyMHRyYW5zZm9ybWVycyUyMGltcG9ydCUyMEF1dG9Nb2RlbEZvclNlcXVlbmNlQ2xhc3NpZmljYXRpb24lMEElMEFtb2RlbCUyMCUzRCUyMEF1dG9Nb2RlbEZvclNlcXVlbmNlQ2xhc3NpZmljYXRpb24uZnJvbV9wcmV0cmFpbmVkKGNoZWNrcG9pbnQlMkMlMjBudW1fbGFiZWxzJTNEMik=",highlighted:`<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=<span class="hljs-number">2</span>)`,wrap:!1}}),N=new y({props:{code:"b3V0cHV0cyUyMCUzRCUyMG1vZGVsKCoqYmF0Y2gpJTBBcHJpbnQob3V0cHV0cy5sb3NzJTJDJTIwb3V0cHV0cy5sb2dpdHMuc2hhcGUp",highlighted:`outputs = model(**batch)
<span class="hljs-built_in">print</span>(outputs.loss, outputs.logits.shape)`,wrap:!1}}),E=new y({props:{code:"dGVuc29yKDAuNTQ0MSUyQyUyMGdyYWRfZm4lM0QlM0NObGxMb3NzQmFja3dhcmQlM0UpJTIwdG9yY2guU2l6ZSglNUI4JTJDJTIwMiU1RCk=",highlighted:'tensor(<span class="hljs-number">0.5441</span>, grad_fn=&lt;NllLossBackward&gt;) torch.Size([<span class="hljs-number">8</span>, <span class="hljs-number">2</span>])',wrap:!1}}),F=new y({props:{code:"ZnJvbSUyMHRyYW5zZm9ybWVycyUyMGltcG9ydCUyMEFkYW1XJTBBJTBBb3B0aW1pemVyJTIwJTNEJTIwQWRhbVcobW9kZWwucGFyYW1ldGVycygpJTJDJTIwbHIlM0Q1ZS01KQ==",highlighted:`<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AdamW

optimizer = AdamW(model.parameters(), lr=<span class="hljs-number">5e-5</span>)`,wrap:!1}}),S=new y({props:{code:"ZnJvbSUyMHRyYW5zZm9ybWVycyUyMGltcG9ydCUyMGdldF9zY2hlZHVsZXIlMEElMEFudW1fZXBvY2hzJTIwJTNEJTIwMyUwQW51bV90cmFpbmluZ19zdGVwcyUyMCUzRCUyMG51bV9lcG9jaHMlMjAqJTIwbGVuKHRyYWluX2RhdGFsb2FkZXIpJTBBbHJfc2NoZWR1bGVyJTIwJTNEJTIwZ2V0X3NjaGVkdWxlciglMEElMjAlMjAlMjAlMjAlMjJsaW5lYXIlMjIlMkMlMEElMjAlMjAlMjAlMjBvcHRpbWl6ZXIlM0RvcHRpbWl6ZXIlMkMlMEElMjAlMjAlMjAlMjBudW1fd2FybXVwX3N0ZXBzJTNEMCUyQyUwQSUyMCUyMCUyMCUyMG51bV90cmFpbmluZ19zdGVwcyUzRG51bV90cmFpbmluZ19zdGVwcyUyQyUwQSklMEFwcmludChudW1fdHJhaW5pbmdfc3RlcHMp",highlighted:`<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> get_scheduler

num_epochs = <span class="hljs-number">3</span>
num_training_steps = num_epochs * <span class="hljs-built_in">len</span>(train_dataloader)
lr_scheduler = get_scheduler(
    <span class="hljs-string">&quot;linear&quot;</span>,
    optimizer=optimizer,
    num_warmup_steps=<span class="hljs-number">0</span>,
    num_training_steps=num_training_steps,
)
<span class="hljs-built_in">print</span>(num_training_steps)`,wrap:!1}}),q=new y({props:{code:"MTM3Nw==",highlighted:'<span class="hljs-number">1377</span>',wrap:!1}}),L=new Ie({props:{title:"Bucla de antrenament",local:"bucla-de-antrenament",headingTag:"h3"}}),P=new y({props:{code:"aW1wb3J0JTIwdG9yY2glMEElMEFkZXZpY2UlMjAlM0QlMjB0b3JjaC5kZXZpY2UoJTIyY3VkYSUyMiklMjBpZiUyMHRvcmNoLmN1ZGEuaXNfYXZhaWxhYmxlKCklMjBlbHNlJTIwdG9yY2guZGV2aWNlKCUyMmNwdSUyMiklMEFtb2RlbC50byhkZXZpY2UpJTBBZGV2aWNl",highlighted:`<span class="hljs-keyword">import</span> torch

device = torch.device(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-keyword">if</span> torch.cuda.is_available() <span class="hljs-keyword">else</span> torch.device(<span class="hljs-string">&quot;cpu&quot;</span>)
model.to(device)
device`,wrap:!1}}),D=new y({props:{code:"ZGV2aWNlKHR5cGUlM0QnY3VkYScp",highlighted:'device(<span class="hljs-built_in">type</span>=<span class="hljs-string">&#x27;cuda&#x27;</span>)',wrap:!1}}),ee=new y({props:{code:"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",highlighted:`<span class="hljs-keyword">from</span> tqdm.auto <span class="hljs-keyword">import</span> tqdm

progress_bar = tqdm(<span class="hljs-built_in">range</span>(num_training_steps))

model.train()
<span class="hljs-keyword">for</span> epoch <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(num_epochs):
    <span class="hljs-keyword">for</span> batch <span class="hljs-keyword">in</span> train_dataloader:
        batch = {k: v.to(device) <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> batch.items()}
        outputs = model(**batch)
        loss = outputs.loss
        loss.backward()

        optimizer.step()
        lr_scheduler.step()
        optimizer.zero_grad()
        progress_bar.update(<span class="hljs-number">1</span>)`,wrap:!1}}),te=new Ie({props:{title:"Bucla de evaluare",local:"bucla-de-evaluare",headingTag:"h3"}}),se=new y({props:{code:"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",highlighted:`<span class="hljs-keyword">import</span> evaluate

metric = evaluate.load(<span class="hljs-string">&quot;glue&quot;</span>, <span class="hljs-string">&quot;mrpc&quot;</span>)
model.<span class="hljs-built_in">eval</span>()
<span class="hljs-keyword">for</span> batch <span class="hljs-keyword">in</span> eval_dataloader:
    batch = {k: v.to(device) <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> batch.items()}
    <span class="hljs-keyword">with</span> torch.no_grad():
        outputs = model(**batch)

    logits = outputs.logits
    predictions = torch.argmax(logits, dim=-<span class="hljs-number">1</span>)
    metric.add_batch(predictions=predictions, references=batch[<span class="hljs-string">&quot;labels&quot;</span>])

metric.compute()`,wrap:!1}}),ne=new y({props:{code:"JTdCJ2FjY3VyYWN5JyUzQSUyMDAuODQzMTM3MjU0OTAxOTYwOCUyQyUyMCdmMSclM0ElMjAwLjg5MDc4NDk4MjkzNTE1MzUlN0Q=",highlighted:'{<span class="hljs-string">&#x27;accuracy&#x27;</span>: <span class="hljs-number">0.8431372549019608</span>, <span class="hljs-string">&#x27;f1&#x27;</span>: <span class="hljs-number">0.8907849829351535</span>}',wrap:!1}}),re=new Ie({props:{title:"Îmbunătățiți circuitul de antrenament cu 🤗 Accelerate",local:"îmbunătățiți-circuitul-de-antrenament-cu-accelerate",headingTag:"h3"}}),pe=new Ut({props:{id:"s7dy8QRgjJ0"}}),Me=new y({props:{code:"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",highlighted:`<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AdamW, AutoModelForSequenceClassification, get_scheduler

model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=<span class="hljs-number">2</span>)
optimizer = AdamW(model.parameters(), lr=<span class="hljs-number">3e-5</span>)

device = torch.device(<span class="hljs-string">&quot;cuda&quot;</span>) <span class="hljs-keyword">if</span> torch.cuda.is_available() <span class="hljs-keyword">else</span> torch.device(<span class="hljs-string">&quot;cpu&quot;</span>)
model.to(device)

num_epochs = <span class="hljs-number">3</span>
num_training_steps = num_epochs * <span class="hljs-built_in">len</span>(train_dataloader)
lr_scheduler = get_scheduler(
    <span class="hljs-string">&quot;linear&quot;</span>,
    optimizer=optimizer,
    num_warmup_steps=<span class="hljs-number">0</span>,
    num_training_steps=num_training_steps,
)

progress_bar = tqdm(<span class="hljs-built_in">range</span>(num_training_steps))

model.train()
<span class="hljs-keyword">for</span> epoch <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(num_epochs):
    <span class="hljs-keyword">for</span> batch <span class="hljs-keyword">in</span> train_dataloader:
        batch = {k: v.to(device) <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> batch.items()}
        outputs = model(**batch)
        loss = outputs.loss
        loss.backward()

        optimizer.step()
        lr_scheduler.step()
        optimizer.zero_grad()
        progress_bar.update(<span class="hljs-number">1</span>)`,wrap:!1}}),me=new y({props:{code:"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",highlighted:`<span class="hljs-addition">+ from accelerate import Accelerator</span>
  from transformers import AdamW, AutoModelForSequenceClassification, get_scheduler

<span class="hljs-addition">+ accelerator = Accelerator()</span>

  model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)
  optimizer = AdamW(model.parameters(), lr=3e-5)

<span class="hljs-deletion">- device = torch.device(&quot;cuda&quot;) if torch.cuda.is_available() else torch.device(&quot;cpu&quot;)</span>
<span class="hljs-deletion">- model.to(device)</span>

<span class="hljs-addition">+ train_dataloader, eval_dataloader, model, optimizer = accelerator.prepare(</span>
<span class="hljs-addition">+     train_dataloader, eval_dataloader, model, optimizer</span>
<span class="hljs-addition">+ )</span>

  num_epochs = 3
  num_training_steps = num_epochs * len(train_dataloader)
  lr_scheduler = get_scheduler(
      &quot;linear&quot;,
      optimizer=optimizer,
      num_warmup_steps=0,
      num_training_steps=num_training_steps
  )

  progress_bar = tqdm(range(num_training_steps))

  model.train()
  for epoch in range(num_epochs):
      for batch in train_dataloader:
<span class="hljs-deletion">-         batch = {k: v.to(device) for k, v in batch.items()}</span>
          outputs = model(**batch)
          loss = outputs.loss
<span class="hljs-deletion">-         loss.backward()</span>
<span class="hljs-addition">+         accelerator.backward(loss)</span>

          optimizer.step()
          lr_scheduler.step()
          optimizer.zero_grad()
          progress_bar.update(1)`,wrap:!1}}),be=new y({props:{code:"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",highlighted:`<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> Accelerator
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AdamW, AutoModelForSequenceClassification, get_scheduler

accelerator = Accelerator()

model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=<span class="hljs-number">2</span>)
optimizer = AdamW(model.parameters(), lr=<span class="hljs-number">3e-5</span>)

train_dl, eval_dl, model, optimizer = accelerator.prepare(
    train_dataloader, eval_dataloader, model, optimizer
)

num_epochs = <span class="hljs-number">3</span>
num_training_steps = num_epochs * <span class="hljs-built_in">len</span>(train_dl)
lr_scheduler = get_scheduler(
    <span class="hljs-string">&quot;linear&quot;</span>,
    optimizer=optimizer,
    num_warmup_steps=<span class="hljs-number">0</span>,
    num_training_steps=num_training_steps,
)

progress_bar = tqdm(<span class="hljs-built_in">range</span>(num_training_steps))

model.train()
<span class="hljs-keyword">for</span> epoch <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(num_epochs):
    <span class="hljs-keyword">for</span> batch <span class="hljs-keyword">in</span> train_dl:
        outputs = model(**batch)
        loss = outputs.loss
        accelerator.backward(loss)

        optimizer.step()
        lr_scheduler.step()
        optimizer.zero_grad()
        progress_bar.update(<span class="hljs-number">1</span>)`,wrap:!1}}),Je=new y({props:{code:"YWNjZWxlcmF0ZSUyMGNvbmZpZw==",highlighted:"accelerate config",wrap:!1}}),Ue=new y({props:{code:"YWNjZWxlcmF0ZSUyMGxhdW5jaCUyMHRyYWluLnB5",highlighted:'accelerate <span class="hljs-built_in">launch</span> train.py',wrap:!1}}),fe=new y({props:{code:"ZnJvbSUyMGFjY2VsZXJhdGUlMjBpbXBvcnQlMjBub3RlYm9va19sYXVuY2hlciUwQSUwQW5vdGVib29rX2xhdW5jaGVyKHRyYWluaW5nX2Z1bmN0aW9uKQ==",highlighted:`<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> notebook_launcher

notebook_launcher(training_function)`,wrap:!1}}),Ze=new Gt({props:{source:"https://github.com/huggingface/course/blob/main/chapters/ro/chapter3/4.mdx"}}),{c(){b=i("meta"),Ge=s(),Be=i("p"),ke=s(),r(w.$$.fragment),Xe=s(),r(U.$$.fragment),$e=s(),r(T.$$.fragment),We=s(),r(h.$$.fragment),ge=s(),f=i("p"),f.innerHTML=Ql,Re=s(),r(C.$$.fragment),Ye=s(),r(Z.$$.fragment),_e=s(),B=i("p"),B.innerHTML=xl,Ae=s(),v=i("ul"),v.innerHTML=Fl,Ve=s(),I=i("p"),I.innerHTML=Hl,ze=s(),r(G.$$.fragment),Ne=s(),k=i("p"),k.textContent=Sl,Ee=s(),r(X.$$.fragment),Qe=s(),$=i("p"),$.textContent=ql,xe=s(),r(W.$$.fragment),Fe=s(),g=i("p"),g.textContent=Ll,He=s(),r(R.$$.fragment),Se=s(),r(Y.$$.fragment),qe=s(),_=i("p"),_.innerHTML=Kl,Le=s(),A=i("p"),A.textContent=Pl,Ke=s(),r(V.$$.fragment),Pe=s(),z=i("p"),z.textContent=Dl,De=s(),r(N.$$.fragment),Oe=s(),r(E.$$.fragment),el=s(),Q=i("p"),Q.innerHTML=Ol,ll=s(),x=i("p"),x.innerHTML=et,tl=s(),r(F.$$.fragment),al=s(),H=i("p"),H.innerHTML=lt,sl=s(),r(S.$$.fragment),nl=s(),r(q.$$.fragment),il=s(),r(L.$$.fragment),rl=s(),K=i("p"),K.innerHTML=tt,pl=s(),r(P.$$.fragment),cl=s(),r(D.$$.fragment),Ml=s(),O=i("p"),O.innerHTML=at,ol=s(),r(ee.$$.fragment),ml=s(),le=i("p"),le.textContent=st,ul=s(),r(te.$$.fragment),dl=s(),ae=i("p"),ae.innerHTML=nt,yl=s(),r(se.$$.fragment),bl=s(),r(ne.$$.fragment),jl=s(),ie=i("p"),ie.textContent=it,Jl=s(),j=i("blockquote"),j.innerHTML=rt,wl=s(),r(re.$$.fragment),Ul=s(),r(pe.$$.fragment),Tl=s(),ce=i("p"),ce.innerHTML=pt,hl=s(),r(Me.$$.fragment),fl=s(),oe=i("p"),oe.textContent=ct,Cl=s(),r(me.$$.fragment),Zl=s(),ue=i("p"),ue.innerHTML=Mt,Bl=s(),de=i("p"),de.innerHTML=ot,vl=s(),J=i("blockquote"),J.innerHTML=mt,Il=s(),ye=i("p"),ye.textContent=ut,Gl=s(),r(be.$$.fragment),kl=s(),je=i("p"),je.innerHTML=dt,Xl=s(),r(Je.$$.fragment),$l=s(),we=i("p"),we.textContent=yt,Wl=s(),r(Ue.$$.fragment),gl=s(),Te=i("p"),Te.textContent=bt,Rl=s(),he=i("p"),he.innerHTML=jt,Yl=s(),r(fe.$$.fragment),_l=s(),Ce=i("p"),Ce.innerHTML=Jt,Al=s(),r(Ze.$$.fragment),Vl=s(),ve=i("p"),this.h()},l(e){const l=Bt("svelte-u9bgzb",document.head);b=p(l,"META",{name:!0,content:!0}),l.forEach(t),Ge=n(e),Be=p(e,"P",{}),wt(Be).forEach(t),ke=n(e),c(w.$$.fragment,e),Xe=n(e),c(U.$$.fragment,e),$e=n(e),c(T.$$.fragment,e),We=n(e),c(h.$$.fragment,e),ge=n(e),f=p(e,"P",{"data-svelte-h":!0}),d(f)!=="svelte-q6dk4a"&&(f.innerHTML=Ql),Re=n(e),c(C.$$.fragment,e),Ye=n(e),c(Z.$$.fragment,e),_e=n(e),B=p(e,"P",{"data-svelte-h":!0}),d(B)!=="svelte-uiid4o"&&(B.innerHTML=xl),Ae=n(e),v=p(e,"UL",{"data-svelte-h":!0}),d(v)!=="svelte-c2aek9"&&(v.innerHTML=Fl),Ve=n(e),I=p(e,"P",{"data-svelte-h":!0}),d(I)!=="svelte-1v3h9d0"&&(I.innerHTML=Hl),ze=n(e),c(G.$$.fragment,e),Ne=n(e),k=p(e,"P",{"data-svelte-h":!0}),d(k)!=="svelte-4ka9th"&&(k.textContent=Sl),Ee=n(e),c(X.$$.fragment,e),Qe=n(e),$=p(e,"P",{"data-svelte-h":!0}),d($)!=="svelte-awckyh"&&($.textContent=ql),xe=n(e),c(W.$$.fragment,e),Fe=n(e),g=p(e,"P",{"data-svelte-h":!0}),d(g)!=="svelte-r4tu2n"&&(g.textContent=Ll),He=n(e),c(R.$$.fragment,e),Se=n(e),c(Y.$$.fragment,e),qe=n(e),_=p(e,"P",{"data-svelte-h":!0}),d(_)!=="svelte-iz9nb8"&&(_.innerHTML=Kl),Le=n(e),A=p(e,"P",{"data-svelte-h":!0}),d(A)!=="svelte-q76el0"&&(A.textContent=Pl),Ke=n(e),c(V.$$.fragment,e),Pe=n(e),z=p(e,"P",{"data-svelte-h":!0}),d(z)!=="svelte-1wnox7m"&&(z.textContent=Dl),De=n(e),c(N.$$.fragment,e),Oe=n(e),c(E.$$.fragment,e),el=n(e),Q=p(e,"P",{"data-svelte-h":!0}),d(Q)!=="svelte-am270m"&&(Q.innerHTML=Ol),ll=n(e),x=p(e,"P",{"data-svelte-h":!0}),d(x)!=="svelte-qdgcv9"&&(x.innerHTML=et),tl=n(e),c(F.$$.fragment,e),al=n(e),H=p(e,"P",{"data-svelte-h":!0}),d(H)!=="svelte-1teovpf"&&(H.innerHTML=lt),sl=n(e),c(S.$$.fragment,e),nl=n(e),c(q.$$.fragment,e),il=n(e),c(L.$$.fragment,e),rl=n(e),K=p(e,"P",{"data-svelte-h":!0}),d(K)!=="svelte-18mygbx"&&(K.innerHTML=tt),pl=n(e),c(P.$$.fragment,e),cl=n(e),c(D.$$.fragment,e),Ml=n(e),O=p(e,"P",{"data-svelte-h":!0}),d(O)!=="svelte-qjyknt"&&(O.innerHTML=at),ol=n(e),c(ee.$$.fragment,e),ml=n(e),le=p(e,"P",{"data-svelte-h":!0}),d(le)!=="svelte-1uejh3i"&&(le.textContent=st),ul=n(e),c(te.$$.fragment,e),dl=n(e),ae=p(e,"P",{"data-svelte-h":!0}),d(ae)!=="svelte-15vnicp"&&(ae.innerHTML=nt),yl=n(e),c(se.$$.fragment,e),bl=n(e),c(ne.$$.fragment,e),jl=n(e),ie=p(e,"P",{"data-svelte-h":!0}),d(ie)!=="svelte-1qxzzsw"&&(ie.textContent=it),Jl=n(e),j=p(e,"BLOCKQUOTE",{class:!0,"data-svelte-h":!0}),d(j)!=="svelte-10f9jsz"&&(j.innerHTML=rt),wl=n(e),c(re.$$.fragment,e),Ul=n(e),c(pe.$$.fragment,e),Tl=n(e),ce=p(e,"P",{"data-svelte-h":!0}),d(ce)!=="svelte-1epvec1"&&(ce.innerHTML=pt),hl=n(e),c(Me.$$.fragment,e),fl=n(e),oe=p(e,"P",{"data-svelte-h":!0}),d(oe)!=="svelte-1a2jma4"&&(oe.textContent=ct),Cl=n(e),c(me.$$.fragment,e),Zl=n(e),ue=p(e,"P",{"data-svelte-h":!0}),d(ue)!=="svelte-nbet76"&&(ue.innerHTML=Mt),Bl=n(e),de=p(e,"P",{"data-svelte-h":!0}),d(de)!=="svelte-rpzdpp"&&(de.innerHTML=ot),vl=n(e),J=p(e,"BLOCKQUOTE",{class:!0,"data-svelte-h":!0}),d(J)!=="svelte-1c0o7ib"&&(J.innerHTML=mt),Il=n(e),ye=p(e,"P",{"data-svelte-h":!0}),d(ye)!=="svelte-1ygvpse"&&(ye.textContent=ut),Gl=n(e),c(be.$$.fragment,e),kl=n(e),je=p(e,"P",{"data-svelte-h":!0}),d(je)!=="svelte-1ch6shh"&&(je.innerHTML=dt),Xl=n(e),c(Je.$$.fragment,e),$l=n(e),we=p(e,"P",{"data-svelte-h":!0}),d(we)!=="svelte-silivt"&&(we.textContent=yt),Wl=n(e),c(Ue.$$.fragment,e),gl=n(e),Te=p(e,"P",{"data-svelte-h":!0}),d(Te)!=="svelte-1a33atb"&&(Te.textContent=bt),Rl=n(e),he=p(e,"P",{"data-svelte-h":!0}),d(he)!=="svelte-3m7ys1"&&(he.innerHTML=jt),Yl=n(e),c(fe.$$.fragment,e),_l=n(e),Ce=p(e,"P",{"data-svelte-h":!0}),d(Ce)!=="svelte-1jw41xa"&&(Ce.innerHTML=Jt),Al=n(e),c(Ze.$$.fragment,e),Vl=n(e),ve=p(e,"P",{}),wt(ve).forEach(t),this.h()},h(){Nl(b,"name","hf:doc:metadata"),Nl(b,"content",$t),Nl(j,"class","tip"),Nl(J,"class","tip")},m(e,l){vt(document.head,b),a(e,Ge,l),a(e,Be,l),a(e,ke,l),M(w,e,l),a(e,Xe,l),M(U,e,l),a(e,$e,l),M(T,e,l),a(e,We,l),M(h,e,l),a(e,ge,l),a(e,f,l),a(e,Re,l),M(C,e,l),a(e,Ye,l),M(Z,e,l),a(e,_e,l),a(e,B,l),a(e,Ae,l),a(e,v,l),a(e,Ve,l),a(e,I,l),a(e,ze,l),M(G,e,l),a(e,Ne,l),a(e,k,l),a(e,Ee,l),M(X,e,l),a(e,Qe,l),a(e,$,l),a(e,xe,l),M(W,e,l),a(e,Fe,l),a(e,g,l),a(e,He,l),M(R,e,l),a(e,Se,l),M(Y,e,l),a(e,qe,l),a(e,_,l),a(e,Le,l),a(e,A,l),a(e,Ke,l),M(V,e,l),a(e,Pe,l),a(e,z,l),a(e,De,l),M(N,e,l),a(e,Oe,l),M(E,e,l),a(e,el,l),a(e,Q,l),a(e,ll,l),a(e,x,l),a(e,tl,l),M(F,e,l),a(e,al,l),a(e,H,l),a(e,sl,l),M(S,e,l),a(e,nl,l),M(q,e,l),a(e,il,l),M(L,e,l),a(e,rl,l),a(e,K,l),a(e,pl,l),M(P,e,l),a(e,cl,l),M(D,e,l),a(e,Ml,l),a(e,O,l),a(e,ol,l),M(ee,e,l),a(e,ml,l),a(e,le,l),a(e,ul,l),M(te,e,l),a(e,dl,l),a(e,ae,l),a(e,yl,l),M(se,e,l),a(e,bl,l),M(ne,e,l),a(e,jl,l),a(e,ie,l),a(e,Jl,l),a(e,j,l),a(e,wl,l),M(re,e,l),a(e,Ul,l),M(pe,e,l),a(e,Tl,l),a(e,ce,l),a(e,hl,l),M(Me,e,l),a(e,fl,l),a(e,oe,l),a(e,Cl,l),M(me,e,l),a(e,Zl,l),a(e,ue,l),a(e,Bl,l),a(e,de,l),a(e,vl,l),a(e,J,l),a(e,Il,l),a(e,ye,l),a(e,Gl,l),M(be,e,l),a(e,kl,l),a(e,je,l),a(e,Xl,l),M(Je,e,l),a(e,$l,l),a(e,we,l),a(e,Wl,l),M(Ue,e,l),a(e,gl,l),a(e,Te,l),a(e,Rl,l),a(e,he,l),a(e,Yl,l),M(fe,e,l),a(e,_l,l),a(e,Ce,l),a(e,Al,l),M(Ze,e,l),a(e,Vl,l),a(e,ve,l),zl=!0},p:ht,i(e){zl||(o(w.$$.fragment,e),o(U.$$.fragment,e),o(T.$$.fragment,e),o(h.$$.fragment,e),o(C.$$.fragment,e),o(Z.$$.fragment,e),o(G.$$.fragment,e),o(X.$$.fragment,e),o(W.$$.fragment,e),o(R.$$.fragment,e),o(Y.$$.fragment,e),o(V.$$.fragment,e),o(N.$$.fragment,e),o(E.$$.fragment,e),o(F.$$.fragment,e),o(S.$$.fragment,e),o(q.$$.fragment,e),o(L.$$.fragment,e),o(P.$$.fragment,e),o(D.$$.fragment,e),o(ee.$$.fragment,e),o(te.$$.fragment,e),o(se.$$.fragment,e),o(ne.$$.fragment,e),o(re.$$.fragment,e),o(pe.$$.fragment,e),o(Me.$$.fragment,e),o(me.$$.fragment,e),o(be.$$.fragment,e),o(Je.$$.fragment,e),o(Ue.$$.fragment,e),o(fe.$$.fragment,e),o(Ze.$$.fragment,e),zl=!0)},o(e){m(w.$$.fragment,e),m(U.$$.fragment,e),m(T.$$.fragment,e),m(h.$$.fragment,e),m(C.$$.fragment,e),m(Z.$$.fragment,e),m(G.$$.fragment,e),m(X.$$.fragment,e),m(W.$$.fragment,e),m(R.$$.fragment,e),m(Y.$$.fragment,e),m(V.$$.fragment,e),m(N.$$.fragment,e),m(E.$$.fragment,e),m(F.$$.fragment,e),m(S.$$.fragment,e),m(q.$$.fragment,e),m(L.$$.fragment,e),m(P.$$.fragment,e),m(D.$$.fragment,e),m(ee.$$.fragment,e),m(te.$$.fragment,e),m(se.$$.fragment,e),m(ne.$$.fragment,e),m(re.$$.fragment,e),m(pe.$$.fragment,e),m(Me.$$.fragment,e),m(me.$$.fragment,e),m(be.$$.fragment,e),m(Je.$$.fragment,e),m(Ue.$$.fragment,e),m(fe.$$.fragment,e),m(Ze.$$.fragment,e),zl=!1},d(e){e&&(t(Ge),t(Be),t(ke),t(Xe),t($e),t(We),t(ge),t(f),t(Re),t(Ye),t(_e),t(B),t(Ae),t(v),t(Ve),t(I),t(ze),t(Ne),t(k),t(Ee),t(Qe),t($),t(xe),t(Fe),t(g),t(He),t(Se),t(qe),t(_),t(Le),t(A),t(Ke),t(Pe),t(z),t(De),t(Oe),t(el),t(Q),t(ll),t(x),t(tl),t(al),t(H),t(sl),t(nl),t(il),t(rl),t(K),t(pl),t(cl),t(Ml),t(O),t(ol),t(ml),t(le),t(ul),t(dl),t(ae),t(yl),t(bl),t(jl),t(ie),t(Jl),t(j),t(wl),t(Ul),t(Tl),t(ce),t(hl),t(fl),t(oe),t(Cl),t(Zl),t(ue),t(Bl),t(de),t(vl),t(J),t(Il),t(ye),t(Gl),t(kl),t(je),t(Xl),t($l),t(we),t(Wl),t(gl),t(Te),t(Rl),t(he),t(Yl),t(_l),t(Ce),t(Al),t(Vl),t(ve)),t(b),u(w,e),u(U,e),u(T,e),u(h,e),u(C,e),u(Z,e),u(G,e),u(X,e),u(W,e),u(R,e),u(Y,e),u(V,e),u(N,e),u(E,e),u(F,e),u(S,e),u(q,e),u(L,e),u(P,e),u(D,e),u(ee,e),u(te,e),u(se,e),u(ne,e),u(re,e),u(pe,e),u(Me,e),u(me,e),u(be,e),u(Je,e),u(Ue,e),u(fe,e),u(Ze,e)}}}const $t='{"title":"O instruire completă","local":"o-instruire-completă","sections":[{"title":"Pregătirea pentru antrenament","local":"pregătirea-pentru-antrenament","sections":[],"depth":3},{"title":"Bucla de antrenament","local":"bucla-de-antrenament","sections":[],"depth":3},{"title":"Bucla de evaluare","local":"bucla-de-evaluare","sections":[],"depth":3},{"title":"Îmbunătățiți circuitul de antrenament cu 🤗 Accelerate","local":"îmbunătățiți-circuitul-de-antrenament-cu-accelerate","sections":[],"depth":3}],"depth":1}';function Wt(El){return ft(()=>{new URLSearchParams(window.location.search).get("fw")}),[]}class zt extends Ct{constructor(b){super(),Zt(this,b,Wt,Xt,Tt,{})}}export{zt as component};