Hugging Face's logo

Buckets:
hf-doc-build
/
doc-dev

Files
xet
hf-doc-build/doc-dev / course /pr_1069 /it /chapter3 /3_tf.html
rtrm's picture
rtrm
download
raw
40.8 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Affinare un modell usando Keras&quot;,&quot;local&quot;:&quot;affinare-un-modell-usando-keras&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Addestramento&quot;,&quot;local&quot;:&quot;addestramento&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Migliorare la performance di addestramento&quot;,&quot;local&quot;:&quot;migliorare-la-performance-di-addestramento&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Predizioni del modello&quot;,&quot;local&quot;:&quot;predizioni-del-modello&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:1}">
 <link href="/docs/course/pr_1069/it/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/entry/start.693d748d.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/scheduler.37c15a92.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/singletons.60b4c7a2.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/index.18351ede.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/paths.43b6516c.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/entry/app.e9cfd099.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/index.2bf4358c.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/nodes/0.bb8a536c.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/each.e59479a4.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/nodes/23.f4f426a0.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/Tip.363c041f.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/Youtube.1e50a667.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/CodeBlock.4e987730.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/DocNotebookDropdown.efc1fb7c.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/FrameworkSwitchCourse.8d4d4ab6.js">
 <link rel="modulepreload" href="/docs/course/pr_1069/it/_app/immutable/chunks/getInferenceSnippets.24b50994.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Affinare un modell usando Keras&quot;,&quot;local&quot;:&quot;affinare-un-modell-usando-keras&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Addestramento&quot;,&quot;local&quot;:&quot;addestramento&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Migliorare la performance di addestramento&quot;,&quot;local&quot;:&quot;migliorare-la-performance-di-addestramento&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Predizioni del modello&quot;,&quot;local&quot;:&quot;predizioni-del-modello&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <div class="bg-white leading-none border border-gray-100 rounded-lg flex p-0.5 w-56 text-sm mb-4"><a class="flex justify-center flex-1 py-1.5 px-2.5 focus:outline-none !no-underline rounded-l bg-red-50 dark:bg-transparent text-red-600" href="?fw=pt"><svg class="mr-1.5" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><defs><clipPath id="a"><rect x="3.05" y="0.5" width="25.73" height="31" fill="none"></rect></clipPath></defs><g clip-path="url(#a)"><path d="M24.94,9.51a12.81,12.81,0,0,1,0,18.16,12.68,12.68,0,0,1-18,0,12.81,12.81,0,0,1,0-18.16l9-9V5l-.84.83-6,6a9.58,9.58,0,1,0,13.55,0ZM20.44,9a1.68,1.68,0,1,1,1.67-1.67A1.68,1.68,0,0,1,20.44,9Z" fill="#ee4c2c"></path></g></svg> Pytorch </a><a class="flex justify-center flex-1 py-1.5 px-2.5 focus:outline-none !no-underline rounded-r text-gray-500 filter grayscale" href="?fw=tf"><svg class="mr-1.5" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width="0.94em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 274"><path d="M145.726 42.065v42.07l72.861 42.07v-42.07l-72.86-42.07zM0 84.135v42.07l36.43 21.03V105.17L0 84.135zm109.291 21.035l-36.43 21.034v126.2l36.43 21.035v-84.135l36.435 21.035v-42.07l-36.435-21.034V105.17z" fill="#E55B2D"></path><path d="M145.726 42.065L36.43 105.17v42.065l72.861-42.065v42.065l36.435-21.03v-84.14zM255.022 63.1l-36.435 21.035v42.07l36.435-21.035V63.1zm-72.865 84.135l-36.43 21.035v42.07l36.43-21.036v-42.07zm-36.43 63.104l-36.436-21.035v84.135l36.435-21.035V210.34z" fill="#ED8E24"></path><path d="M145.726 0L0 84.135l36.43 21.035l109.296-63.105l72.861 42.07L255.022 63.1L145.726 0zm0 126.204l-36.435 21.03l36.435 21.036l36.43-21.035l-36.43-21.03z" fill="#F8BF3C"></path></svg> TensorFlow </a></div> <h1 class="relative group"><a id="affinare-un-modell-usando-keras" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#affinare-un-modell-usando-keras"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Affinare un modell usando Keras</span></h1> <div class="flex space-x-1 absolute z-10 right-0 top-0"> <a href="https://colab.research.google.com/github/huggingface/notebooks/blob/master/course/it/chapter3/section3_tf.ipynb" target="_blank"><img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"></a> <a href="https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/master/course/it/chapter3/section3_tf.ipynb" target="_blank"><img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"></a></div> <p data-svelte-h="svelte-mazm0f">Dopo tutto il lavoro di preprocessing nella sezione precedente, rimangono giusto gli ultimi passi per addestrare il modello. Attenzione tuttavia che il comando <code>model.fit()</code> sarà molto lento su una CPU. Se non avete una GPU a disposizione, potete avere accesso gratuitamente a GPU o TPU su <a href="https://colab.research.google.com/" rel="nofollow">Google Colab</a>.</p> <p data-svelte-h="svelte-11al6ht">Gli esempi di codice qui sotto partono dal presupposto che gli esempi nella sezione precedente siano già stati eseguiti. Ecco un breve riassunto di cosa serve:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><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
<span class="hljs-keyword">import</span> numpy <span class="hljs-keyword">as</span> np
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, return_tensors=<span class="hljs-string">&quot;tf&quot;</span>)
tf_train_dataset = tokenized_datasets[<span class="hljs-string">&quot;train&quot;</span>].to_tf_dataset(
 columns=[<span class="hljs-string">&quot;attention_mask&quot;</span>, <span class="hljs-string">&quot;input_ids&quot;</span>, <span class="hljs-string">&quot;token_type_ids&quot;</span>],
 label_cols=[<span class="hljs-string">&quot;labels&quot;</span>],
 shuffle=<span class="hljs-literal">True</span>,
 collate_fn=data_collator,
 batch_size=<span class="hljs-number">8</span>,
)
tf_validation_dataset = tokenized_datasets[<span class="hljs-string">&quot;validation&quot;</span>].to_tf_dataset(
 columns=[<span class="hljs-string">&quot;attention_mask&quot;</span>, <span class="hljs-string">&quot;input_ids&quot;</span>, <span class="hljs-string">&quot;token_type_ids&quot;</span>],
 label_cols=[<span class="hljs-string">&quot;labels&quot;</span>],
 shuffle=<span class="hljs-literal">False</span>,
 collate_fn=data_collator,
 batch_size=<span class="hljs-number">8</span>,
)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="addestramento" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#addestramento"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Addestramento</span></h3> <p data-svelte-h="svelte-29hd6y">I modelli di TensorFlow importati da 🤗 Transformers sono già dei modelli Keras. Ecco una breve introduzione a Keras.</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/rnTGBy2ax1c" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-t32qum">Ciò significa che una volta che si hanno i dati, è richiesto poco lavoro aggiuntivo per cominciare l’addestramento.</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/AUozVp78dhk" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-rpqwya">Come nel <a href="/course/chapter2">capitolo precedente</a>, verrà utilizzata la classe <code>TFAutoModelForSequenceClassification</code> con due etichette:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> TFAutoModelForSequenceClassification
model = TFAutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=<span class="hljs-number">2</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-12uqbjy">Diversamente dal <a href="/course/chapter2">Capitolo 2</a>, un avviso di avvertimento verrà visualizzato dopo aver istanziato questo modello pre-addestrato. Ciò avviene perché BERT non è stato pre-addestrato per classificare coppie di frasi, quindi la testa del modello pre-addestrato viene scartata e una nuova testa adeguata per il compito di classificazione di sequenze è stata inserita. Gli avvertimenti indicano che alcuni pesi non verranno usati (quelli corrispondenti alla testa scartata del modello pre-addestrato) e che altri pesi sono stati inizializzati con valori casuali (quelli per la nuova testa). L’avvertimento viene concluso con un’esortazione ad addestrare il modello, che è esattamente ciò che stiamo per fare.</p> <p data-svelte-h="svelte-elvb6f">Per affinare il modello sul dataset, bisogna solo chiamare <code>compile()</code> (compila) sul modello e passare i dati al metodo <code>fit()</code>. Ciò farà partire il processo di affinamento (che dovrebbe richiedere un paio di minuti su una GPU) e fare il report della funzione obiettivo di addestramento, in aggiunta alla funzione obiettivo di validazione alla fine di ogni epoca.</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-ez7phw">I modelli 🤗 Transformers hanno un’abilità speciale che manca alla maggior parte dei modelli Keras – possono usare in maniera automatica una funzione obiettivo appropriata, calcolata internamente. Questa funzione obiettivo verrà usata di default a meno che non venga definito l’argomento di funzione obiettivo nel metodo <code>compile()</code>. Per usare la funzione obiettivo interna è necessario passare le etichette come parte dell’input, non separatamente, che è l’approccio normale con i modelli Keras. Verranno mostrati esempi di ciò nella Parte 2 del corso, dove definire la funzione obiettivo correttamente può essere difficile. Per la classificazione di sequenze, invece, la funzione obiettivo standard di Keras funziona bene, quindi verrà utilizzata quella.</p></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> tensorflow.keras.losses <span class="hljs-keyword">import</span> SparseCategoricalCrossentropy
model.<span class="hljs-built_in">compile</span>(
 optimizer=<span class="hljs-string">&quot;adam&quot;</span>,
 loss=SparseCategoricalCrossentropy(from_logits=<span class="hljs-literal">True</span>),
 metrics=[<span class="hljs-string">&quot;accuracy&quot;</span>],
)
model.fit(
 tf_train_dataset,
 validation_data=tf_validation_dataset,
)<!-- HTML_TAG_END --></pre></div> <div class="course-tip course-tip-orange bg-gradient-to-br dark:bg-gradient-to-r before:border-orange-500 dark:before:border-orange-800 from-orange-50 dark:from-gray-900 to-white dark:to-gray-950 border border-orange-50 text-orange-700 dark:text-gray-400"><p data-svelte-h="svelte-1knd3lr">Attenzione ad un errore comune — si <em>può</em> passare solo il nome della funzione obiettivo a Keras come una stringa, ma di default Keras si aspetta che softmax sia già stato applicato ai risultati. Molti modelli invece forniscono come risultato i valori prima dell’applicazione del softmax, chiamati <em>logits</em>. Bisogna informare la funzione obiettivo che il nostro modello fa esattamente questo, e il solo modo di farlo è invocandola direttamente, non tramite la stringa che rappresenta il suo nome.</p></div> <h3 class="relative group"><a id="migliorare-la-performance-di-addestramento" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#migliorare-la-performance-di-addestramento"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Migliorare la performance di addestramento</span></h3> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/cpzq6ESSM5c" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-14aoo6j">Il codice presentato qui sopra sicuramente funziona, ma la funzione obiettivo scende in maniera molto lenta e sporadica. La causa principale di ciò è la <em>learning rate</em> (tasso di apprendimento). Come con la funzione obiettivo, quando si passa il nome di un ottimizzatore a Keras tramite una stringa, Keras inizializza quell’ottimizzatore con i valori di default per tutti i parametri, inclusa la learning rate. Grazie alla nostra esperienza, però, sappiamo che i modelli transformer beneficiano da una learning rate molto più bassa del valore di default per Adam, che è 1e-3, scritto anche come 10 alla -3, o 0.001. Il valore 5e-5 (0.00005), circa venti volte più basso, è un punto di partenza migliore.</p> <p data-svelte-h="svelte-1jd05ug">In aggiunta a diminuire la learning rate, abbiamo un secondo asso nella manica: possiamo ridurre lentamente la learning rate durante l’addestramento. Nella letteratura, questo approccio viene spesso indicato come <em>decaying</em> (decadimento) o <em>annealing</em> (ricottura) della learning rate. In Keras, il modo migliore per ottenere ciò è attraverso un <em>learning rate scheduler</em> (pianificatore del tasso di addestramento). Un buon esempio è <code>PolynomialDecay</code> (decadimento polinomiale) — nonostante il nome, con le impostazioni di default ha un semplice decadimento lineare dal valore iniziale a quello finale durante l’addestramento, che è esattamente ciò che cerchiamo. Per utilizzare lo scheduler in maniera corretta, però, bisogna dirgli quanto lungo sarà l’addestramento. Lo calcoliamo tramite la variabile <code>num_train_steps</code> nell’esempio qui sotto.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> tensorflow.keras.optimizers.schedules <span class="hljs-keyword">import</span> PolynomialDecay
batch_size = <span class="hljs-number">8</span>
num_epochs = <span class="hljs-number">3</span>
<span class="hljs-comment"># The number of training steps is the number of samples in the dataset, divided by the batch size then multiplied</span>
<span class="hljs-comment"># by the total number of epochs. Note that the tf_train_dataset here is a batched tf.data.Dataset,</span>
<span class="hljs-comment"># not the original Hugging Face Dataset, so its len() is already num_samples // batch_size.</span>
num_train_steps = <span class="hljs-built_in">len</span>(tf_train_dataset) * num_epochs
lr_scheduler = PolynomialDecay(
 initial_learning_rate=<span class="hljs-number">5e-5</span>, end_learning_rate=<span class="hljs-number">0.0</span>, decay_steps=num_train_steps
)
<span class="hljs-keyword">from</span> tensorflow.keras.optimizers <span class="hljs-keyword">import</span> Adam
opt = Adam(learning_rate=lr_scheduler)<!-- HTML_TAG_END --></pre></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-12088bb">La libreria 🤗 Transformers fornisce anche una funzione <code>create_optimizer()</code> che crea un ottimizzatore <code>AdamW</code> con decadimento della learning rate. Questa può essere una scorciatoia utile che verrà presentata nel dettaglio nelle sezioni future del corso.</p></div> <p data-svelte-h="svelte-oxsztq">Adesso che abbiamo il nostro ottimizzatore nuovo di zecca, possiamo provare con un addestramento. Per prima cosa, ricarichiamo il modello, per resettare i cambiamenti ai pesi dall’addestramento precedente, dopodiché lo possiamo compilare con nuovo ottimizzatore.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> tensorflow <span class="hljs-keyword">as</span> tf
model = TFAutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=<span class="hljs-number">2</span>)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=<span class="hljs-literal">True</span>)
model.<span class="hljs-built_in">compile</span>(optimizer=opt, loss=loss, metrics=[<span class="hljs-string">&quot;accuracy&quot;</span>])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-8qduz1">Ora chiamiamo di nuovo fit</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->model.fit(tf_train_dataset, validation_data=tf_validation_dataset, epochs=<span class="hljs-number">3</span>)<!-- HTML_TAG_END --></pre></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1b5s7xj">💡 Se vuoi caricare il modello in maniera automatica all’Hub durante l’addestramento, puoi passare <code>PushToHubCallback</code> al metodo <code>model.fit()</code>. Maggiori dettagli verranno forniti nel <a href="/course/chapter4/3">Capitolo 4</a></p></div> <h3 class="relative group"><a id="predizioni-del-modello" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#predizioni-del-modello"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Predizioni del modello</span></h3> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/nx10eh4CoOs" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <p data-svelte-h="svelte-1xxp91">Vedere la funzione obiettivo che diminuisce durante l’addestramento è bello, ma se volessimo ottenere dei risultati dal modello addestrato, ad esempio per calcolare delle metriche o per usare il modello in produzione? Per questo, si può usare il metodo <code>predict()</code>. Questo metodo restituisce i `<em>logits</em> dalla testa del modello, uno per classe.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->preds = model.predict(tf_validation_dataset)[<span class="hljs-string">&quot;logits&quot;</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1ylry8s">I logits possono essere convertiti nelle predizioni delle classi del modello usando <code>argmax</code> per trovare il logit più grande, che corrisponde alla classe più probabile.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->class_preds = np.argmax(preds, axis=<span class="hljs-number">1</span>)
<span class="hljs-built_in">print</span>(preds.shape, class_preds.shape)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-number">408</span>, <span class="hljs-number">2</span>) (<span class="hljs-number">408</span>,)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-oorxl5">Ora usiamo le <code>preds</code> per calcolare delle metriche! Si possono caricare le metriche associate al dataset MRPC in maniera facile tanto quanto caricare il dataset in sé, usando la funzione <code>load_metric()</code>. L’oggetto restituito ha un metodo <code>compute()</code> che può essere usato per calcolare le metriche.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> load_metric
metric = load_metric(<span class="hljs-string">&quot;glue&quot;</span>, <span class="hljs-string">&quot;mrpc&quot;</span>)
metric.compute(predictions=class_preds, references=raw_datasets[<span class="hljs-string">&quot;validation&quot;</span>][<span class="hljs-string">&quot;label&quot;</span>])<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->{<span class="hljs-string">&#x27;accuracy&#x27;</span>: <span class="hljs-number">0.8578431372549019</span>, <span class="hljs-string">&#x27;f1&#x27;</span>: <span class="hljs-number">0.8996539792387542</span>}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-nmfkbr">L’esatto valore dei risultati ottenuti può variare, poiché l’inizializzazione casuale della testa del modello può cambiare le metriche ottenute. Qui vediamo che il nostro modello ha una accuratezza del 87.78% sul validation set e valore F1 di 89.97. Queste sono le due metriche utilizzare per valutare risultati sul dataset MRPC per il benchmark GLUE. La tabella nell’<a href="https://arxiv.org/pdf/1810.04805.pdf" rel="nofollow">articolo du BERT</a> riportava un valore F1 di 88.9 per il modello di base. Quello era il modello <code>uncased</code>, mentre qui stiamo usando il modello <code>cased</code>, il che spiega i migliori risultati.</p> <p data-svelte-h="svelte-1xhes2n">Questo conclude l’introduzione all’affinamento usando l’API Keras. Un esempio per i compiti di NLP più comuni verrà fornito nel Capitolo 7. Per migliorare le vostre abilità con l’API Keras, provate ad affinare un modello sul dataset GLUE SST-2, usando il processing dei dati spiegato nella sezione 2.</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/course/blob/main/chapters/it/chapter3/3_tf.mdx" target="_blank"><span data-svelte-h="svelte-1kd6by1">&lt;</span> <span data-svelte-h="svelte-x0xyl0">&gt;</span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>
<script>
 {
 __sveltekit_u8ez91 = {
 assets: "/docs/course/pr_1069/it",
 base: "/docs/course/pr_1069/it",
 env: {}
 };
 const element = document.currentScript.parentElement;
 const data = [null,null];
 Promise.all([
 import("/docs/course/pr_1069/it/_app/immutable/entry/start.693d748d.js"),
 import("/docs/course/pr_1069/it/_app/immutable/entry/app.e9cfd099.js")
 ]).then(([kit, app]) => {
 kit.start(app, element, {
 node_ids: [0, 23],
 data,
 form: null,
 error: null
 });
 });
 }
 </script>

Xet Storage Details

Size:
40.8 kB
·
Xet hash:
b319958b8f6d275671ac70ed1ba3394704ba912958ba381f52e22d17c0851cd4

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.