Buckets:

hf-doc-build/doc / peft /main /en /task_guides /semantic_segmentation_lora.html
rtrm's picture
download
raw
72.3 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Semantic segmentation using LoRA&quot;,&quot;local&quot;:&quot;semantic-segmentation-using-lora&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Install dependencies&quot;,&quot;local&quot;:&quot;install-dependencies&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Authenticate to share your model&quot;,&quot;local&quot;:&quot;authenticate-to-share-your-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Load a dataset&quot;,&quot;local&quot;:&quot;load-a-dataset&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Prepare label maps&quot;,&quot;local&quot;:&quot;prepare-label-maps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Prepare datasets for training and evaluation&quot;,&quot;local&quot;:&quot;prepare-datasets-for-training-and-evaluation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Create evaluation function&quot;,&quot;local&quot;:&quot;create-evaluation-function&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Load a base model&quot;,&quot;local&quot;:&quot;load-a-base-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Wrap the base model as a PeftModel for LoRA training&quot;,&quot;local&quot;:&quot;wrap-the-base-model-as-a-peftmodel-for-lora-training&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Train the model&quot;,&quot;local&quot;:&quot;train-the-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Save the model and run inference&quot;,&quot;local&quot;:&quot;save-the-model-and-run-inference&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
<link href="/docs/peft/main/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/entry/start.f1bde2f6.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/chunks/scheduler.d627b047.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/chunks/singletons.eea5c746.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/chunks/index.a57a1c33.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/chunks/paths.e43091d4.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/entry/app.182b7f64.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/chunks/index.d48c4817.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/nodes/0.8a16be44.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/chunks/each.e59479a4.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/nodes/39.20ee778b.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/chunks/CodeBlock.5da89496.js">
<link rel="modulepreload" href="/docs/peft/main/en/_app/immutable/chunks/Heading.47e562a9.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Semantic segmentation using LoRA&quot;,&quot;local&quot;:&quot;semantic-segmentation-using-lora&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Install dependencies&quot;,&quot;local&quot;:&quot;install-dependencies&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Authenticate to share your model&quot;,&quot;local&quot;:&quot;authenticate-to-share-your-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Load a dataset&quot;,&quot;local&quot;:&quot;load-a-dataset&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Prepare label maps&quot;,&quot;local&quot;:&quot;prepare-label-maps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Prepare datasets for training and evaluation&quot;,&quot;local&quot;:&quot;prepare-datasets-for-training-and-evaluation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Create evaluation function&quot;,&quot;local&quot;:&quot;create-evaluation-function&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Load a base model&quot;,&quot;local&quot;:&quot;load-a-base-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Wrap the base model as a PeftModel for LoRA training&quot;,&quot;local&quot;:&quot;wrap-the-base-model-as-a-peftmodel-for-lora-training&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Train the model&quot;,&quot;local&quot;:&quot;train-the-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Save the model and run inference&quot;,&quot;local&quot;:&quot;save-the-model-and-run-inference&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="semantic-segmentation-using-lora" 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="#semantic-segmentation-using-lora"><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>Semantic segmentation using LoRA</span></h1> <p data-svelte-h="svelte-ptgpdg">This guide demonstrates how to use LoRA, a low-rank approximation technique, to finetune a SegFormer model variant for semantic segmentation.
By using LoRA from 🤗 PEFT, we can reduce the number of trainable parameters in the SegFormer model to only 14% of the original trainable parameters.</p> <p data-svelte-h="svelte-1q529js">LoRA achieves this reduction by adding low-rank “update matrices” to specific blocks of the model, such as the attention
blocks. During fine-tuning, only these matrices are trained, while the original model parameters are left unchanged.
At inference time, the update matrices are merged with the original model parameters to produce the final classification result.</p> <p data-svelte-h="svelte-e5fkf5">For more information on LoRA, please refer to the <a href="https://arxiv.org/abs/2106.09685" rel="nofollow">original LoRA paper</a>.</p> <h2 class="relative group"><a id="install-dependencies" 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="#install-dependencies"><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>Install dependencies</span></h2> <p data-svelte-h="svelte-1hz22x2">Install the libraries required for model training:</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 -->!pip install transformers accelerate evaluate datasets peft -q<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="authenticate-to-share-your-model" 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="#authenticate-to-share-your-model"><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>Authenticate to share your model</span></h2> <p data-svelte-h="svelte-1t4proi">To share the finetuned model with the community at the end of the training, authenticate using your 🤗 token.
You can obtain your token from your <a href="https://huggingface.co/settings/token" rel="nofollow">account settings</a>.</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> huggingface_hub <span class="hljs-keyword">import</span> notebook_login
notebook_login()<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="load-a-dataset" 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="#load-a-dataset"><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>Load a dataset</span></h2> <p data-svelte-h="svelte-1w5yso9">To ensure that this example runs within a reasonable time frame, here we are limiting the number of instances from the training
set of the <a href="https://huggingface.co/datasets/scene_parse_150" rel="nofollow">SceneParse150 dataset</a> to 150.</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
ds = load_dataset(<span class="hljs-string">&quot;scene_parse_150&quot;</span>, split=<span class="hljs-string">&quot;train[:150]&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1d7ffhy">Next, split the dataset into train and test sets.</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 -->ds = ds.train_test_split(test_size=<span class="hljs-number">0.1</span>)
train_ds = ds[<span class="hljs-string">&quot;train&quot;</span>]
test_ds = ds[<span class="hljs-string">&quot;test&quot;</span>]<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="prepare-label-maps" 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="#prepare-label-maps"><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>Prepare label maps</span></h2> <p data-svelte-h="svelte-getpmi">Create a dictionary that maps a label id to a label class, which will be useful when setting up the model later:</p> <ul data-svelte-h="svelte-phg2cz"><li><code>label2id</code>: maps the semantic classes of the dataset to integer ids.</li> <li><code>id2label</code>: maps integer ids back to the semantic classes.</li></ul> <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> json
<span class="hljs-keyword">from</span> huggingface_hub <span class="hljs-keyword">import</span> cached_download, hf_hub_url
repo_id = <span class="hljs-string">&quot;huggingface/label-files&quot;</span>
filename = <span class="hljs-string">&quot;ade20k-id2label.json&quot;</span>
id2label = json.load(<span class="hljs-built_in">open</span>(cached_download(hf_hub_url(repo_id, filename, repo_type=<span class="hljs-string">&quot;dataset&quot;</span>)), <span class="hljs-string">&quot;r&quot;</span>))
id2label = {<span class="hljs-built_in">int</span>(k): v <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> id2label.items()}
label2id = {v: k <span class="hljs-keyword">for</span> k, v <span class="hljs-keyword">in</span> id2label.items()}
num_labels = <span class="hljs-built_in">len</span>(id2label)<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="prepare-datasets-for-training-and-evaluation" 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="#prepare-datasets-for-training-and-evaluation"><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>Prepare datasets for training and evaluation</span></h2> <p data-svelte-h="svelte-x8866f">Next, load the SegFormer image processor to prepare the images and annotations for the model. This dataset uses the
zero-index as the background class, so make sure to set <code>do_reduce_labels=True</code> to subtract one from all labels since the
background class is not among the 150 classes.</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> AutoImageProcessor
checkpoint = <span class="hljs-string">&quot;nvidia/mit-b0&quot;</span>
image_processor = AutoImageProcessor.from_pretrained(checkpoint, do_reduce_labels=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-fe2rtc">Add a function to apply data augmentation to the images, so that the model is more robust against overfitting. Here we use the
<a href="https://pytorch.org/vision/stable/generated/torchvision.transforms.ColorJitter.html" rel="nofollow">ColorJitter</a> function from
<a href="https://pytorch.org/vision/stable/index.html" rel="nofollow">torchvision</a> to randomly change the color properties of an image.</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> torchvision.transforms <span class="hljs-keyword">import</span> ColorJitter
jitter = ColorJitter(brightness=<span class="hljs-number">0.25</span>, contrast=<span class="hljs-number">0.25</span>, saturation=<span class="hljs-number">0.25</span>, hue=<span class="hljs-number">0.1</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-vm1b9v">Add a function to handle grayscale images and ensure that each input image has three color channels, regardless of
whether it was originally grayscale or RGB. The function converts RGB images to array as is, and for grayscale images
that have only one color channel, the function replicates the same channel three times using <code>np.tile()</code> before converting
the image into an array.</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> numpy <span class="hljs-keyword">as</span> np
<span class="hljs-keyword">def</span> <span class="hljs-title function_">handle_grayscale_image</span>(<span class="hljs-params">image</span>):
np_image = np.array(image)
<span class="hljs-keyword">if</span> np_image.ndim == <span class="hljs-number">2</span>:
tiled_image = np.tile(np.expand_dims(np_image, -<span class="hljs-number">1</span>), <span class="hljs-number">3</span>)
<span class="hljs-keyword">return</span> Image.fromarray(tiled_image)
<span class="hljs-keyword">else</span>:
<span class="hljs-keyword">return</span> Image.fromarray(np_image)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-yxifz1">Finally, combine everything in two functions that you’ll use to transform training and validation data. The two functions
are similar except data augmentation is applied only to the training data.</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> PIL <span class="hljs-keyword">import</span> Image
<span class="hljs-keyword">def</span> <span class="hljs-title function_">train_transforms</span>(<span class="hljs-params">example_batch</span>):
images = [jitter(handle_grayscale_image(x)) <span class="hljs-keyword">for</span> x <span class="hljs-keyword">in</span> example_batch[<span class="hljs-string">&quot;image&quot;</span>]]
labels = [x <span class="hljs-keyword">for</span> x <span class="hljs-keyword">in</span> example_batch[<span class="hljs-string">&quot;annotation&quot;</span>]]
inputs = image_processor(images, labels)
<span class="hljs-keyword">return</span> inputs
<span class="hljs-keyword">def</span> <span class="hljs-title function_">val_transforms</span>(<span class="hljs-params">example_batch</span>):
images = [handle_grayscale_image(x) <span class="hljs-keyword">for</span> x <span class="hljs-keyword">in</span> example_batch[<span class="hljs-string">&quot;image&quot;</span>]]
labels = [x <span class="hljs-keyword">for</span> x <span class="hljs-keyword">in</span> example_batch[<span class="hljs-string">&quot;annotation&quot;</span>]]
inputs = image_processor(images, labels)
<span class="hljs-keyword">return</span> inputs<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-sagovx">To apply the preprocessing functions over the entire dataset, use the 🤗 Datasets <code>set_transform</code> function:</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 -->train_ds.set_transform(train_transforms)
test_ds.set_transform(val_transforms)<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="create-evaluation-function" 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="#create-evaluation-function"><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>Create evaluation function</span></h2> <p data-svelte-h="svelte-1x2616c">Including a metric during training is helpful for evaluating your model’s performance. You can load an evaluation
method with the <a href="https://huggingface.co/docs/evaluate/index" rel="nofollow">🤗 Evaluate</a> library. For this task, use
the <a href="https://huggingface.co/spaces/evaluate-metric/accuracy" rel="nofollow">mean Intersection over Union (IoU)</a> metric (see the 🤗 Evaluate
<a href="https://huggingface.co/docs/evaluate/a_quick_tour" rel="nofollow">quick tour</a> to learn more about how to load and compute a metric):</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> torch
<span class="hljs-keyword">from</span> torch <span class="hljs-keyword">import</span> nn
<span class="hljs-keyword">import</span> evaluate
metric = evaluate.load(<span class="hljs-string">&quot;mean_iou&quot;</span>)
<span class="hljs-keyword">def</span> <span class="hljs-title function_">compute_metrics</span>(<span class="hljs-params">eval_pred</span>):
<span class="hljs-keyword">with</span> torch.no_grad():
logits, labels = eval_pred
logits_tensor = torch.from_numpy(logits)
logits_tensor = nn.functional.interpolate(
logits_tensor,
size=labels.shape[-<span class="hljs-number">2</span>:],
mode=<span class="hljs-string">&quot;bilinear&quot;</span>,
align_corners=<span class="hljs-literal">False</span>,
).argmax(dim=<span class="hljs-number">1</span>)
pred_labels = logits_tensor.detach().cpu().numpy()
<span class="hljs-comment"># currently using _compute instead of compute</span>
<span class="hljs-comment"># see this issue for more info: https://github.com/huggingface/evaluate/pull/328#issuecomment-1286866576</span>
metrics = metric._compute(
predictions=pred_labels,
references=labels,
num_labels=<span class="hljs-built_in">len</span>(id2label),
ignore_index=<span class="hljs-number">0</span>,
reduce_labels=image_processor.do_reduce_labels,
)
per_category_accuracy = metrics.pop(<span class="hljs-string">&quot;per_category_accuracy&quot;</span>).tolist()
per_category_iou = metrics.pop(<span class="hljs-string">&quot;per_category_iou&quot;</span>).tolist()
metrics.update({<span class="hljs-string">f&quot;accuracy_<span class="hljs-subst">{id2label[i]}</span>&quot;</span>: v <span class="hljs-keyword">for</span> i, v <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(per_category_accuracy)})
metrics.update({<span class="hljs-string">f&quot;iou_<span class="hljs-subst">{id2label[i]}</span>&quot;</span>: v <span class="hljs-keyword">for</span> i, v <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(per_category_iou)})
<span class="hljs-keyword">return</span> metrics<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="load-a-base-model" 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="#load-a-base-model"><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>Load a base model</span></h2> <p data-svelte-h="svelte-z7c6p7">Before loading a base model, let’s define a helper function to check the total number of parameters a model has, as well
as how many of them are trainable.</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">def</span> <span class="hljs-title function_">print_trainable_parameters</span>(<span class="hljs-params">model</span>):
<span class="hljs-string">&quot;&quot;&quot;
Prints the number of trainable parameters in the model.
&quot;&quot;&quot;</span>
trainable_params = <span class="hljs-number">0</span>
all_param = <span class="hljs-number">0</span>
<span class="hljs-keyword">for</span> _, param <span class="hljs-keyword">in</span> model.named_parameters():
all_param += param.numel()
<span class="hljs-keyword">if</span> param.requires_grad:
trainable_params += param.numel()
<span class="hljs-built_in">print</span>(
<span class="hljs-string">f&quot;trainable params: <span class="hljs-subst">{trainable_params}</span> || all params: <span class="hljs-subst">{all_param}</span> || trainable%: <span class="hljs-subst">{<span class="hljs-number">100</span> * trainable_params / all_param:<span class="hljs-number">.2</span>f}</span>&quot;</span>
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-134egv8">Choose a base model checkpoint. For this example, we use the <a href="https://huggingface.co/nvidia/mit-b0" rel="nofollow">SegFormer B0 variant</a>.
In addition to the checkpoint, pass the <code>label2id</code> and <code>id2label</code> dictionaries to let the <code>AutoModelForSemanticSegmentation</code> class know that we’re
interested in a custom base model where the decoder head should be randomly initialized using the classes from the custom dataset.</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> AutoModelForSemanticSegmentation, TrainingArguments, Trainer
model = AutoModelForSemanticSegmentation.from_pretrained(
checkpoint, id2label=id2label, label2id=label2id, ignore_mismatched_sizes=<span class="hljs-literal">True</span>
)
print_trainable_parameters(model)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1hyd5lx">At this point you can check with the <code>print_trainable_parameters</code> helper function that all 100% parameters in the base
model (aka <code>model</code>) are trainable.</p> <h2 class="relative group"><a id="wrap-the-base-model-as-a-peftmodel-for-lora-training" 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="#wrap-the-base-model-as-a-peftmodel-for-lora-training"><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>Wrap the base model as a PeftModel for LoRA training</span></h2> <p data-svelte-h="svelte-h2ilia">To leverage the LoRa method, you need to wrap the base model as a <code>PeftModel</code>. This involves two steps:</p> <ol data-svelte-h="svelte-gdfjac"><li>Defining LoRa configuration with <code>LoraConfig</code></li> <li>Wrapping the original <code>model</code> with <code>get_peft_model()</code> using the config defined in the step above.</li></ol> <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> peft <span class="hljs-keyword">import</span> LoraConfig, get_peft_model
config = LoraConfig(
r=<span class="hljs-number">32</span>,
lora_alpha=<span class="hljs-number">32</span>,
target_modules=[<span class="hljs-string">&quot;query&quot;</span>, <span class="hljs-string">&quot;value&quot;</span>],
lora_dropout=<span class="hljs-number">0.1</span>,
bias=<span class="hljs-string">&quot;lora_only&quot;</span>,
modules_to_save=[<span class="hljs-string">&quot;decode_head&quot;</span>],
)
lora_model = get_peft_model(model, config)
print_trainable_parameters(lora_model)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1dir0x5">Let’s review the <code>LoraConfig</code>. To enable LoRA technique, we must define the target modules within <code>LoraConfig</code> so that
<code>PeftModel</code> can update the necessary matrices. Specifically, we want to target the <code>query</code> and <code>value</code> matrices in the
attention blocks of the base model. These matrices are identified by their respective names, “query” and “value”.
Therefore, we should specify these names in the <code>target_modules</code> argument of <code>LoraConfig</code>.</p> <p data-svelte-h="svelte-l3bbz">After we wrap our base model <code>model</code> with <code>PeftModel</code> along with the config, we get
a new model where only the LoRA parameters are trainable (so-called “update matrices”) while the pre-trained parameters
are kept frozen. These include the parameters of the randomly initialized classifier parameters too. This is NOT we want
when fine-tuning the base model on our custom dataset. To ensure that the classifier parameters are also trained, we
specify <code>modules_to_save</code>. This also ensures that these modules are serialized alongside the LoRA trainable parameters
when using utilities like <code>save_pretrained()</code> and <code>push_to_hub()</code>.</p> <p data-svelte-h="svelte-1coy915">In addition to specifying the <code>target_modules</code> within <code>LoraConfig</code>, we also need to specify the <code>modules_to_save</code>. When
we wrap our base model with <code>PeftModel</code> and pass the configuration, we obtain a new model in which only the LoRA parameters
are trainable, while the pre-trained parameters and the randomly initialized classifier parameters are kept frozen.
However, we do want to train the classifier parameters. By specifying the <code>modules_to_save</code> argument, we ensure that the
classifier parameters are also trainable, and they will be serialized alongside the LoRA trainable parameters when we
use utility functions like <code>save_pretrained()</code> and <code>push_to_hub()</code>.</p> <p data-svelte-h="svelte-ss5ia">Let’s review the rest of the parameters:</p> <ul data-svelte-h="svelte-1m2rv79"><li><code>r</code>: The dimension used by the LoRA update matrices.</li> <li><code>alpha</code>: Scaling factor.</li> <li><code>bias</code>: Specifies if the <code>bias</code> parameters should be trained. <code>None</code> denotes none of the <code>bias</code> parameters will be trained.</li></ul> <p data-svelte-h="svelte-10wmx8i">When all is configured, and the base model is wrapped, the <code>print_trainable_parameters</code> helper function lets us explore
the number of trainable parameters. Since we’re interested in performing <strong>parameter-efficient fine-tuning</strong>,
we should expect to see a lower number of trainable parameters from the <code>lora_model</code> in comparison to the original <code>model</code>
which is indeed the case here.</p> <p data-svelte-h="svelte-1o171ky">You can also manually verify what modules are trainable in the <code>lora_model</code>.</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">for</span> name, param <span class="hljs-keyword">in</span> lora_model.named_parameters():
<span class="hljs-keyword">if</span> param.requires_grad:
<span class="hljs-built_in">print</span>(name, param.shape)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1g7xqd2">This confirms that only the LoRA parameters appended to the attention blocks and the <code>decode_head</code> parameters are trainable.</p> <h2 class="relative group"><a id="train-the-model" 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="#train-the-model"><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>Train the model</span></h2> <p data-svelte-h="svelte-sicdsw">Start by defining your training hyperparameters in <code>TrainingArguments</code>. You can change the values of most parameters however
you prefer. Make sure to set <code>remove_unused_columns=False</code>, otherwise the image column will be dropped, and it’s required here.
The only other required parameter is <code>output_dir</code> which specifies where to save your model.
At the end of each epoch, the <code>Trainer</code> will evaluate the IoU metric and save the training checkpoint.</p> <p data-svelte-h="svelte-sj51v2">Note that this example is meant to walk you through the workflow when using PEFT for semantic segmentation. We didn’t
perform extensive hyperparameter tuning to achieve optimal results.</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_name = checkpoint.split(<span class="hljs-string">&quot;/&quot;</span>)[-<span class="hljs-number">1</span>]
training_args = TrainingArguments(
output_dir=<span class="hljs-string">f&quot;<span class="hljs-subst">{model_name}</span>-scene-parse-150-lora&quot;</span>,
learning_rate=<span class="hljs-number">5e-4</span>,
num_train_epochs=<span class="hljs-number">50</span>,
per_device_train_batch_size=<span class="hljs-number">4</span>,
per_device_eval_batch_size=<span class="hljs-number">2</span>,
save_total_limit=<span class="hljs-number">3</span>,
evaluation_strategy=<span class="hljs-string">&quot;epoch&quot;</span>,
save_strategy=<span class="hljs-string">&quot;epoch&quot;</span>,
logging_steps=<span class="hljs-number">5</span>,
remove_unused_columns=<span class="hljs-literal">False</span>,
push_to_hub=<span class="hljs-literal">True</span>,
label_names=[<span class="hljs-string">&quot;labels&quot;</span>],
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-nimjtw">Pass the training arguments to <code>Trainer</code> along with the model, dataset, and <code>compute_metrics</code> function.
Call <code>train()</code> to finetune your model.</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 -->trainer = Trainer(
model=lora_model,
args=training_args,
train_dataset=train_ds,
eval_dataset=test_ds,
compute_metrics=compute_metrics,
)
trainer.train()<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="save-the-model-and-run-inference" 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="#save-the-model-and-run-inference"><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>Save the model and run inference</span></h2> <p data-svelte-h="svelte-wzzfy">Use the <code>save_pretrained()</code> method of the <code>lora_model</code> to save the <em>LoRA-only parameters</em> locally.
Alternatively, use the <code>push_to_hub()</code> method to upload these parameters directly to the Hugging Face Hub
(as shown in the <a href="image_classification_lora">Image classification using LoRA</a> task guide).</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_id = <span class="hljs-string">&quot;segformer-scene-parse-150-lora&quot;</span>
lora_model.save_pretrained(model_id)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-8jrf93">We can see that the LoRA-only parameters are just <strong>2.2 MB in size</strong>! This greatly improves the portability when using very large models.</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-built_in">ls</span> -lh {model_id}
total 2.2M
-rw-r--r-- 1 root root 369 Feb 8 03:09 adapter_config.json
-rw-r--r-- 1 root root 2.2M Feb 8 03:09 adapter_model.bin<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1ws2x6q">Let’s now prepare an <code>inference_model</code> and run inference.</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> peft <span class="hljs-keyword">import</span> PeftConfig
config = PeftConfig.from_pretrained(model_id)
model = AutoModelForSemanticSegmentation.from_pretrained(
checkpoint, id2label=id2label, label2id=label2id, ignore_mismatched_sizes=<span class="hljs-literal">True</span>
)
inference_model = PeftModel.from_pretrained(model, model_id)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1k70j7o">Get an image:</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> requests
url = <span class="hljs-string">&quot;https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/semantic-seg-image.png&quot;</span>
image = Image.<span class="hljs-built_in">open</span>(requests.get(url, stream=<span class="hljs-literal">True</span>).raw)
image<!-- HTML_TAG_END --></pre></div> <div class="flex justify-center" data-svelte-h="svelte-runw56"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/semantic-seg-image.png" alt="photo of a room"></div> <p data-svelte-h="svelte-1a29ol4">Preprocess the image to prepare for inference.</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 -->encoding = image_processor(image.convert(<span class="hljs-string">&quot;RGB&quot;</span>), return_tensors=<span class="hljs-string">&quot;pt&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-k2a84">Run inference with the encoded image.</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">with</span> torch.no_grad():
outputs = inference_model(pixel_values=encoding.pixel_values)
logits = outputs.logits
upsampled_logits = nn.functional.interpolate(
logits,
size=image.size[::-<span class="hljs-number">1</span>],
mode=<span class="hljs-string">&quot;bilinear&quot;</span>,
align_corners=<span class="hljs-literal">False</span>,
)
pred_seg = upsampled_logits.argmax(dim=<span class="hljs-number">1</span>)[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-paeq5">Next, visualize the results. We need a color palette for this. Here, we use ade_palette(). As it is a long array, so
we don’t include it in this guide, please copy it from <a href="https://github.com/tensorflow/models/blob/3f1ca33afe3c1631b733ea7e40c294273b9e406d/research/deeplab/utils/get_dataset_colormap.py#L51" rel="nofollow">the TensorFlow Model Garden repository</a>.</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> matplotlib.pyplot <span class="hljs-keyword">as</span> plt
color_seg = np.zeros((pred_seg.shape[<span class="hljs-number">0</span>], pred_seg.shape[<span class="hljs-number">1</span>], <span class="hljs-number">3</span>), dtype=np.uint8)
palette = np.array(ade_palette())
<span class="hljs-keyword">for</span> label, color <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(palette):
color_seg[pred_seg == label, :] = color
color_seg = color_seg[..., ::-<span class="hljs-number">1</span>] <span class="hljs-comment"># convert to BGR</span>
img = np.array(image) * <span class="hljs-number">0.5</span> + color_seg * <span class="hljs-number">0.5</span> <span class="hljs-comment"># plot the image with the segmentation map</span>
img = img.astype(np.uint8)
plt.figure(figsize=(<span class="hljs-number">15</span>, <span class="hljs-number">10</span>))
plt.imshow(img)
plt.show()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1rbxn15">As you can see, the results are far from perfect, however, this example is designed to illustrate the end-to-end workflow of
fine-tuning a semantic segmentation model with LoRa technique, and is not aiming to achieve state-of-the-art
results. The results you see here are the same as you would get if you performed full fine-tuning on the same setup (same
model variant, same dataset, same training schedule, etc.), except LoRA allows to achieve them with a fraction of total
trainable parameters and in less time.</p> <p data-svelte-h="svelte-1bj2ejn">If you wish to use this example and improve the results, here are some things that you can try:</p> <ul data-svelte-h="svelte-1herdxp"><li>Increase the number of training samples.</li> <li>Try a larger SegFormer model variant (explore available model variants on the <a href="https://huggingface.co/models?search=segformer" rel="nofollow">Hugging Face Hub</a>).</li> <li>Try different values for the arguments available in <code>LoraConfig</code>.</li> <li>Tune the learning rate and batch size.</li></ul> <p></p>
<script>
{
__sveltekit_c8wad6 = {
assets: "/docs/peft/main/en",
base: "/docs/peft/main/en",
env: {}
};
const element = document.currentScript.parentElement;
const data = [null,null];
Promise.all([
import("/docs/peft/main/en/_app/immutable/entry/start.f1bde2f6.js"),
import("/docs/peft/main/en/_app/immutable/entry/app.182b7f64.js")
]).then(([kit, app]) => {
kit.start(app, element, {
node_ids: [0, 39],
data,
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
72.3 kB
·
Xet hash:
50d8b906eb8587c31e4487212526ae1c4548284d33e3bb2ec6fd5671e57de994

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