Buckets:

hf-doc-build
/

doc-dev

Files
xet
hf-doc-build/doc-dev / transformers /pr_48873 /en /_app /immutable /nodes /102.BvCSqyO2.js
HuggingFaceDocBuilder
29 days ago
download
raw
33.3 kB
import"../chunks/DsnmJJEf.js";import{i as N,h as q,C as X,H as t,b as Y,a as d,D as n,E as S,s as E}from"../chunks/B89P_j9O.js";import{p as H,o as A,s as e,f as u,a as r,b as D,c as l,d as c,r as i,n as I}from"../chunks/v-ay6ctG.js";import{T as L}from"../chunks/DTuqSbyr.js";import{E as K}from"../chunks/opS7o4Lw.js";import{H as $}from"../chunks/CJWayN8L.js";const O='{"title":"ColPali","local":"colpali","sections":[{"title":"Notes","local":"notes","sections":[],"depth":2},{"title":"ColPaliConfig","local":"transformers.ColPaliConfig","sections":[],"depth":2},{"title":"ColPaliProcessor","local":"transformers.ColPaliProcessor","sections":[],"depth":2},{"title":"ColPaliForRetrieval","local":"transformers.ColPaliForRetrieval","sections":[],"depth":2}],"depth":1}';var ee=c('<meta name="hf:doc:metadata"/>'),oe=c("<!> <p>If you have issue with loading the images with PIL, you can use the following code to create dummy images:</p> <!>",1),se=c("<p>Example:</p> <!>",1),ae=c(`<p>Although the recipe for forward pass needs to be defined within this function, one should call the <code>Module</code> instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.</p>`),te=c(`<p></p> <p><em>This model was published in HF papers on 2024-06-27 and contributed to Hugging Face Transformers on 2024-12-17.</em></p> <!> <!> <p><a href="https://huggingface.co/papers/2407.01449" rel="nofollow">ColPali</a> is a model designed to retrieve documents by analyzing their visual features. Unlike traditional systems that rely heavily on text extraction and OCR, ColPali treats each page as an image. It uses <a href="./paligemma">Paligemma-3B</a> to capture not only text, but also the layout, tables, charts, and other visual elements to create detailed multi-vector embeddings that can be used for retrieval by computing pairwise late interaction similarity scores. This offers a more comprehensive understanding of documents and enables more efficient and accurate retrieval.</p> <p>This model was contributed by <a href="https://huggingface.co/tonywu71" rel="nofollow">@tonywu71</a> (ILLUIN Technology) and <a href="https://huggingface.co/yonigozlan" rel="nofollow">@yonigozlan</a> (HuggingFace).</p> <p>You can find all the original ColPali checkpoints under Vidore’s <a href="https://huggingface.co/collections/vidore/hf-native-colvision-models-6755d68fc60a8553acaa96f7" rel="nofollow">Hf-native ColVision Models</a> collection.</p> <blockquote class="tip"><p>Click on the ColPali models in the right sidebar for more examples of how to use ColPali for image retrieval.</p></blockquote> <!> <p>Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the <a href="../quantization/overview">Quantization</a> overview for more available quantization backends.</p> <p>The example below uses <a href="../quantization/bitsandbytes">bitsandbytes</a> to quantize the weights to int4.</p> <!> <!> <ul><li><code>score_retrieval()</code> returns a 2D tensor where the first dimension is the number of queries and the second dimension is the number of images. A higher score indicates more similarity between the query and image.</li></ul> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>This is the configuration class to store the configuration of a ColpaliModel. It is used to instantiate a Colpali
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the <a href="https://huggingface.co/vidore/colpali-v1.2" rel="nofollow">vidore/colpali-v1.2</a></p> <p>Configuration objects inherit from <a href="/docs/transformers/pr_48873/en/main_classes/configuration#transformers.PreTrainedConfig">PreTrainedConfig</a> and can be used to control the model outputs. Read the
documentation from <a href="/docs/transformers/pr_48873/en/main_classes/configuration#transformers.PreTrainedConfig">PreTrainedConfig</a> for more information.</p> <!></div> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>Constructs a ColPaliProcessor which wraps a image processor and a tokenizer into a single processor.</p> <p><a href="/docs/transformers/pr_48873/en/model_doc/colpali#transformers.ColPaliProcessor">ColPaliProcessor</a> offers all the functionalities of <a href="/docs/transformers/pr_48873/en/model_doc/siglip#transformers.SiglipImageProcessor">SiglipImageProcessor</a> and <code>tokenizer_class</code>. See the <a href="/docs/transformers/pr_48873/en/model_doc/siglip#transformers.SiglipImageProcessor">~SiglipImageProcessor</a> and <code>~tokenizer_class</code> for more information.</p> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!></div></div> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>The ColPali architecture leverages VLMs to construct efficient multi-vector embeddings directly
from document images (“screenshots”) for document retrieval. The model is trained to maximize the similarity
between these document embeddings and the corresponding query embeddings, using the late interaction method
introduced in ColBERT.</p> <p>Using ColPali removes the need for potentially complex and brittle layout recognition and OCR pipelines with a
single model that can take into account both the textual and visual content (layout, charts, etc.) of a document.</p> <p>ColPali is part of the ColVision model family, which was first introduced in the following paper: <a href="https://huggingface.co/papers/2407.01449" rel="nofollow"><em>ColPali: Efficient Document Retrieval with Vision Language Models</em></a>.</p> <p>This model inherits from <a href="/docs/transformers/pr_48873/en/main_classes/model#transformers.PreTrainedModel">PreTrainedModel</a>. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)</p> <p>This model is also a PyTorch <a href="https://pytorch.org/docs/stable/nn.html#torch.nn.Module" rel="nofollow">torch.nn.Module</a> subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.</p> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>The <a href="/docs/transformers/pr_48873/en/model_doc/colpali#transformers.ColPaliForRetrieval">ColPaliForRetrieval</a> forward method, overrides the <code>__call__</code> special method.</p> <!> <ul><li><p><strong>loss</strong> (<code>torch.FloatTensor</code> of shape <code>(1,)</code>, <em>optional</em>, returned when <code>labels</code> is provided) — Language modeling loss (for next-token prediction).</p></li> <li><p><strong>embeddings</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, sequence_length, hidden_size)</code>) — The embeddings of the model.</p></li> <li><p><strong>past_key_values</strong> (<code>Cache</code>, <em>optional</em>, returned when <code>use_cache=True</code> is passed or when <code>config.use_cache=True</code>) — It is a <a href="/docs/transformers/pr_48873/en/internal/generation_utils#transformers.Cache">Cache</a> instance. For more details, see our <a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>.</p> <p>Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see <code>past_key_values</code> input) to speed up sequential decoding.</p></li> <li><p><strong>hidden_states</strong> (<code>tuple[torch.FloatTensor]</code>, <em>optional</em>, returned when <code>output_hidden_states=True</code> is passed or when <code>config.output_hidden_states=True</code>) — Tuple of <code>torch.FloatTensor</code> (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape <code>(batch_size, sequence_length, hidden_size)</code>.</p> <p>Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.</p></li> <li><p><strong>attentions</strong> (<code>tuple[torch.FloatTensor]</code>, <em>optional</em>, returned when <code>output_attentions=True</code> is passed or when <code>config.output_attentions=True</code>) — Tuple of <code>torch.FloatTensor</code> (one for each layer) of shape <code>(batch_size, num_heads, sequence_length, sequence_length)</code>.</p> <p>Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.</p></li> <li><p><strong>image_hidden_states</strong> (<code>torch.FloatTensor</code>, <em>optional</em>) — A <code>torch.FloatTensor</code> of size <code>(batch_size, num_images, sequence_length, hidden_size)</code>.
image_hidden_states of the model produced by the vision encoder after projecting last hidden state.</p></li></ul></div></div> <!> <p></p>`,1);function pe(P,x){H(x,!1),A(()=>{new URLSearchParams(window.location.search).get("fw")}),N();var g=te();q("1femb22",o=>{var a=ee();E(a,"content",O),r(o,a)});var f=e(u(g),4);X(f,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var y=e(f,2);t(y,{title:"ColPali",local:"colpali",headingTag:"h1"});var M=e(y,10);Y(M,{id:"usage",options:["image retrieval"],children:(o,a)=>{$(o,{id:"usage",option:"image retrieval",children:(s,k)=>{var B=oe(),R=u(B);d(R,{code:"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",highlighted:`<span class="hljs-keyword">import</span> requests
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> PIL <span class="hljs-keyword">import</span> Image
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> ColPaliForRetrieval, ColPaliProcessor
<span class="hljs-comment"># Load the model and the processor</span>
model_name = <span class="hljs-string">&quot;vidore/colpali-v1.3-hf&quot;</span>
model = ColPaliForRetrieval.from_pretrained(
    model_name,
    device_map=<span class="hljs-string">&quot;auto&quot;</span>,  <span class="hljs-comment"># &quot;cpu&quot;, &quot;cuda&quot;, &quot;xpu&quot;, or &quot;mps&quot; for Apple Silicon</span>
)
processor = ColPaliProcessor.from_pretrained(model_name)
<span class="hljs-comment"># The document page screenshots from your corpus</span>
url1 = <span class="hljs-string">&quot;https://upload.wikimedia.org/wikipedia/commons/8/89/US-original-Declaration-1776.jpg&quot;</span>
url2 = <span class="hljs-string">&quot;https://upload.wikimedia.org/wikipedia/commons/thumb/4/4c/Romeoandjuliet1597.jpg/500px-Romeoandjuliet1597.jpg&quot;</span>
images = [
    Image.<span class="hljs-built_in">open</span>(requests.get(url1, stream=<span class="hljs-literal">True</span>).raw),
    Image.<span class="hljs-built_in">open</span>(requests.get(url2, stream=<span class="hljs-literal">True</span>).raw),
]
<span class="hljs-comment"># The queries you want to retrieve documents for</span>
queries = [
    <span class="hljs-string">&quot;When was the United States Declaration of Independence proclaimed?&quot;</span>,
    <span class="hljs-string">&quot;Who printed the edition of Romeo and Juliet?&quot;</span>,
]
<span class="hljs-comment"># Process the inputs</span>
inputs_images = processor(images=images).to(model.device)
inputs_text = processor(text=queries).to(model.device)
<span class="hljs-comment"># Forward pass</span>
<span class="hljs-keyword">with</span> torch.no_grad():
    image_embeddings = model(**inputs_images).embeddings
    query_embeddings = model(**inputs_text).embeddings
<span class="hljs-comment"># Score the queries against the images</span>
scores = processor.score_retrieval(query_embeddings, image_embeddings)
<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Retrieval scores (query x image):&quot;</span>)
<span class="hljs-built_in">print</span>(scores)`,lang:"python",wrap:!1});var Q=e(R,4);d(Q,{code:"aW1hZ2VzJTIwJTNEJTIwJTVCJTBBJTIwJTIwJTIwJTIwSW1hZ2UubmV3KCUyMlJHQiUyMiUyQyUyMCgxMjglMkMlMjAxMjgpJTJDJTIwY29sb3IlM0QlMjJ3aGl0ZSUyMiklMkMlMEElMjAlMjAlMjAlMjBJbWFnZS5uZXcoJTIyUkdCJTIyJTJDJTIwKDY0JTJDJTIwMzIpJTJDJTIwY29sb3IlM0QlMjJibGFjayUyMiklMkMlMEElNUQ=",highlighted:`images = [
    Image.new(<span class="hljs-string">&quot;RGB&quot;</span>, (<span class="hljs-number">128</span>, <span class="hljs-number">128</span>), color=<span class="hljs-string">&quot;white&quot;</span>),
    Image.new(<span class="hljs-string">&quot;RGB&quot;</span>, (<span class="hljs-number">64</span>, <span class="hljs-number">32</span>), color=<span class="hljs-string">&quot;black&quot;</span>),
]`,lang:"python",wrap:!1}),r(s,B)},$$slots:{default:!0}})},$$slots:{default:!0}});var b=e(M,6);d(b,{code:"aW1wb3J0JTIwcmVxdWVzdHMlMEFpbXBvcnQlMjB0b3JjaCUwQWZyb20lMjBQSUwlMjBpbXBvcnQlMjBJbWFnZSUwQSUwQWZyb20lMjB0cmFuc2Zvcm1lcnMlMjBpbXBvcnQlMjBCaXRzQW5kQnl0ZXNDb25maWclMkMlMjBDb2xQYWxpRm9yUmV0cmlldmFsJTJDJTIwQ29sUGFsaVByb2Nlc3NvciUwQSUwQSUwQW1vZGVsX25hbWUlMjAlM0QlMjAlMjJ2aWRvcmUlMkZjb2xwYWxpLXYxLjMtaGYlMjIlMEElMEElMjMlMjA0LWJpdCUyMHF1YW50aXphdGlvbiUyMGNvbmZpZ3VyYXRpb24lMEFibmJfY29uZmlnJTIwJTNEJTIwQml0c0FuZEJ5dGVzQ29uZmlnKCUwQSUyMCUyMCUyMCUyMGxvYWRfaW5fNGJpdCUzRFRydWUlMkMlMEElMjAlMjAlMjAlMjBibmJfNGJpdF91c2VfZG91YmxlX3F1YW50JTNEVHJ1ZSUyQyUwQSUyMCUyMCUyMCUyMGJuYl80Yml0X3F1YW50X3R5cGUlM0QlMjJuZjQlMjIlMkMlMEElMjAlMjAlMjAlMjBibmJfNGJpdF9jb21wdXRlX2R0eXBlJTNEdG9yY2guZmxvYXQxNiUyQyUwQSklMEElMEFtb2RlbCUyMCUzRCUyMENvbFBhbGlGb3JSZXRyaWV2YWwuZnJvbV9wcmV0cmFpbmVkKCUwQSUyMCUyMCUyMCUyMG1vZGVsX25hbWUlMkMlMEElMjAlMjAlMjAlMjBxdWFudGl6YXRpb25fY29uZmlnJTNEYm5iX2NvbmZpZyUyQyUwQSUyMCUyMCUyMCUyMGRldmljZV9tYXAlM0QlMjJhdXRvJTIyJTJDJTBBKSUwQSUwQXByb2Nlc3NvciUyMCUzRCUyMENvbFBhbGlQcm9jZXNzb3IuZnJvbV9wcmV0cmFpbmVkKG1vZGVsX25hbWUpJTBBJTBBdXJsMSUyMCUzRCUyMCUyMmh0dHBzJTNBJTJGJTJGdXBsb2FkLndpa2ltZWRpYS5vcmclMkZ3aWtpcGVkaWElMkZjb21tb25zJTJGOCUyRjg5JTJGVVMtb3JpZ2luYWwtRGVjbGFyYXRpb24tMTc3Ni5qcGclMjIlMEF1cmwyJTIwJTNEJTIwJTIyaHR0cHMlM0ElMkYlMkZ1cGxvYWQud2lraW1lZGlhLm9yZyUyRndpa2lwZWRpYSUyRmNvbW1vbnMlMkZ0aHVtYiUyRjQlMkY0YyUyRlJvbWVvYW5kanVsaWV0MTU5Ny5qcGclMkY1MDBweC1Sb21lb2FuZGp1bGlldDE1OTcuanBnJTIyJTBBJTBBaW1hZ2VzJTIwJTNEJTIwJTVCJTBBJTIwJTIwJTIwJTIwSW1hZ2Uub3BlbihyZXF1ZXN0cy5nZXQodXJsMSUyQyUyMHN0cmVhbSUzRFRydWUpLnJhdyklMkMlMEElMjAlMjAlMjAlMjBJbWFnZS5vcGVuKHJlcXVlc3RzLmdldCh1cmwyJTJDJTIwc3RyZWFtJTNEVHJ1ZSkucmF3KSUyQyUwQSU1RCUwQSUwQXF1ZXJpZXMlMjAlM0QlMjAlNUIlMEElMjAlMjAlMjAlMjAlMjJXaGVuJTIwd2FzJTIwdGhlJTIwVW5pdGVkJTIwU3RhdGVzJTIwRGVjbGFyYXRpb24lMjBvZiUyMEluZGVwZW5kZW5jZSUyMHByb2NsYWltZWQlM0YlMjIlMkMlMEElMjAlMjAlMjAlMjAlMjJXaG8lMjBwcmludGVkJTIwdGhlJTIwZWRpdGlvbiUyMG9mJTIwUm9tZW8lMjBhbmQlMjBKdWxpZXQlM0YlMjIlMkMlMEElNUQlMEElMEElMjMlMjBQcm9jZXNzJTIwdGhlJTIwaW5wdXRzJTBBaW5wdXRzX2ltYWdlcyUyMCUzRCUyMHByb2Nlc3NvcihpbWFnZXMlM0RpbWFnZXMlMkMlMjByZXR1cm5fdGVuc29ycyUzRCUyMnB0JTIyKS50byhtb2RlbC5kZXZpY2UpJTBBaW5wdXRzX3RleHQlMjAlM0QlMjBwcm9jZXNzb3IodGV4dCUzRHF1ZXJpZXMlMkMlMjByZXR1cm5fdGVuc29ycyUzRCUyMnB0JTIyKS50byhtb2RlbC5kZXZpY2UpJTBBJTBBJTIzJTIwRm9yd2FyZCUyMHBhc3MlMEF3aXRoJTIwdG9yY2gubm9fZ3JhZCgpJTNBJTBBJTIwJTIwJTIwJTIwaW1hZ2VfZW1iZWRkaW5ncyUyMCUzRCUyMG1vZGVsKCoqaW5wdXRzX2ltYWdlcykuZW1iZWRkaW5ncyUwQSUyMCUyMCUyMCUyMHF1ZXJ5X2VtYmVkZGluZ3MlMjAlM0QlMjBtb2RlbCgqKmlucHV0c190ZXh0KS5lbWJlZGRpbmdzJTBBJTBBJTIzJTIwU2NvcmUlMjB0aGUlMjBxdWVyaWVzJTIwYWdhaW5zdCUyMHRoZSUyMGltYWdlcyUwQXNjb3JlcyUyMCUzRCUyMHByb2Nlc3Nvci5zY29yZV9yZXRyaWV2YWwocXVlcnlfZW1iZWRkaW5ncyUyQyUyMGltYWdlX2VtYmVkZGluZ3MpJTBBJTBBcHJpbnQoJTIyUmV0cmlldmFsJTIwc2NvcmVzJTIwKHF1ZXJ5JTIweCUyMGltYWdlKSUzQSUyMiklMEFwcmludChzY29yZXMp",highlighted:`<span class="hljs-keyword">import</span> requests
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> PIL <span class="hljs-keyword">import</span> Image
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BitsAndBytesConfig, ColPaliForRetrieval, ColPaliProcessor
model_name = <span class="hljs-string">&quot;vidore/colpali-v1.3-hf&quot;</span>
<span class="hljs-comment"># 4-bit quantization configuration</span>
bnb_config = BitsAndBytesConfig(
    load_in_4bit=<span class="hljs-literal">True</span>,
    bnb_4bit_use_double_quant=<span class="hljs-literal">True</span>,
    bnb_4bit_quant_type=<span class="hljs-string">&quot;nf4&quot;</span>,
    bnb_4bit_compute_dtype=torch.float16,
)
model = ColPaliForRetrieval.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map=<span class="hljs-string">&quot;auto&quot;</span>,
)
processor = ColPaliProcessor.from_pretrained(model_name)
url1 = <span class="hljs-string">&quot;https://upload.wikimedia.org/wikipedia/commons/8/89/US-original-Declaration-1776.jpg&quot;</span>
url2 = <span class="hljs-string">&quot;https://upload.wikimedia.org/wikipedia/commons/thumb/4/4c/Romeoandjuliet1597.jpg/500px-Romeoandjuliet1597.jpg&quot;</span>
images = [
    Image.<span class="hljs-built_in">open</span>(requests.get(url1, stream=<span class="hljs-literal">True</span>).raw),
    Image.<span class="hljs-built_in">open</span>(requests.get(url2, stream=<span class="hljs-literal">True</span>).raw),
]
queries = [
    <span class="hljs-string">&quot;When was the United States Declaration of Independence proclaimed?&quot;</span>,
    <span class="hljs-string">&quot;Who printed the edition of Romeo and Juliet?&quot;</span>,
]
<span class="hljs-comment"># Process the inputs</span>
inputs_images = processor(images=images, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>).to(model.device)
inputs_text = processor(text=queries, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>).to(model.device)
<span class="hljs-comment"># Forward pass</span>
<span class="hljs-keyword">with</span> torch.no_grad():
    image_embeddings = model(**inputs_images).embeddings
    query_embeddings = model(**inputs_text).embeddings
<span class="hljs-comment"># Score the queries against the images</span>
scores = processor.score_retrieval(query_embeddings, image_embeddings)
<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Retrieval scores (query x image):&quot;</span>)
<span class="hljs-built_in">print</span>(scores)`,lang:"python",wrap:!1});var _=e(b,2);t(_,{title:"Notes",local:"notes",headingTag:"h2"});var w=e(_,4);t(w,{title:"ColPaliConfig",local:"transformers.ColPaliConfig",headingTag:"h2"});var p=e(w,2),U=l(p);n(U,{name:"class transformers.ColPaliConfig",anchor:"transformers.ColPaliConfig",source:"https://github.com/huggingface/transformers/blob/vr_48873/src/transformers/models/colpali/configuration_colpali.py#L28",parameters:[{name:"transformers_version",val:": str | None = None"},{name:"architectures",val:": list[str] | None = None"},{name:"output_hidden_states",val:": bool | None = False"},{name:"return_dict",val:": bool | None = True"},{name:"dtype",val:": str | torch.dtype | None = None"},{name:"chunk_size_feed_forward",val:": int = 0"},{name:"is_encoder_decoder",val:": bool = False"},{name:"id2label",val:": dict[int, str] | dict[str, str] | None = None"},{name:"label2id",val:": dict[str, int] | dict[str, str] | None = None"},{name:"problem_type",val:": Literal['regression', 'single_label_classification', 'multi_label_classification'] | None = None"},{name:"vlm_config",val:": dict | transformers.configuration_utils.PreTrainedConfig | None = None"},{name:"text_config",val:": dict | transformers.configuration_utils.PreTrainedConfig | None = None"},{name:"embedding_dim",val:": int = 128"}],parametersDescription:[{anchor:"transformers.ColPaliConfig.vlm_config",description:`<strong>vlm_config</strong> (<code>Union[dict, ~configuration_utils.PreTrainedConfig]</code>, <em>optional</em>) &#x2014;
The config object or dictionary of the vision-language backbone.`,name:"vlm_config"},{anchor:"transformers.ColPaliConfig.text_config",description:`<strong>text_config</strong> (<code>Union[dict, ~configuration_utils.PreTrainedConfig]</code>, <em>optional</em>) &#x2014;
The config object or dictionary of the text backbone.`,name:"text_config"},{anchor:"transformers.ColPaliConfig.embedding_dim",description:`<strong>embedding_dim</strong> (<code>int</code>, <em>optional</em>, defaults to <code>128</code>) &#x2014;
Dimensionality of the embeddings and hidden states.`,name:"embedding_dim"}]});var F=e(U,6);K(F,{anchor:"transformers.ColPaliConfig.example",children:(o,a)=>{var s=se(),k=e(u(s),2);d(k,{code:"ZnJvbSUyMHRyYW5zZm9ybWVycy5tb2RlbHMuY29scGFsaSUyMGltcG9ydCUyMENvbFBhbGlDb25maWclMkMlMjBDb2xQYWxpRm9yUmV0cmlldmFsJTBBJTBBY29uZmlnJTIwJTNEJTIwQ29sUGFsaUNvbmZpZygpJTBBbW9kZWwlMjAlM0QlMjBDb2xQYWxpRm9yUmV0cmlldmFsKGNvbmZpZyk=",highlighted:`<span class="hljs-keyword">from</span> transformers.models.colpali <span class="hljs-keyword">import</span> ColPaliConfig, ColPaliForRetrieval
config = ColPaliConfig()
model = ColPaliForRetrieval(config)`,lang:"python",wrap:!1}),r(o,s)},$$slots:{default:!0}}),i(p);var T=e(p,2);t(T,{title:"ColPaliProcessor",local:"transformers.ColPaliProcessor",headingTag:"h2"});var m=e(T,2),v=l(m);n(v,{name:"class transformers.ColPaliProcessor",anchor:"transformers.ColPaliProcessor",source:"https://github.com/huggingface/transformers/blob/vr_48873/src/transformers/models/colpali/processing_colpali.py#L54",parameters:[{name:"image_processor",val:" = None"},{name:"tokenizer",val:" = None"},{name:"chat_template",val:" = None"},{name:"visual_prompt_prefix",val:": str = 'Describe the image.'"},{name:"query_prefix",val:": str = 'Question: '"}],parametersDescription:[{anchor:"transformers.ColPaliProcessor.image_processor",description:`<strong>image_processor</strong> (<code>SiglipImageProcessor</code>) &#x2014;
The image processor is a required input.`,name:"image_processor"},{anchor:"transformers.ColPaliProcessor.tokenizer",description:`<strong>tokenizer</strong> (<code>tokenizer_class</code>) &#x2014;
The tokenizer is a required input.`,name:"tokenizer"},{anchor:"transformers.ColPaliProcessor.chat_template",description:`<strong>chat_template</strong> (<code>str</code>) &#x2014;
A Jinja template to convert lists of messages in a chat into a tokenizable string.`,name:"chat_template"},{anchor:"transformers.ColPaliProcessor.visual_prompt_prefix",description:`<strong>visual_prompt_prefix</strong> (<code>str</code>, <em>optional</em>, defaults to <code>&quot;Describe the image.&quot;</code>) &#x2014;
A string that gets tokenized and prepended to the image tokens.`,name:"visual_prompt_prefix"},{anchor:"transformers.ColPaliProcessor.query_prefix",description:`<strong>query_prefix</strong> (<code>str</code>, <em>optional</em>, defaults to <code>&quot;Question -- &quot;</code>):
A prefix to be used for the query.`,name:"query_prefix"}]});var J=e(v,6),G=l(J);n(G,{name:"__call__",anchor:"transformers.ColPaliProcessor.__call__",source:"https://github.com/huggingface/transformers/blob/vr_48873/src/transformers/models/colpali/processing_colpali.py#L94",parameters:[{name:"images",val:": typing.Union[ForwardRef('PIL.Image.Image'), numpy.ndarray, ForwardRef('torch.Tensor'), list['PIL.Image.Image'], list[numpy.ndarray], list['torch.Tensor'], NoneType] = None"},{name:"text",val:": str | list[str] | list[list[str]] = None"},{name:"**kwargs",val:": Unpack"}],parametersDescription:[{anchor:"transformers.ColPaliProcessor.__call__.images",description:`<strong>images</strong> (<code>Union[PIL.Image.Image, numpy.ndarray, torch.Tensor, list[PIL.Image.Image], list[numpy.ndarray], list[torch.Tensor]]</code>, <em>optional</em>) &#x2014;
Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
passing in images with pixel values between 0 and 1, set <code>do_rescale=False</code>.`,name:"images"},{anchor:"transformers.ColPaliProcessor.__call__.text",description:`<strong>text</strong> (<code>Union[str, list[str], list[list[str]]]</code>, <em>optional</em>) &#x2014;
The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
(pretokenized string). If you pass a pretokenized input, set <code>is_split_into_words=True</code> to avoid ambiguity with batched inputs.`,name:"text"},{anchor:"transformers.ColPaliProcessor.__call__.return_tensors",description:`<strong>return_tensors</strong> (<code>str</code> or <a href="/docs/transformers/pr_48873/en/internal/file_utils#transformers.TensorType">TensorType</a>, <em>optional</em>) &#x2014;
If set, will return tensors of a particular framework. Acceptable values are:</p>
<ul>
<li><code>&apos;pt&apos;</code>: Return PyTorch <code>torch.Tensor</code> objects.</li>
<li><code>&apos;np&apos;</code>: Return NumPy <code>np.ndarray</code> objects.</li>
</ul>`,name:"return_tensors"},{anchor:"transformers.ColPaliProcessor.__call__.*kwargs",description:`*<strong>*kwargs</strong> (<a href="/docs/transformers/pr_48873/en/main_classes/processors#transformers.ProcessingKwargs">ProcessingKwargs</a>, <em>optional</em>) &#x2014;
Additional processing options for each modality (text, images, videos, audio). Model-specific parameters
are listed above; see the TypedDict class for the complete list of supported arguments.`,name:"*kwargs"}],returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>
<p>A <a
  href="/docs/transformers/pr_48873/en/main_classes/image_processor#transformers.BatchFeature"
>BatchFeature</a> with the following fields:</p>
<ul>
<li><strong>input_ids</strong> — List of token ids to be fed to a model. Returned when <code>text</code> is not <code>None</code>. If <code>suffix</code>
is provided, the <code>input_ids</code> will also contain the suffix input ids.</li>
<li><strong>attention_mask</strong> — List of indices specifying which tokens should be attended to by the model (when
<code>return_attention_mask=True</code> or if <em>“attention_mask”</em> is in <code>self.model_input_names</code> and if <code>text</code> is not
<code>None</code>).</li>
<li><strong>pixel_values</strong> — Pixel values to be fed to a model. Returned when <code>images</code> is not <code>None</code>.</li>
<li><strong>labels</strong> — Labels compatible with training if <code>suffix</code> is not None</li>
</ul>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>
<p><a
  href="/docs/transformers/pr_48873/en/main_classes/image_processor#transformers.BatchFeature"
>BatchFeature</a></p>
`}),i(J),i(m);var j=e(m,2);t(j,{title:"ColPaliForRetrieval",local:"transformers.ColPaliForRetrieval",headingTag:"h2"});var h=e(j,2),C=l(h);n(C,{name:"class transformers.ColPaliForRetrieval",anchor:"transformers.ColPaliForRetrieval",source:"https://github.com/huggingface/transformers/blob/vr_48873/src/transformers/models/colpali/modeling_colpali.py#L90",parameters:[{name:"config",val:": ColPaliConfig"}],parametersDescription:[{anchor:"transformers.ColPaliForRetrieval.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_48873/en/model_doc/colpali#transformers.ColPaliConfig">ColPaliConfig</a>) &#x2014;
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
<a href="/docs/transformers/pr_48873/en/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"}]});var Z=e(C,12),W=l(Z);n(W,{name:"forward",anchor:"transformers.ColPaliForRetrieval.forward",source:"https://github.com/huggingface/transformers/blob/vr_48873/src/transformers/models/colpali/modeling_colpali.py#L108",parameters:[{name:"input_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"pixel_values",val:": typing.Optional[torch.FloatTensor] = None"},{name:"attention_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"**kwargs",val:": Unpack"}],parametersDescription:[{anchor:"transformers.ColPaliForRetrieval.forward.input_ids",description:`<strong>input_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_48873/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_48873/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_48873/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.__call__">PreTrainedTokenizer.<strong>call</strong>()</a> for details.</p>
<p><a href="../glossary#input-ids">What are input IDs?</a>`,name:"input_ids"},{anchor:"transformers.ColPaliForRetrieval.forward.pixel_values",description:`<strong>pixel_values</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, num_channels, image_size, image_size)</code>, <em>optional</em>) &#x2014;
The tensors corresponding to the input images. Pixel values can be obtained using
<code>image_processor_class</code>. See <code>image_processor_class.__call__</code> for details (<code>processor_class</code> uses
<code>image_processor_class</code> for processing images).`,name:"pixel_values"},{anchor:"transformers.ColPaliForRetrieval.forward.attention_mask",description:`<strong>attention_mask</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Mask to avoid performing attention on padding token indices. Mask values selected in <code>[0, 1]</code>:</p>
<ul>
<li>1 for tokens that are <strong>not masked</strong>,</li>
<li>0 for tokens that are <strong>masked</strong>.</li>
</ul>
<p><a href="../glossary#attention-mask">What are attention masks?</a>`,name:"attention_mask"}],returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>
<p>A <code>ColPaliForRetrievalOutput</code> or a tuple of
<code>torch.FloatTensor</code> (if <code>return_dict=False</code> is passed or when <code>config.return_dict=False</code>) comprising various
elements depending on the configuration (<code>None</code>) and inputs.</p>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>
<p><code>ColPaliForRetrievalOutput</code> or <code>tuple(torch.FloatTensor)</code></p>
`});var V=e(W,4);L(V,{children:(o,a)=>{var s=ae();r(o,s)},$$slots:{default:!0}}),I(2),i(Z),i(h);var z=e(h,2);S(z,{source:"https://github.com/huggingface/transformers/blob/main/docs/source/en/model_doc/colpali.md"}),I(2),r(P,g),D()}export{pe as component};
Xet Storage Details

Size:: 33.3 kB
Xet hash:: 1444371b862135e680bc998149e994f8c96fa3b200aef7b8a3e869e7b36ccd99
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.