diffusers/pr_13966/en/_app/immutable/nodes/94.sjczvbvv.js

import"../chunks/DsnmJJEf.js";import{i as x,h as w,C as k,H as t,D as s,E as N,s as O}from"../chunks/BtE7mKSK.js";import{p as L,o as z,s as e,f as A,a as g,b as C,c as r,d as b,n as a,r as i}from"../chunks/jDjavuwI.js";const q='{"title":"Transformer2DModel","local":"transformer2dmodel","sections":[{"title":"Transformer2DModel","local":"diffusers.Transformer2DModel","sections":[],"depth":2},{"title":"Transformer2DModelOutput","local":"diffusers.models.modeling_outputs.Transformer2DModelOutput","sections":[],"depth":2}],"depth":1}';var I=b('<meta name="hf:doc:metadata"/>'),P=b('<p></p> <!> <!> <p>A Transformer model for image-like data from <a href="https://huggingface.co/CompVis" rel="nofollow">CompVis</a> that is based on the <a href="https://huggingface.co/papers/2010.11929" rel="nofollow">Vision Transformer</a> introduced by Dosovitskiy et al. The <a href="/docs/diffusers/pr_13966/en/api/models/transformer2d#diffusers.Transformer2DModel">Transformer2DModel</a> accepts discrete (classes of vector embeddings) or continuous (actual embeddings) inputs.</p> <p>When the input is <strong>continuous</strong>:</p> <ol><li>Project the input and reshape it to <code>(batch_size, sequence_length, feature_dimension)</code>.</li> <li>Apply the Transformer blocks in the standard way.</li> <li>Reshape to image.</li></ol> <p>When the input is <strong>discrete</strong>:</p> <blockquote class="tip"><p>It is assumed one of the input classes is the masked latent pixel. The predicted classes of the unnoised image don’t contain a prediction for the masked pixel because the unnoised image cannot be masked.</p></blockquote> <ol><li>Convert input (classes of latent pixels) to embeddings and apply positional embeddings.</li> <li>Apply the Transformer blocks in the standard way.</li> <li>Predict classes of unnoised image.</li></ol> <!> <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>A 2D Transformer model for image-like data.</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/diffusers/pr_13966/en/api/models/transformer2d#diffusers.Transformer2DModel">Transformer2DModel</a> forward method.</p></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 output of <a href="/docs/diffusers/pr_13966/en/api/models/transformer2d#diffusers.Transformer2DModel">Transformer2DModel</a>.</p></div> <!> <p></p>',1);function W(T,v){L(v,!1),z(()=>{new URLSearchParams(window.location.search).get("fw")}),x();var d=P();w("pau42",h=>{var _=I();O(_,"content",q),g(h,_)});var c=e(A(d),2);k(c,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var l=e(c,2);t(l,{title:"Transformer2DModel",local:"transformer2dmodel",headingTag:"h1"});var m=e(l,14);t(m,{title:"Transformer2DModel",local:"diffusers.Transformer2DModel",headingTag:"h2"});var o=e(m,2),f=r(o);s(f,{name:"class diffusers.Transformer2DModel",anchor:"diffusers.Transformer2DModel",source:"https://github.com/huggingface/diffusers/blob/vr_13966/src/diffusers/models/transformers/transformer_2d.py#L39",parameters:[{name:"num_attention_heads",val:": int = 16"},{name:"attention_head_dim",val:": int = 88"},{name:"in_channels",val:": int | None = None"},{name:"out_channels",val:": int | None = None"},{name:"num_layers",val:": int = 1"},{name:"dropout",val:": float = 0.0"},{name:"norm_num_groups",val:": int = 32"},{name:"cross_attention_dim",val:": int | None = None"},{name:"attention_bias",val:": bool = False"},{name:"sample_size",val:": int | None = None"},{name:"num_vector_embeds",val:": int | None = None"},{name:"patch_size",val:": int | None = None"},{name:"activation_fn",val:": str = 'geglu'"},{name:"num_embeds_ada_norm",val:": int | None = None"},{name:"use_linear_projection",val:": bool = False"},{name:"only_cross_attention",val:": bool = False"},{name:"double_self_attention",val:": bool = False"},{name:"upcast_attention",val:": bool = False"},{name:"norm_type",val:": str = 'layer_norm'"},{name:"norm_elementwise_affine",val:": bool = True"},{name:"norm_eps",val:": float = 1e-05"},{name:"attention_type",val:": str = 'default'"},{name:"caption_channels",val:": int = None"},{name:"interpolation_scale",val:": float = None"},{name:"use_additional_conditions",val:": bool | None = None"}],parametersDescription:[{anchor:"diffusers.Transformer2DModel.num_attention_heads",description:"<strong>num_attention_heads</strong> (<code>int</code>, <em>optional</em>, defaults to 16) &#x2014; The number of heads to use for multi-head attention.",name:"num_attention_heads"},{anchor:"diffusers.Transformer2DModel.attention_head_dim",description:"<strong>attention_head_dim</strong> (<code>int</code>, <em>optional</em>, defaults to 88) &#x2014; The number of channels in each head.",name:"attention_head_dim"},{anchor:"diffusers.Transformer2DModel.in_channels",description:`<strong>in_channels</strong> (<code>int</code>, <em>optional</em>) &#x2014;
The number of channels in the input and output (specify if the input is <strong>continuous</strong>).`,name:"in_channels"},{anchor:"diffusers.Transformer2DModel.num_layers",description:"<strong>num_layers</strong> (<code>int</code>, <em>optional</em>, defaults to 1) &#x2014; The number of layers of Transformer blocks to use.",name:"num_layers"},{anchor:"diffusers.Transformer2DModel.dropout",description:"<strong>dropout</strong> (<code>float</code>, <em>optional</em>, defaults to 0.0) &#x2014; The dropout probability to use.",name:"dropout"},{anchor:"diffusers.Transformer2DModel.cross_attention_dim",description:"<strong>cross_attention_dim</strong> (<code>int</code>, <em>optional</em>) &#x2014; The number of <code>encoder_hidden_states</code> dimensions to use.",name:"cross_attention_dim"},{anchor:"diffusers.Transformer2DModel.sample_size",description:`<strong>sample_size</strong> (<code>int</code>, <em>optional</em>) &#x2014; The width of the latent images (specify if the input is <strong>discrete</strong>).
This is fixed during training since it is used to learn a number of position embeddings.`,name:"sample_size"},{anchor:"diffusers.Transformer2DModel.num_vector_embeds",description:`<strong>num_vector_embeds</strong> (<code>int</code>, <em>optional</em>) &#x2014;
The number of classes of the vector embeddings of the latent pixels (specify if the input is <strong>discrete</strong>).
Includes the class for the masked latent pixel.`,name:"num_vector_embeds"},{anchor:"diffusers.Transformer2DModel.activation_fn",description:"<strong>activation_fn</strong> (<code>str</code>, <em>optional</em>, defaults to <code>&quot;geglu&quot;</code>) &#x2014; Activation function to use in feed-forward.",name:"activation_fn"},{anchor:"diffusers.Transformer2DModel.num_embeds_ada_norm",description:`<strong>num_embeds_ada_norm</strong> ( <code>int</code>, <em>optional</em>) &#x2014;
The number of diffusion steps used during training. Pass if at least one of the norm_layers is
<code>AdaLayerNorm</code>. This is fixed during training since it is used to learn a number of embeddings that are
added to the hidden states.</p>
<p>During inference, you can denoise for up to but not more steps than <code>num_embeds_ada_norm</code>.`,name:"num_embeds_ada_norm"},{anchor:"diffusers.Transformer2DModel.attention_bias",description:`<strong>attention_bias</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Configure if the <code>TransformerBlocks</code> attention should contain a bias parameter.`,name:"attention_bias"}]});var p=e(f,4),D=r(p);s(D,{name:"forward",anchor:"diffusers.Transformer2DModel.forward",source:"https://github.com/huggingface/diffusers/blob/vr_13966/src/diffusers/models/transformers/transformer_2d.py#L324",parameters:[{name:"hidden_states",val:": Tensor"},{name:"encoder_hidden_states",val:": typing.Optional[torch.Tensor] = None"},{name:"timestep",val:": typing.Optional[torch.LongTensor] = None"},{name:"added_cond_kwargs",val:": dict = None"},{name:"class_labels",val:": typing.Optional[torch.LongTensor] = None"},{name:"cross_attention_kwargs",val:": dict = None"},{name:"attention_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"encoder_attention_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"return_dict",val:": bool = True"}],parametersDescription:[{anchor:"diffusers.Transformer2DModel.forward.hidden_states",description:`<strong>hidden_states</strong> (<code>torch.LongTensor</code> of shape <code>(batch size, num latent pixels)</code> if discrete, <code>torch.Tensor</code> of shape <code>(batch size, channel, height, width)</code> if continuous) &#x2014;
Input <code>hidden_states</code>.`,name:"hidden_states"},{anchor:"diffusers.Transformer2DModel.forward.encoder_hidden_states",description:`<strong>encoder_hidden_states</strong> ( <code>torch.Tensor</code> of shape <code>(batch size, sequence len, embed dims)</code>, <em>optional</em>) &#x2014;
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
self-attention.`,name:"encoder_hidden_states"},{anchor:"diffusers.Transformer2DModel.forward.timestep",description:`<strong>timestep</strong> ( <code>torch.LongTensor</code>, <em>optional</em>) &#x2014;
Used to indicate denoising step. Optional timestep to be applied as an embedding in <code>AdaLayerNorm</code>.`,name:"timestep"},{anchor:"diffusers.Transformer2DModel.forward.class_labels",description:`<strong>class_labels</strong> ( <code>torch.LongTensor</code> of shape <code>(batch size, num classes)</code>, <em>optional</em>) &#x2014;
Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
<code>AdaLayerZeroNorm</code>.`,name:"class_labels"},{anchor:"diffusers.Transformer2DModel.forward.cross_attention_kwargs",description:`<strong>cross_attention_kwargs</strong> ( <code>dict[str, Any]</code>, <em>optional</em>) &#x2014;
A kwargs dictionary that if specified is passed along to the <code>AttentionProcessor</code> as defined under
<code>self.processor</code> in
<a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py" rel="nofollow">diffusers.models.attention_processor</a>.`,name:"cross_attention_kwargs"},{anchor:"diffusers.Transformer2DModel.forward.attention_mask",description:`<strong>attention_mask</strong> ( <code>torch.Tensor</code>, <em>optional</em>) &#x2014;
An attention mask of shape <code>(batch, key_tokens)</code> is applied to <code>encoder_hidden_states</code>. If <code>1</code> the mask
is kept, otherwise if <code>0</code> it is discarded. Mask will be converted into a bias, which adds large
negative values to the attention scores corresponding to &#x201C;discard&#x201D; tokens.`,name:"attention_mask"},{anchor:"diffusers.Transformer2DModel.forward.encoder_attention_mask",description:`<strong>encoder_attention_mask</strong> ( <code>torch.Tensor</code>, <em>optional</em>) &#x2014;
Cross-attention mask applied to <code>encoder_hidden_states</code>. Two formats supported:</p>
<ul>
<li>Mask <code>(batch, sequence_length)</code> True = keep, False = discard.</li>
<li>Bias <code>(batch, 1, sequence_length)</code> 0 = keep, -10000 = discard.</li>
</ul>
<p>If <code>ndim == 2</code>: will be interpreted as a mask, then converted into a bias consistent with the format
above. This bias will be added to the cross-attention scores.`,name:"encoder_attention_mask"},{anchor:"diffusers.Transformer2DModel.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) &#x2014;
Whether or not to return a <a href="/docs/diffusers/pr_13966/en/api/models/unet2d-cond#diffusers.models.unets.unet_2d_condition.UNet2DConditionOutput">UNet2DConditionOutput</a> instead of a plain
tuple.`,name:"return_dict"}],returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>If <code>return_dict</code> is True, an <code>Transformer2DModelOutput</code> is returned,
otherwise a <code>tuple</code> where the first element is the sample tensor.</p>
`}),a(2),i(p),i(o);var u=e(o,2);t(u,{title:"Transformer2DModelOutput",local:"diffusers.models.modeling_outputs.Transformer2DModelOutput",headingTag:"h2"});var n=e(u,2),M=r(n);s(M,{name:"class diffusers.models.modeling_outputs.Transformer2DModelOutput",anchor:"diffusers.models.modeling_outputs.Transformer2DModelOutput",source:"https://github.com/huggingface/diffusers/blob/vr_13966/src/diffusers/models/modeling_outputs.py#L21",parameters:[{name:"sample",val:": torch.Tensor"}],parametersDescription:[{anchor:"diffusers.models.modeling_outputs.Transformer2DModelOutput.sample",description:`<strong>sample</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, num_channels, height, width)</code> or <code>(batch size, num_vector_embeds - 1, num_latent_pixels)</code> if <a href="/docs/diffusers/pr_13966/en/api/models/transformer2d#diffusers.Transformer2DModel">Transformer2DModel</a> is discrete) &#x2014;
The hidden states output conditioned on the <code>encoder_hidden_states</code> input. If discrete, returns probability
distributions for the unnoised latent pixels.`,name:"sample"}]}),a(2),i(n);var y=e(n,2);N(y,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/en/api/models/transformer2d.md"}),a(2),g(T,d),C()}export{W as component};