diffusers/pr_13966/en/_app/immutable/nodes/85.CwcNggM2.js

import"../chunks/DsnmJJEf.js";import{i as M,h as k,C as N,H as g,D as o,E as z,s as j}from"../chunks/BtE7mKSK.js";import{p as q,o as L,s as e,f as C,a as x,b as F,c as n,d as b,n as t,r}from"../chunks/jDjavuwI.js";const I='{"title":"PixArtTransformer2DModel","local":"pixarttransformer2dmodel","sections":[{"title":"PixArtTransformer2DModel","local":"diffusers.PixArtTransformer2DModel","sections":[],"depth":2}],"depth":1}';var O=b('<meta name="hf:doc:metadata"/>'),S=b(`<p></p> <!> <!> <p>A Transformer model for image-like data from <a href="https://huggingface.co/papers/2310.00426" rel="nofollow">PixArt-Alpha</a> and <a href="https://huggingface.co/papers/2403.04692" rel="nofollow">PixArt-Sigma</a>.</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>A 2D Transformer model as introduced in PixArt family of models (<a href="https://huggingface.co/papers/2310.00426" rel="nofollow">https://huggingface.co/papers/2310.00426</a>, <a href="https://huggingface.co/papers/2403.04692" rel="nofollow">https://huggingface.co/papers/2403.04692</a>).</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/pixart_transformer2d#diffusers.PixArtTransformer2DModel">PixArtTransformer2DModel</a> forward method.</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>Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
are fused. For cross-attention modules, key and value projection matrices are fused.</p> <blockquote class="warning"><p>> This API is 🧪 experimental.</p></blockquote></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>Disables custom attention processors and sets the default attention implementation.</p> <p>Safe to just use <code>AttnProcessor()</code> as PixArt doesn’t have any exotic attention processors in default model.</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>Disables the fused QKV projection if enabled.</p> <blockquote class="warning"><p>> This API is 🧪 experimental.</p></blockquote></div></div> <!> <p></p>`,1);function W(v,T){q(T,!1),L(()=>{new URLSearchParams(window.location.search).get("fw")}),M();var l=S();k("1mw7l43",_=>{var h=O();j(h,"content",I),x(_,h)});var c=e(C(l),2);N(c,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var m=e(c,2);g(m,{title:"PixArtTransformer2DModel",local:"pixarttransformer2dmodel",headingTag:"h1"});var f=e(m,4);g(f,{title:"PixArtTransformer2DModel",local:"diffusers.PixArtTransformer2DModel",headingTag:"h2"});var s=e(f,2),p=n(s);o(p,{name:"class diffusers.PixArtTransformer2DModel",anchor:"diffusers.PixArtTransformer2DModel",source:"https://github.com/huggingface/diffusers/blob/vr_13966/src/diffusers/models/transformers/pixart_transformer_2d.py#L32",parameters:[{name:"num_attention_heads",val:": int = 16"},{name:"attention_head_dim",val:": int = 72"},{name:"in_channels",val:": int = 4"},{name:"out_channels",val:": int | None = 8"},{name:"num_layers",val:": int = 28"},{name:"dropout",val:": float = 0.0"},{name:"norm_num_groups",val:": int = 32"},{name:"cross_attention_dim",val:": int | None = 1152"},{name:"attention_bias",val:": bool = True"},{name:"sample_size",val:": int = 128"},{name:"patch_size",val:": int = 2"},{name:"activation_fn",val:": str = 'gelu-approximate'"},{name:"num_embeds_ada_norm",val:": int | None = 1000"},{name:"upcast_attention",val:": bool = False"},{name:"norm_type",val:": str = 'ada_norm_single'"},{name:"norm_elementwise_affine",val:": bool = False"},{name:"norm_eps",val:": float = 1e-06"},{name:"interpolation_scale",val:": int | None = None"},{name:"use_additional_conditions",val:": bool | None = None"},{name:"caption_channels",val:": int | None = None"},{name:"attention_type",val:": str | None = 'default'"}],parametersDescription:[{anchor:"diffusers.PixArtTransformer2DModel.num_attention_heads",description:"<strong>num_attention_heads</strong> (int, optional, defaults to 16) &#x2014; The number of heads to use for multi-head attention.",name:"num_attention_heads"},{anchor:"diffusers.PixArtTransformer2DModel.attention_head_dim",description:"<strong>attention_head_dim</strong> (int, optional, defaults to 72) &#x2014; The number of channels in each head.",name:"attention_head_dim"},{anchor:"diffusers.PixArtTransformer2DModel.in_channels",description:"<strong>in_channels</strong> (int, defaults to 4) &#x2014; The number of channels in the input.",name:"in_channels"},{anchor:"diffusers.PixArtTransformer2DModel.out_channels",description:`<strong>out_channels</strong> (int, optional) &#x2014;
The number of channels in the output. Specify this parameter if the output channel number differs from the
input.`,name:"out_channels"},{anchor:"diffusers.PixArtTransformer2DModel.num_layers",description:"<strong>num_layers</strong> (int, optional, defaults to 28) &#x2014; The number of layers of Transformer blocks to use.",name:"num_layers"},{anchor:"diffusers.PixArtTransformer2DModel.dropout",description:"<strong>dropout</strong> (float, optional, defaults to 0.0) &#x2014; The dropout probability to use within the Transformer blocks.",name:"dropout"},{anchor:"diffusers.PixArtTransformer2DModel.norm_num_groups",description:`<strong>norm_num_groups</strong> (int, optional, defaults to 32) &#x2014;
Number of groups for group normalization within Transformer blocks.`,name:"norm_num_groups"},{anchor:"diffusers.PixArtTransformer2DModel.cross_attention_dim",description:`<strong>cross_attention_dim</strong> (int, optional) &#x2014;
The dimensionality for cross-attention layers, typically matching the encoder&#x2019;s hidden dimension.`,name:"cross_attention_dim"},{anchor:"diffusers.PixArtTransformer2DModel.attention_bias",description:`<strong>attention_bias</strong> (bool, optional, defaults to True) &#x2014;
Configure if the Transformer blocks&#x2019; attention should contain a bias parameter.`,name:"attention_bias"},{anchor:"diffusers.PixArtTransformer2DModel.sample_size",description:`<strong>sample_size</strong> (int, defaults to 128) &#x2014;
The width of the latent images. This parameter is fixed during training.`,name:"sample_size"},{anchor:"diffusers.PixArtTransformer2DModel.patch_size",description:`<strong>patch_size</strong> (int, defaults to 2) &#x2014;
Size of the patches the model processes, relevant for architectures working on non-sequential data.`,name:"patch_size"},{anchor:"diffusers.PixArtTransformer2DModel.activation_fn",description:`<strong>activation_fn</strong> (str, optional, defaults to &#x201C;gelu-approximate&#x201D;) &#x2014;
Activation function to use in feed-forward networks within Transformer blocks.`,name:"activation_fn"},{anchor:"diffusers.PixArtTransformer2DModel.num_embeds_ada_norm",description:`<strong>num_embeds_ada_norm</strong> (int, optional, defaults to 1000) &#x2014;
Number of embeddings for AdaLayerNorm, fixed during training and affects the maximum denoising steps during
inference.`,name:"num_embeds_ada_norm"},{anchor:"diffusers.PixArtTransformer2DModel.upcast_attention",description:`<strong>upcast_attention</strong> (bool, optional, defaults to False) &#x2014;
If true, upcasts the attention mechanism dimensions for potentially improved performance.`,name:"upcast_attention"},{anchor:"diffusers.PixArtTransformer2DModel.norm_type",description:`<strong>norm_type</strong> (str, optional, defaults to &#x201C;ada_norm_zero&#x201D;) &#x2014;
Specifies the type of normalization used, can be &#x2018;ada_norm_zero&#x2019;.`,name:"norm_type"},{anchor:"diffusers.PixArtTransformer2DModel.norm_elementwise_affine",description:`<strong>norm_elementwise_affine</strong> (bool, optional, defaults to False) &#x2014;
If true, enables element-wise affine parameters in the normalization layers.`,name:"norm_elementwise_affine"},{anchor:"diffusers.PixArtTransformer2DModel.norm_eps",description:`<strong>norm_eps</strong> (float, optional, defaults to 1e-6) &#x2014;
A small constant added to the denominator in normalization layers to prevent division by zero.`,name:"norm_eps"},{anchor:"diffusers.PixArtTransformer2DModel.interpolation_scale",description:"<strong>interpolation_scale</strong> (int, optional) &#x2014; Scale factor to use during interpolating the position embeddings.",name:"interpolation_scale"},{anchor:"diffusers.PixArtTransformer2DModel.use_additional_conditions",description:"<strong>use_additional_conditions</strong> (bool, optional) &#x2014; If we&#x2019;re using additional conditions as inputs.",name:"use_additional_conditions"},{anchor:"diffusers.PixArtTransformer2DModel.attention_type",description:"<strong>attention_type</strong> (str, optional, defaults to &#x201C;default&#x201D;) &#x2014; Kind of attention mechanism to be used.",name:"attention_type"},{anchor:"diffusers.PixArtTransformer2DModel.caption_channels",description:`<strong>caption_channels</strong> (int, optional, defaults to None) &#x2014;
Number of channels to use for projecting the caption embeddings.`,name:"caption_channels"},{anchor:"diffusers.PixArtTransformer2DModel.use_linear_projection",description:`<strong>use_linear_projection</strong> (bool, optional, defaults to False) &#x2014;
Deprecated argument. Will be removed in a future version.`,name:"use_linear_projection"},{anchor:"diffusers.PixArtTransformer2DModel.num_vector_embeds",description:`<strong>num_vector_embeds</strong> (bool, optional, defaults to False) &#x2014;
Deprecated argument. Will be removed in a future version.`,name:"num_vector_embeds"}]});var a=e(p,4),A=n(a);o(A,{name:"forward",anchor:"diffusers.PixArtTransformer2DModel.forward",source:"https://github.com/huggingface/diffusers/blob/vr_13966/src/diffusers/models/transformers/pixart_transformer_2d.py#L227",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:"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.PixArtTransformer2DModel.forward.hidden_states",description:`<strong>hidden_states</strong> (<code>torch.FloatTensor</code> of shape <code>(batch size, channel, height, width)</code>) &#x2014;
Input <code>hidden_states</code>.`,name:"hidden_states"},{anchor:"diffusers.PixArtTransformer2DModel.forward.encoder_hidden_states",description:`<strong>encoder_hidden_states</strong> (<code>torch.FloatTensor</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.PixArtTransformer2DModel.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.PixArtTransformer2DModel.forward.added_cond_kwargs",description:"<strong>added_cond_kwargs</strong> &#x2014; (<code>dict[str, Any]</code>, <em>optional</em>): Additional conditions to be used as inputs.",name:"added_cond_kwargs"},{anchor:"diffusers.PixArtTransformer2DModel.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.PixArtTransformer2DModel.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.PixArtTransformer2DModel.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.PixArtTransformer2DModel.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>~models.transformer_2d.Transformer2DModelOutput</code> is returned, otherwise a
<code>tuple</code> where the first element is the sample tensor.</p>
`}),t(2),r(a);var i=e(a,2),D=n(i);o(D,{name:"fuse_qkv_projections",anchor:"diffusers.PixArtTransformer2DModel.fuse_qkv_projections",source:"https://github.com/huggingface/diffusers/blob/vr_13966/src/diffusers/models/transformers/pixart_transformer_2d.py#L196",parameters:[]}),t(4),r(i);var d=e(i,2),P=n(d);o(P,{name:"set_default_attn_processor",anchor:"diffusers.PixArtTransformer2DModel.set_default_attn_processor",source:"https://github.com/huggingface/diffusers/blob/vr_13966/src/diffusers/models/transformers/pixart_transformer_2d.py#L187",parameters:[]}),t(4),r(d);var u=e(d,2),w=n(u);o(w,{name:"unfuse_qkv_projections",anchor:"diffusers.PixArtTransformer2DModel.unfuse_qkv_projections",source:"https://github.com/huggingface/diffusers/blob/vr_13966/src/diffusers/models/transformers/pixart_transformer_2d.py#L218",parameters:[]}),t(4),r(u),r(s);var y=e(s,2);z(y,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/en/api/models/pixart_transformer2d.md"}),t(2),x(v,l),F()}export{W as component};