A Diffusion Transformer model for 3D data from Latte.
( num_attention_heads: int = 16 attention_head_dim: int = 88 in_channels: Optional = None out_channels: Optional = None num_layers: int = 1 dropout: float = 0.0 cross_attention_dim: Optional = None attention_bias: bool = False sample_size: int = 64 patch_size: Optional = None activation_fn: str = 'geglu' num_embeds_ada_norm: Optional = None norm_type: str = 'layer_norm' norm_elementwise_affine: bool = True norm_eps: float = 1e-05 caption_channels: int = None video_length: int = 16 )
( hidden_states: Tensor timestep: Optional = None encoder_hidden_states: Optional = None encoder_attention_mask: Optional = None enable_temporal_attentions: bool = True return_dict: bool = True )
Parameters
(batch size, channel, num_frame, height, width)
—
Input hidden_states
. torch.LongTensor
, optional) —
Used to indicate denoising step. Optional timestep to be applied as an embedding in AdaLayerNorm
. torch.FloatTensor
of shape (batch size, sequence len, embed dims)
, optional) —
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
self-attention. torch.Tensor
, optional) —
Cross-attention mask applied to encoder_hidden_states
. Two formats supported:
(batcheight, sequence_length)
True = keep, False = discard.(batcheight, 1, sequence_length)
0 = keep, -10000 = discard.If ndim == 2
: 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.
enable_temporal_attentions —
(bool
, optional, defaults to True
): Whether to enable temporal attentions.
bool
, optional, defaults to True
) —
Whether or not to return a ~models.unet_2d_condition.UNet2DConditionOutput
instead of a plain
tuple. The LatteTransformer3DModel forward method.