Struct Transformer2DModel 

pub struct Transformer2DModel<T: Float> {
    pub norm: GroupNorm<T>,
    pub proj_in: Conv2d<T>,
    pub transformer_blocks: Vec<BasicTransformerBlock<T>>,
    pub proj_out: Conv2d<T>,
    /* private fields */
}

Diffusers’ Transformer2DModel configured the way SD-1.5’s UNet uses it:

h = norm(x)                              # GroupNorm(32, in_channels)
h = proj_in(h)                           # Conv2d(C, inner, k=1) [use_linear_projection=False]
h = h.permute(0, 2, 3, 1).reshape(B, H*W, inner)
for block in transformer_blocks:
    h = block(h, encoder_hidden_states)
h = h.reshape(B, H, W, inner).permute(0, 3, 1, 2)
h = proj_out(h)                          # Conv2d(inner, C, k=1)
return h + residual

SD-1.5 v1 uses Conv2d (not Linear) for proj_in/proj_out (use_linear_projection=False). transformer_layers_per_block=1 (the diffusers default and the SD-1.5 v1 setting).

Fields

norm: GroupNorm<T>

GroupNorm before proj_in.

proj_in: Conv2d<T>

proj_in: Conv2d(C, inner, k=1).

transformer_blocks: Vec<BasicTransformerBlock<T>>

N × BasicTransformerBlock.

proj_out: Conv2d<T>

proj_out: Conv2d(inner, C, k=1).

Implementations

impl<T: Float> Transformer2DModel<T>

pub fn new( in_channels: usize, heads: usize, dim_head: usize, num_layers: usize, cross_attention_dim: usize, norm_num_groups: usize, ) -> FerrotorchResult<Self>

Build a randomly-initialized Transformer2DModel.

inner_dim = heads * dim_head for the SD UNet (proj_in expands only when these disagree; for SD it’s always equal to in_channels).

Errors

Returns the underlying FerrotorchError for invalid dims.

pub fn forward_xattn( &self, x: &Tensor<T>, encoder_hidden_states: &Tensor<T>, ) -> FerrotorchResult<Tensor<T>>

Forward with encoder hidden states for cross-attn.

x has shape [B, C, H, W]. The result has the same shape.

Errors

Returns FerrotorchError::ShapeMismatch when the input is not [B, channels, H, W].

Trait Implementations

impl<T: Debug + Float> Debug for Transformer2DModel<T>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<T: Float> Module<T> for Transformer2DModel<T>

fn forward(&self, _input: &Tensor<T>) -> FerrotorchResult<Tensor<T>>

Forward pass. Takes input tensor, returns output tensor.

fn parameters(&self) -> Vec<&Parameter<T>>

Iterate over all learnable parameters.

fn parameters_mut(&mut self) -> Vec<&mut Parameter<T>>

Iterate over all learnable parameters mutably.

fn named_parameters(&self) -> Vec<(String, &Parameter<T>)>

Named parameters for state dict serialization.

fn train(&mut self)

Set training mode. Affects dropout, batchnorm, etc.

fn eval(&mut self)

Set evaluation mode.

fn is_training(&self) -> bool

Whether the module is in training mode.

fn load_state_dict( &mut self, state: &StateDict<T>, strict: bool, ) -> FerrotorchResult<()>

Load parameters from a state dict.

fn to_device(&mut self, device: Device) -> Result<(), FerrotorchError>

Move all parameters and buffers to a device.

fn state_dict(&self) -> HashMap<String, Tensor<T>>

Export parameters and buffers as a state dict (torch parity).

fn buffers(&self) -> Vec<&Buffer<T>>

Iterate over all non-trainable buffers (e.g. running mean / variance in BatchNorm). Default returns empty — concrete modules with buffers override.

fn buffers_mut(&mut self) -> Vec<&mut Buffer<T>>

Mutable iteration over all buffers. Default returns empty.

fn named_buffers(&self) -> Vec<(String, &Buffer<T>)>

Named buffers (dot-separated paths for nested modules). Default returns empty.

fn as_any(&self) -> Option<&(dyn Any + 'static)>

Downcast hook for type-erased buffer-loader dispatch. (#984)

fn children(&self) -> Vec<&dyn Module<T>>

Direct child modules. Default returns empty (leaf module).

fn named_children(&self) -> Vec<(String, &dyn Module<T>)>

Direct child modules with their attribute names. Default returns empty.

fn modules(&self) -> Vec<&dyn Module<T>>

where Self: Sized,

All modules in this subtree, depth-first (self first, then each child’s descendants in order).

fn descendants_dyn(&self) -> Vec<&dyn Module<T>>

All strict descendants of self in depth-first order. Object-safe.

fn named_modules(&self) -> Vec<(String, &dyn Module<T>)>

where Self: Sized,

All modules in this subtree with dot-separated path names. The root is named ""; children paths are joined with ..

fn named_descendants_dyn(&self) -> Vec<(String, &dyn Module<T>)>

Strict descendants with dot-paths. Object-safe.

fn with_forward_hook( self, hook: Box<dyn Fn(&Tensor<T>, &Tensor<T>) + Send + Sync>, ) -> (HookedModule<Self, T>, HookHandle)

where Self: Sized,

Wrap this module in a HookedModule and register a forward hook. Returns the wrapper paired with a HookHandle that can be used to remove the hook later. The wrapper implements Module<T> itself, so it slots into any place the original module did. Mirrors torch.nn.Module.register_forward_hook.

fn with_forward_pre_hook( self, hook: Box<dyn Fn(&Tensor<T>) -> Result<Tensor<T>, FerrotorchError> + Send + Sync>, ) -> (HookedModule<Self, T>, HookHandle)

where Self: Sized,

Wrap this module in a HookedModule and register a forward pre-hook. See Self::with_forward_hook. Mirrors torch.nn.Module.register_forward_pre_hook.

fn with_backward_hook( self, hook: Box<dyn Fn(&Tensor<T>, &Tensor<T>) + Send + Sync>, ) -> (HookedModule<Self, T>, HookHandle)

where Self: Sized,

Wrap this module in a HookedModule and register a backward hook. See Self::with_forward_hook. Mirrors torch.nn.Module.register_backward_hook.

fn zero_grad(&self) -> Result<(), FerrotorchError>

Set the gradient of every parameter to None.

fn requires_grad_(&mut self, requires_grad: bool)

Toggle requires_grad on every parameter (freeze / unfreeze the module). Mirrors torch.nn.Module.requires_grad_.

fn apply_to_parameters(&mut self, f: &mut dyn FnMut(&mut Parameter<T>))

Apply a function to every parameter in this module. Mirrors torch.nn.Module.apply for the parameter case (true apply recurses over all submodules; the recursive form requires &mut dyn Module which conflicts with this trait’s &mut self borrow).