Expand description
ai layers and value types relating to the burn library
Re-exports§
pub use burn as lib;
Structs§
- Attention
- layer for computing attention from [key,query,value] inputs
- Attention
Config - layer for computing attention from [key,query,value] inputs
- Batch
Stacker - batcher that stacks things
- Bias
Config - layer for adding bias somewhere
- Cache
- layer for caching kv values from kqv when run mutably. cats along d1 and outputs the concatenated keys and values.
- Classification
- wrapper for converting loss to classification output
- Classification
Layer - layer for converting loss to classification output
- Dont
Render - metrics renderer implementation that doesn’t actually do anything
- Identity
- identity version that knows what backend
- KQV
- layer for linear splitting into [key,query,value] for attention purposes
- KQVConfig
- layer for linear splitting into [key,query,value] for attention purposes
- Loss
Output - general loss output for being converted into other loss outputs
- Power
Mask Info - power mask information
- Regression
- wrapper for converting loss to regression output
- Regression
Layer - layer for converting loss to regression output
- Train
Config - configuration for convenient training through the wrapper
- Wrapped
- wraps in a burn wrapper
Enums§
- Attention
Mask - Config
- enumerates config for some burn layers
- Kind
- enumerates kinds for values
- Layer
- enumerates some burn layers
- Reshape
- value reshaping arguments
- Shape
- tensor shapes for Value
- Value
- enumerates burn tensors up to 8 dimensions, along with a variant to represent operation compatibility errors, and a variant for multiple tensors. An empty multi variant can be used to represent a lack of data. Once a the depth of a multi variant is enough for an operation to take full effect, further nesting should result in the same as applying separately
Traits§
- Shortcuts
- chained method shortcut trait
- ToBackend
- trait for switching the backend of a module
- Wrappable
- higher kinded type trait to allow rewrapping burn modules in different backends to implement some wrapper features
Functions§
- apply_
depthwise - helper function for applying operations that apply to a specific depth of multiple structure such that wrapping multiple appropriate inputs with a multi outputs the output of the function applied to all inputs. 0 depth for empty, 1 for single, 2+ for multi
- new
- starts the building of an ai structure in chained method style from an identity operation