tinybig.reconciliation
This module provides the parameter reconciliation functions that can be used to build the RPN model within the tinyBIG toolkit.
Parameter Reconciliation Function
Formally, given the underlying data distribution mapping \(f: {R}^m \to {R}^n\) to be learned, the parameter reconciliation function \(\psi\) adjusts the available parameter vector of length \(l\) by fabricating a new parameter matrix of size \(n \times D\) to accommodate the expansion space dimension \(D\) as follows:
\[ \psi: {R}^l \to {R}^{n \times D}, \]
which is defined only on the parameters without any input data.
In most of the cases, the parameter vector length \(l\) is much smaller than the output matrix size \(n \times D\), i.e., \(l \ll n \times D\). Meanwhile, in practice, we can also define function \(\psi\) to fabricate a longer parameter vector into a smaller parameter matrix, i.e., \(l > n \times D\). To unify these different cases, the data reconciliation function can also be referred to as the "parameter fabrication function", and these function names will be used interchangeably.
Classes in this Module
This module contains the following categories of parameter reconciliation functions:
- Basic reconciliation functions
- Lowrank reconciliation functions
- Hypernet reconciliation functions
Organization of this Module
Parameter Reconciliation/Fabrication
- fabrication (defined in the tinybig.module directory)
Basic Reconciliation
- constant_reconciliation
- zero_reconciliation
- one_reconciliation
- constant_eye_reconciliation
- identity_reconciliation
- masking_reconciliation
- duplicated_padding_reconciliation