The original size of the lora model is 1.7g, in order to save hard disk space and video memory space, through sv_fro method for matrix compression, reduced to the original 1/100 size, and the loss of accuracy control within 5%, if you need the full version of 1.7g can join our QQ group: 895633778

Introduction to compression algorithm:

sv_fro method is a method of deep compression of matrix, which is mainly used to reduce the storage space and computational complexity of matrix.

The basic idea of the sv_fro method is to decompose A matrix into the product of three matrices using Singular Value Decomposition (SVD) : A = U SV ^T, where U and V are orthogonal matrices and S is a singular value matrix. A singular value matrix S is a diagonal matrix whose diagonal elements are called singular values.

In sv_fro method, only the part with large singular value in matrix A is kept, and the smaller singular value is set to 0, so as to realize the deep compression of matrix. The specific steps are as follows:

1. Perform singular value decomposition on matrix A to obtain U, S and V.

2. According to the preset threshold, set the singular value in S that is less than the threshold to 0.

3. Reconstruct the original matrix A using the compressed singular value matrix S, U and V.

After deep compression by sv_fro method, the storage space of matrix can be greatly reduced, and the computational complexity can be reduced to a certain extent. This has important practical significance for the tasks that need to deal with large-scale matrices.

模型简介

本lora模型原大小为1.7g，为了节省硬盘空间和显存空间，通过sv_fro方法进行矩阵压缩，缩小到原来的1/100大小，而损失精度控制在百分之5以内，如果你需要完整版1.7g的可以加入我们的QQ群：895633778

压缩算法简介：

sv_fro方法是一种对矩阵进行深度压缩的方法，主要用于减少矩阵的存储空间和计算复杂度。

sv_fro方法的基本思想是利用奇异值分解（Singular Value Decomposition，SVD）将矩阵分解为三个矩阵的乘积：A = U S V^T，其中U和V是正交矩阵，S是奇异值矩阵。奇异值矩阵S是一个对角矩阵，其对角线上的元素称为奇异值。

在sv_fro方法中，只保留矩阵A中奇异值较大的部分，将较小的奇异值设为0，从而实现对矩阵的深度压缩。具体步骤如下：

1. 对矩阵A进行奇异值分解，得到U、S和V。

2. 根据预设的阈值，将S中小于阈值的奇异值设为0。

3. 使用压缩后的奇异值矩阵S以及U和V重构原始矩阵A。

通过sv_fro方法进行深度压缩后，可大幅减少矩阵的存储空间，同时还能在一定程度上降低计算复杂度。这对于需要处理大规模矩阵的任务来说，具有重要的实际应用意义。

Alpha1

本版本的封面你可以非常清晰的看到XL与SD1.5之间的差距。

在原训练素材中，有衣服上带文字的图，如同封面你看到的一样，XL有个文字概念，而且非常的清晰！

这是对XL的第一次探索，希望大家能够多多返图！