jintao-huang/ml_alg

jintao的machine learning库

ML-ALG

Introduction

1. 这个仓库是一个机器学习算法, 传统算法的集成库, 现在主要是自己在使用. 所在文件夹为libs/, 下面将会介绍各个文件的用途.

文件功能介绍

1. mini_lightning部分, 现已移置: https://github.com/ustcml/mini-lightning
   1. 含: mini-lightning轻量级的深度学习训练框架.
   2. 含: Examples: cv, nlp, dqn, gan, contrastive_learning, gnn, ae, vae; ddp等.
2. leetcode-alg部分: 现已移置: https://github.com/Jintao-Huang/LeetCode-Py
   1. 含: leetcode-alg数据结构和算法库
   2. 含: 基于leetcode-alg的leetcode(python)题目的解答
3. libs/ml/_ml_alg/*: 机器学习中的算法实现
   1. _metrics.py: ml中的metrics的torch实现. (faster than torchmetrics.functional, sklearn, 使用torch实现, 支持cuda加速)
      1. 含accuracy, confusion_matrix, precision, recall, f1_score, fbeta_score, PR_curve, AP, roc_curve, AUC, r2_score, cosine_similarity, euclidean_distance, kl_divergence, pearson_corrcoef, spearman_corrcoef.
   2. _nn_functional.py: 实现torch.nn.functional包中的算法. (没啥实用性, 用于学习)
      1. 含激活函数, 损失, batch_norm, layer_norm, dropout, linear, conv2d, conv_transpose2d, conv1d, avg_pool2d, max_pool2d, rnn_relu_cell, rnn_tanh_cell, lstm_cell, gru_cell, multi-head attention, interpolate(nearest, bilinear), adaptive_avg_pool2d, adaptive_max_pool2d.
   3. _ml_alg.py: 传统ml算法的torch实现 (faster than sklearn, 支持cuda加速). (开发中...)
      1. 含归一化方法, LinearRegression, Ridge, LogisticRegression, PCA, KMeans, NearestNeighbors等
   4. _optim_functional.py: 优化器的实现. (没啥实用性, 用于学习)
      1. 含sgd, adam, adamw.
   5. _tvt_functional_tensor.py: torchvision.transforms._functional_tensor的实现. (没啥实用性, 用于学习)
      1. 含: to_tensor, normalize, pad, hflip, vflip, rgb_to_grayscale, crop, center_crop, resize, resized_crop, adjust_brightness, adjust_contrast, adjust_saturation, adjust_hue, rotate, affine
   6. _tvt_functional.py: torchvision.transforms.functional; torchvision.transforms的实现. (没啥实用性, 用于学习)
      1. 含: random_horizontal_flip, random_resized_crop...
   7. _linalg.py: 线性代数算法. (没啥实用性, 用于学习)
      1. 含pinv, solve, lstsq, cholesky_solve, lu_solve等
   8. _functional/*: 一些torch的函数实现. (没啥实用性, 用于学习)
      1. 含logsumexp, softmax, var, cov, corrcoef, bincount, unique_consecutive
      2. 含div, fmod, remainder
   9. _rand.py: (没啥实用性, 用于学习)
      1. 含normal, uniform, randperm, multivariate_normal
   10. _pygnn_functional.py: 图网络的实现. (开发中...)
   11. _class_impl/: pytorch的常见base类: Module, Optimizer, _LRScheduler的简化版
4. libs/alg_fast/*: 传统算法库的numba/cython版本 (开发中...)
5. examples/*: 一些代表性的examples
6. libs/_plt/*, 可视化的库.
   1. _2d.py:
      1. 含plot, scatter, imshow, hist, bar, text, contour等.
      2. 含config_ax, config_plt, config_fig等.
   2. _3d.py
7. libs/ml/
   1. _pd/*: torch pandas库. (开发中)
   2. _models: 一些模型的实现.
8. libs/utils/*: 一些工具函数的实现

Installation and Use

# Installation
# 下载仓库到本地, 进入setup.py所在文件夹. 输入以下命令即可(会自动安装依赖, pytorch请手动安装, 避免cuda版本不匹配)
pip install -e .

# Use
from libs import *

TODO

1. tvtF: adjust_hue; rotate; affine
2. pyg: pygnn的函数