skylark0924 / machine-learning-is-all-you-need

🔥🌟《Machine Learning 格物志》: ML + DL + RL basic codes and notes by sklearn, PyTorch, TensorFlow, Keras & the most important, from scratch!💪 This repository is ALL You Need!

Machine-Learning-Basic-Codes🏆

朱子云：

所谓致知在格物者，言欲致吾之知，在即物而穷其理也。盖人心之灵，莫不有知，而天下之物，莫不有理。惟于理有未穷，故其知有不尽也。是以大学始教，必使学者即凡天下之物，莫不因其已知之理而益穷之，以求至乎其极。至于用力之久，而一时豁然贯通焉，则众物之表里精粗无不到，而吾心之全体大用无不明矣。

📐📏

格物 (Ko Wu) which means 'investigate the essence of things' in English is a key method for study and better understanding of the knowledge. It is proposed by ancient Chinese philosophers about 2000 years ago and has a profound impact on later generations. The spirit of Ko Wu asks us to not only learn how to use knowledge, but also clearly understand the intrinsic theory. Therefore, it is necessary to re-implement ML algorithms by ourselves to figure out what exactly they did and why they succeed.

This repository aims to implement popular Machine Learning and Deep Learning algorithms by both pure python and use open-source frameworks.

• Common Machine Learning Part: switch by use_sklearn flag in the main function；
• Deep Learning Part: four implement methods for each algorithm (use_sklearn, use_keras, use_torch and self_implement)；
• Applications Part: RL + NLP + CV
• New trend: GNNs

Welcome everyone to help me finish this Ko Wu project by pulling requests or giving me some suggestions and issues!!!

关联知乎专栏 Associated Zhihu Blog

RL in Robotics

Machine Learning 格物志

代码目录 Code Catalog

Regression

1. Single Linear Regression
2. Multiple Linear Regression

Classification

3. Logistic Regression
4. KNN
5. Support Vector Machine
6. Naive Bayes

Regression & Classification

7. Decision Tree
8. Random Forest

Neural Network

9. Feedforward Neural Network
10. Convolutional Neural Network
11. LSTM

Unsupervised Learning

12. PCA
13. K-Means

Ensemble Model

14. Boosting

Reinforcement Learning

1. Value Based Methods: Q-learning(Tabular), DQN
2. Policy Based Methods: Vanilla Policy Gradient, TRPO, PPO
3. Actor-Critic Structure: AC, A2C, A3C
4. Deep Deterministic Policy Gradient: DDPG, DDPG C++ (Undone), TD3
5. Soft Actor-Critic

Computer Vision

1. GAN
2. Resnet: Pytorch version, libtorch C++ version
3. VGG
4. FlowNet

Natural Language Processing

1. Attention mechanism
2. Transformer
3. BERT

Graph Neural Networks

1. Graph Neural Network (GNN)
2. Graph Convolutional Neural Network (GCN)
3. Graph Attention Networks (GAT)
4. GraphSAGE
5. GraphRNN
6. Variational Graph Auto-Encoders (GAE)

---

If you're interested in reinforcement learning, we encourage you to check out our latest library of reinforcement learning and imitation learning in (humanoid) robotics.

Repository address: https://github.com/Skylark0924/Rofunc

machine-learning-is-all-you-need's Issues

the implement of TD error in Actor_Critic

Machine-Learning-is-ALL-You-Need/RL_Actor_Critic/17Actor_Critic.py

Lines 80 to 92 in 766a50b

	R = 0
	saved_actions = self.model.saved_actions
	policy_losses = [] # list to save actor (policy) loss
	value_losses = [] # list to save critic (value) loss
	returns = [] # list to save the true values
	# calculate the true value using rewards returned from the environment
	for r in self.model.rewards[::-1]:
	# calculate the discounted value
	R = r + self.gamma * R
	returns.insert(0, R)
	returns = torch.tensor(returns)

Shouldn't R here be set as the value estimation of the last state?
If I understand correctly, if R is set to 0, the returns for the later t become less accurate, and it also fails to reflect the temporal difference.