Machine Learning Demos

This repo is a collection of code snippets to demostrate common models and applications of Machine Learning. Each script is a standalone program. The intention is to capture the core of the idea with the minimal amount of code. Hope you find it helpful as a quick reference to understand ML techniques.

Natural Language Processing

• nlp/text_generation_gpt_2_beam_search.py: GPT-2, text generation, pre-trained model, beam search
• nlp/text_generation_gpt_2_top_k_top_p_filtering.py: GPT-2, text generation, pre-trained model, top-k top-p filtering
• nlp/text_generation_gpt_2_greedy_search.py: GPT-2, text generation, pre-trained model, greedy search
• nlp/transformer.py: the original Transformer model implemented with PyTorch
• nlp/imdb_bert.py: BERT, sentiment analysis, transfer learning, text classification
• nlp/imdb_lstm.py: LSTM, embedding, text classification
• nlp/imdb_tf_hub.py: embedding, TF-Hub, text classification
• nlp/imdb.py: embedding, text classification

Speech

• speech/rnnt.py: RNN transducer, speech recognition, end-to-end model. This is only to demostrate the implemenation of the model, because I don't have the massive hardware resource required to train it (300GB data + 1 GPU month).

Vision

• vision/oxford_iiit_pet_unet.py: U-Net, transfer learning, image segmentation
• vision/cats_vs_dogs.py: ResNet, transfer learning, image classification
• vision/cats_vs_dogs_heatmap.py: ResNet, CNN activation map, transfer learning
• vision/mnist_cnn.py: simple CNN, image classification
• vision/mnist.py: DNN, Dropout, image classification

Generative Adversarial Networks

• gan/mnist.py: Deep Convolutional GAN, image generation

Reinforcement Learning

All were written with PyTorch and run with OpenAI Gym.

• rl/cartpole_pytorch_dqn.py: Deep Q Learning, experience replay
• rl/cartpole_pytorch_actor_critic.py: Actor-Critic architecture
• rl/cartpole_pytorch_ddpg.py: DDPG, Actor-Critic architecture
• rl/cartpole_pytorch_policy_gradient.py: vanilla policy gradient method
• rl/cartpole_random_pytorch.py: randomized weights (toy algorithm)
• rl/mountain_car.py: Deep Q Learning, experience replay
• rl/mountain_car_continuous_ddpg.py: DDPG, Actor-Critic architecture
• rl/robotics.py: an attempt to get some more complex cases work. Failed due to not having enough computing power. Used stable_baselines, hindsight experience replay and DDPG.

Machine Learning Fundementals

ml-fundementals: A series of scripts to demostrate the core and heart of Machine Learning: backpropagation, SGD, losses, optimizers, layers, ReLU and automatic differentiation. All were written with PyTorch.