jeju-dl-camp-2018

Authors: Valliappa Chockalingam and Rishab Gargeya

Code Overview

• Trainer

  • train.py: Execute this file to run the entire project.
    • Flags:
    • Configuration Script: Location of configuration file.
  • Manager Class:
    • Self Variables: Env, Agent.
    • Run Method: Train multiple episodes. Visualize distributions (saves graph) with distagent.

• Configuration

  • All variabes for configuration: Network, Environment, Agent, Optimizer, Experience Replay.
  • Configuration file is propagated through trainer.py into each submodule folder (Agent, Optimizer, Memory, Environment).

• Agent

  • Agent creates the Network, Optimizer, Memory and explores the Environment.
  • Types: Agent, Categorical Agent, Quantile Regression, VAE
  • Methods: Act, Greedly Learning, Distribution, Learn

• Action Policy

  • Implements Policies
  • Methods: Policy, Epsilon Greedy

• Optimizer

  • Implements Optimizers for Network
  • Methods: Adam
  • Features: Gradient Clipping

• Memory

  • Implements Experience Replay
  • Methods: Add, Sample Size

• Environment

  • Wrapper for OpenAI Gym
  • Methods: Step, Reset, Render, Num_Actions, Observation_Dims

• Util

  • Util Functions
  • Methods: TF Copy OP

• GCP Config

  • Config files for running experiments on GCP
  • setup.py, config.yml, submit.sh : To run GCP CMLE ./submit.sh

• Function Approximator

  • Networks to learn relationships. Agent creates the network,
  • Head: Action Layer for Policy or Value.
  • Network: Algorithmic Function Approximator. Generic MLP / CNN implementation.
  • Types: Categorical DQN, Quantile Regression DQN, Implicit Quantile Networks (IQN)

• environment.yml : dependencies list

• run on GCP: nohup xvfb-run -s "-screen 0 1400x900x24" python -m trainer.train params_*.json &

This was supported by Deep Learning Camp Jeju 2018 which was organized by TensorFlow Korea User Group. We also thank our mentors Yu-Han Liu and Taehoon Kim.