/tl_paper

Primary LanguagePython

Codes for ACM MM Open Source Software Competition

This repo summarizes the codes that mention in "TensorLayer: A Versatile Library for Efficient Deep Learning Development".

Set Up Environment

Before you start, you need to set up the environment for all examples in this repo.

  • Layer and Model Modules
  • Deep Reinforcement Learning Environment
    • To run the DRL example, you will need to install OpenAI gym for game environment, lz4 for data compression and the dataset module.
    • pip install gym lz4
    • sudo apt-get install swig cmake
    • pip install gym[atari] or pip install gym[all]
  • Dataset and Workflow Modules
    • For Deep Reinforcement Learning and Hyper-parameter selection and cross-validation you may want to use dataset and workflow modules.
    • Install MongoDB
      • Follow MongoDB docs
      • We recommend to use one machine as dataset server.
    • Install eAE (Optional)

Raw Performance

You only need to install TensorFlow to run these examples. This a raw performance comparsion between TensorLayer and original TensorFlow engine, to prove TensorLayer's simplicity would not sacrifice the performance.

Deep Reinforcement Learning

You need to install all modules to run this example. This is a simple asynchronous DRL example, you can run this example in one machine or multiple machines with dataset module.

  • About the codes
  • Run the experiment
    • Before you run it in distributed mode, we higly recommend you to run one trainer with multiple data generators in a standalone machine.
    • For distributed mode, one machine run one trainer with GPU, all other machines run data generator.
      • Start multiple agents that generate training samples for the trainer. Run the following script uses to start multiple generators. Note: all of the generated data will be stored in MongoDB, which will be fetched by the trainer
        • python submit_job_eae.py
      • After start the generators, run the following script to train a model.
        • python tutorial_tensordb_atari_pong_trainer.py
      • To terminate all of the generators, run the following scripts.
        • python terminate_running_jobs.py

Hyper-parameter selection and cross-validation

You need to install all modules to run this example.

Generative Adversarial Network

You only need to install TensorFlow to run these examples.

More