README.md

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
  • These are the basic modules of TensorLayer, you can run all single machine examples with this module only.
  • TensorFlow 1.0+
    • pip install tensorflow-gpu or follow the official installation guide
  • TensorLayer
    • It is self-contained
    • You can also get the latest version in Github
• 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)
    • You may need this environement to distribute the different jobs. An installation process is available at that address: https://github.com/aoehmichen/eae-docker/blob/master/install_eae_hybrid.txt

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.

• CIFAR-10 classification
  • TensorFlow Implementation
  • TensorLayer Implementation, Optimized Version (5x faster, optimize CPU/GPU operation, tested on Titan X Pascal)
• PTB language modelling
  • TensorFlow Implementation
  • TensorLayer Implementation
• Word2vec
  • TensorFlow Implementation
  • TensorLayer Implementation

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
  • tutorial_tensordb_atari_pong_generator.py is the data generator.
  • tutorial_tensordb_atari_pong_trainer.py is the GPU trainer.
• 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.

• DeepSleepNet

Generative Adversarial Network

You only need to install TensorFlow to run these examples.

• DCGAN
  • TensorFlow Implementation
  • TensorLayer Implementation
• Text to image synthesis
  • TensorFlow Implementation
  • TensorLayer Implementation

More

• Documentation
• Examples and Applications
• Github