songyouwei / TextGAN-PyTorch

TextGAN is a PyTorch framework for Generative Adversarial Networks (GANs) based text generation models.

TextGAN-PyTorch

TextGAN is a PyTorch framework for Generative Adversarial Networks (GANs) based text generation models. TextGAN serves as a benchmarking platform to support research on GAN-based text generation models. Since most GAN-based text generation models are implemented by Tensorflow, TextGAN can help those who get used to PyTorch to enter the text generation field faster.

For now, only few GANs-based models are implemented, including SeqGAN (Yu et. al, 2017), LeakGAN (Guo et. al, 2018) and RelGAN (Nie et. al, 2018). If you find any mistake in my implementation, please let me know! Also, please feel free to contribute to this repository if you want to add other models.

Requirements

• PyTorch >= 1.0.0
• Python 3.6
• Numpy 1.14.5
• CUDA 7.5+ (For GPU)
• nltk 3.4
• tqdm 4.32.1

To install, run pip install -r requirements.txt. In case of CUDA problems, consult the official PyTorch Get Started guide.

Implemented Models and Original Papers

• SeqGAN - SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient
• LeakGAN - Long Text Generation via Adversarial Training with Leaked Information
• RelGAN - RelGAN: Relational Generative Adversarial Networks for Text Generation

Get Started

• Get Started

git clone https://github.com/williamSYSU/TextGAN-PyTorch.git
cd TextGAN-PyTorch

• For real data experiments, Image COCO and EMNLP news dataset can be downloaded from here.
• Run with SeqGAN

cd run
python3 run_seqgan.py 0 0	# The first 0 is job_id, the second 0 is gpu_id

• Run with LeakGAN

cd run
python3 run_leakgan.py 0 0

• Run with RelGAN

cd run
python3 run_relgan.py 0 0

Features

1. Instructor

   For each model, the entire runing process is defined in instructor/oracle_data/seqgan_instructor.py. (Take SeqGAN in Synthetic data experiment for example). Some basic functions like init_model()and optimize() are defined in the base class BasicInstructor in instructor.py. If you want to add a new GAN-based text generation model, please create a new instructor under instructor/oracle_data and define the training process for the model.

2. Visualization

   Use utils/visualization.py to visualize the log file, including model loss and metrics scores. Custom your log files in log_file_list, no more than len(color_list). The log filename should exclude .txt.

3. Logging

   The TextGAN-PyTorch use the logging module in Python to record the running process, like generator's loss and metric scores. For the convenience of visualization, there would be two same log file saved in log/log_****_****.txt and save/**/log.txt respectively. Furthermore, The code would automatically save the state dict of models and a batch-size of generator's samples in ./save/**/models and ./save/**/samples per log step, where ** depends on your hyper-parameters.

4. Running Signal

   You can easily control the training process with the class Signal (please refer to utils/helpers.py) based on dictionary file run_signal.txt.

   For using the Signal, just edit the local file run_signal.txt and set pre_sig to Fasle for example, the program will stop pre-training process and step into next training phase. It is convenient to early stop the training if you think the current training is enough.

5. Automatiaclly select GPU

   In config.py, the program would automatically select a GPU device with the least GPU-Util in nvidia-smi. This feature is enabled by default. If you want to manually select a GPU device, please uncomment the --device args in run_[run_model].py and specify a GPU device with command.

Reproduction Results

Please note the log step of each model is different. See run_[run_model].py for details of log step.

Synthetic data: Oracle data

• NLL_oracle

  LeakGAN suprisely outperforms RelGAN due to its temperature control, but LeakGAN’s samples suffered from severe mode collapse.

  Though both LeakGAN and RelGAN would suffer from mode collapse, the pattern of collapse is different. LeakGAN will generate a sentence with only a few words. RelGAN will generate repeated sentences with different words.

  

• NLL_gen

  

Real data: Image COCO data

• BLEU-3​

  

• NLL_gen

  

TODO

• Add Experiment Results
• Fix bugs in LeakGAN model
• Add instructors of SeqGAN and LeakGAN in instrutor/real_data
• Fix logging bugs for save_root
• Fix issues of $NLL_{oracle}$ in SeqGAN model