Train LeakGAN with a custom dataset? about textgan-pytorch

[williamSYSU]

I have added the LeakGAN instructor leakgan_instructor.py in `instructor/real_data', please refer to it.

For a custom dataset, the steps should be as follows:

1. Preprocess dataset files, one sentence per line, ending with '\n'. Put them into dataset folder, training data should be placed in the root of dataset and testing data should be placed in the testdata.

2. Please refer to the function init_dict() in utils/text_process.py, similarly add the code to load your custom dataset files. (This function will automatically initialing dictionary based on dataset args in the future.)

3. Now you can design your own Instructor. The main differences of instructors between oracle data and real data are:

   • self.oracle_data should load training data (config.train_data) and add a self.test_data loading testing data (config.test_data).

   • Add BLEU metrics.

   • The positive samples for discriminator should be `self.oracle_data.target' now.

   • Remove the evaluation process of discriminator since there is no validation data.

   • Change the way of saving generator's samples. Before changing write_tensor to write_tokens, use tensor_to_tokens() to transform Tensor to word tokens.

See 'instructor/real_data/leakgan_instructor.py' for more details.

Please don't worry that these questions will bother me. These also help me to improve my code.

[ishalyminov]

Perfect, thanks a lot! 🙏