PyTorch implementation of a research paper MaskGAN: Better Text Generation via Filling in the______ within the Deep Learning course at the HSE.
-
MaskMLE: dl_cshse_2019/seq2seq/maskmle.py
-
Algorithm:
- Pretrain rnn cells (lstm) on language modelling task.
- Use the weights in generator and discrimantor encoders and decoders
- Pretrain generator and discriminator in MaskMLE mode
- Train generator and discriminator in GAN actor-crititc mode with following loop:
- Fix discriminator parameters, train generator in actor-critic mode
- Fix generator parameters, sample real and fake sentences, train discriminator