#Professor forcing paper code

https://arxiv.org/pdf/1610.09038.pdf

Progress

• Generator and discriminator implemented
• Char-level PTB tested
• Jupyter Notebook with results
• Word-level tests
• Sequential MNIST (also mentioned in the paper)
• Publish pre-trained model

Requirements

• PyTorch 0.4 (older versions will not work)
• TensorboardX for tensorboard usage (optional)

Preprocessing

No special preprocessing is required, just create a file with tokens separated by space. In the original paper, the car-level language modelling is used, so in this case you should create a single file with content like 'h e l l o _ w o r l d !'

Training

python train.py Useful command-line arguments:

• -cuda for GPU
• -adversarial for training both generator and discriminator (otherwise model.generator will not be initialized and trained)
• -data_path
• -vocab_path. Vocab file is created if not provided and saved in vocab.pt file.
• -save_path Path to save the model. Model is saved after each epoch and info about the model itself and its results is appended to its name.

For more parameters, consult opts.py

Evaluating

python sampler.py

For sampling, -checkpoint, data_path and vocab_path arguments must be provided.