vencent-debug / seqgan_pytorch

seqGAN in pytorch

SeqGAN in pytorch

Requirements

pytorch 0.4

jieba 0.39 (if you want to tokenize with wordseg.py)

Background

Written based on SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient, - Lantao Yu, Weinan Zhang, Jun Wang, Yong Yu.

Translated from the original tensorflow repo, and adjusted for wider usability.

Many thanks to the original authors.

Inputs and outputs

Input text as shown in sample data. Supports variable lengths input and batch processing, however for word segmentation you'd need to make some changes to the data_processing.py.

Train:

• Input sequence length is 'START' + tokens, output sequence y is tokens + 'END'. You need to specify in config.py the SEQ_LENGTH, which is the usual embedding size plus one for the 'START' token.
• Tokenized training file can be created by running the wordseg.py on the original input file london.txt. Default training file name is real_data_london.pkl.
• Training file path can be specified in config.py. Training file name can be set up in data_processing.py (currently as a default argument).
• To train based on input file: run sequenceGAN.py. The generated text from training will be saved in genTxt.txt file.
• To train in batches of e.g. 4: run sequenceGAN.py 4.
• During training, to read in only limited number of samples, e.g. 8: run sequenceGAN.py 4 8. If it's left out then all samples will be used for training.
• After training, the vocabulary for decoding, as well as the generator network will be saved in PATH. Training performance will be saved in record.txt file.

Generate new text:

• After training, to generate new text: run sequenceGAN_generate.py.
• To generate more than one sample (e.g. 4 samples): run sequenceGAN_generate.py 4.

Test:

• To individually test the modules, run the files separately (the sanityCheck functions will be run). Result will be saved in test.txt file in PATH.

Params

• SEQ_LENGTH: uniform length of sentences. Flexible input sequence length is not tested. Default to 8 because sample Chinese poem data has 7 characters per line, plus the START token.
• EMB_SIZE: embedded size, size of word2vec.
• BATCH_SIZE: training batch size.
• GENERATE_NUM: number of sentences to be generated for training.
• FILTER_SIZE = [ 1, 2, 3, 4, 5, 6, 7, 8]: different sizes of filters used in the discriminator. equivalent to n-grams.
• NUM_FILTER = [100,200,200,200,160,160,160,100]: corresponding filter size (or number of features to be extracted).
• DIS_NUM_EPOCH: number of epochs for discriminator training
• DIS_NUM_EPOCH_PRETRAIN: number of epochs for discriminator pretraining
• GEN_NUM_EPOCH: number of epochs for generator training
• GEN_NUM_EPOCH_PRETRAIN: number of epochs for generator pretraining
• GEN_HIDDEN_DIM: generator (LSTM) hidden dimension.
• ROLLOUT_ITER: number of iterations to roll out, for calculation of rewards
• TOTAL_BATCH: number of batches in adversarial training
• MAXINT: used as penalty factor to avoid choosing start and end tokens for the sequence. The penalty factor is applied before calculating the probability of output tokens.

Intro to SeqGAN

For basic understanding please refer to the original paper.

What's not in the paper, is my simplified view of the world:

The generator is not much different from a vanilla LSTM, except for a custom loss function with rewards (basically giving higher probability to generated sequences which are labeled as "real data" by the discriminator).

The rollout module is almost a copy of the generator - if the r_update_date is set to 0 then it IS the exact copy. Reason why the update is set to non-zero, is explained here and here. the rollout module does not participate in the gradient.

The discriminator module is a CNN with a series of parallel conv layers, acting similar to n-grams: