RankGAN Generator¶

class textbox.module.Generator.RankGANGenerator.RankGANGenerator(config, dataset)[source]¶

Bases: UnconditionalGenerator

RankGANGenerator is a generative model with the LSTMs.

adversarial_loss(ref_data, discriminator_func)[source]¶

Calculate the adversarial generator loss guided by discriminator. The Monte Carlo rollouts methods is utilized to simulate intermediate rewards when a sequence is incomplete. For the partial sequence, the average ranking score is used to approximate the expected future reward.

Parameters: discriminator_func (function) – The function provided from discriminator to calculated the ranking score.
Returns: The calculated adversarial loss, shape: [].
Return type: torch.Tensor

calculate_loss(corpus, nll_test=False)[source]¶

Calculate the generated loss of corpus.

Parameters

corpus (Corpus) – The corpus to be calculated.
nll_test (Bool) – Optional; if nll_test is True the loss is calculated in sentence level rather than in word level.

Returns

The calculated loss of corpus, shape: [].

Return type

torch.Tensor

generate(batch_data, eval_data)[source]¶

Generate tokens of sentences using eval_data.

Parameters

batch_data (Corpus) – Single batch corpus information of evaluation data.
eval_data – Common information of all evaluation data.

Returns

The generated tokens of each sentence.

Return type

List[List[str]]

sample(sample_num)[source]¶

Sample sample_num generated sentence indice.

Parameters: sample_num (int) – The number to generate.
Returns: The generated sentence indice, shape: [sample_num, max_length].
Return type: torch.Tensor

sample_batch()[source]¶

Sample a batch of generated sentence indice.

Returns: The generated sentence indice, shape: [batch_size, max_length].
Return type: torch.Tensor

training: bool¶