RankGAN Discriminator

class textbox.module.Discriminator.RankGANDiscriminator.RankGANDiscriminator(config, dataset)

Bases: UnconditionalGenerator

RankGANDiscriminator is a ranker which can endow a relative rank among the sequences when given a reference. The ranker is designed with the convolutional neural network.

calculate_loss(real_data, fake_data, ref_data)

Calculate the loss for real data and fake data. To rank the human_written sentences higher than the machine-written sentences.

Parameters

• real_data (torch.Tensor) – The realistic sentence data, shape: [batch_size, max_seq_len].

• fake_data (torch.Tensor) – The generated sentence data, shape: [batch_size, max_seq_len].

• ref_data (torch.Tensor) – The reference sentence data, shape: [ref_size, max_seq_len].

Returns

The calculated loss of real data and fake data, shape: [].

Return type

torch.Tensor

forward(data)

Maps concatenated sequence matrices into the embedded feature vectors.

Parameters

data (torch.Tensor) – The sentence data, shape: [batch_size, max_seq_len].

Returns

The embedded feature vectors, shape: [batch_size, total_filter_num].

Return type

torch.Tensor

get_rank_scores(sample_data, ref_data)

Get the ranking score (before softmax) for sample s given reference u.

\[\alpha(s|u) = cosine(y_s,y_u) = \frac{y_s \cdot y_u}{\parallel y_s \parallel \parallel y_u \parallel}\]

Parameters

• sample_data (torch.Tensor) – The realistic or generated sentence data, shape: [sample_size, max_seq_len].

• ref_data (torch.Tensor) – The reference sentence data, shape: [ref_size, max_seq_len].

Returns

The ranking score of sample data, shape: [batch_size].

Return type

torch.Tensor

highway(data)

Apply the highway net to data.

Parameters

data (torch.Tensor) – The original data, shape: [batch_size, total_filter_num].

Returns

The data processed after highway net, shape: [batch_size, total_filter_num].

Return type

torch.Tensor

training: bool