CNN Decoder

class textbox.module.Decoder.cnn_decoder.BasicCNNDecoder(input_size, latent_size, decoder_kernel_size, decoder_dilations, dropout_ratio)

Bases: Module

Basic Convolution Neural Network (CNN) decoder. Code Reference: https://github.com/kefirski/contiguous-succotash

forward(decoder_input, noise)

Implement the decoding process.

Parameters

• decoder_input (Torch.Tensor) – target sequence embedding, shape: [batch_size, sequence_length, embedding_size].

• noise (Torch.Tensor) – latent code, shape: [batch_size, latent_size].

Returns

output features, shape: [batch_size, sequence_length, feature_size].

Return type

torch.Tensor

training: bool

class textbox.module.Decoder.cnn_decoder.HybridDecoder(embedding_size, latent_size, hidden_size, num_dec_layers, rnn_type, vocab_size)

Bases: Module

Hybrid Convolution Neural Network (CNN) and Recurrent Neural Network (RNN) decoder. Code Reference: https://github.com/kefirski/hybrid_rvae

conv_decoder(latent_variable)

Implement the CNN decoder.

Parameters

latent_variable (Torch.Tensor) – latent code, shape: [batch_size, latent_size].

Returns

output features, shape: [batch_size, sequence_length, feature_size].

Return type

torch.Tensor

forward(decoder_input, latent_variable)

Implement the decoding process.

Parameters

• decoder_input (Torch.Tensor) – target sequence embedding, shape: [batch_size, sequence_length, embedding_size].

• latent_variable (Torch.Tensor) – latent code, shape: [batch_size, latent_size].

Returns

• torch.Tensor: RNN output features, shape: [batch_size, sequence_length, feature_size].

• torch.Tensor: CNN output features, shape: [batch_size, sequence_length, feature_size].

Return type

tuple

rnn_decoder(cnn_logits, decoder_input, initial_state=None)

Implement the RNN decoder using CNN output.

Parameters

• cnn_logits (Torch.Tensor) – latent code, shape: [batch_size, sequence_length, feature_size].

• decoder_input (Torch.Tensor) – target sequence embedding, shape: [batch_size, sequence_length, embedding_size].

• initial_state (Torch.Tensor) – initial hidden states, default: None.

Returns

• Torch.Tensor: output features, shape: [batch_size, sequence_length, num_directions * hidden_size].

• Torch.Tensor: hidden states, shape: [batch_size, num_layers * num_directions, hidden_size].

Return type

tuple

training: bool