def __init__(self,
word_embeddings_size,
hidden_size,
embeddings_freeze=False):
super(LanguageEncodingLSTM, self).__init__()
self.lstm_cell = LSTMCell(word_embeddings_size, hidden_size, bias=True)
self.freeze(embeddings_freeze)
def __init__(self, joint_embed_size, language_encoding_lstm_size,
hidden_size):
super(LanguageGenerationLSTM, self).__init__()
self.lstm_cell = LSTMCell(language_encoding_lstm_size +
joint_embed_size,
hidden_size,
bias=True)
def __init__(self,
word_map,
embed_dim=512,
encoder_output_dim=2048,
pretrained_embeddings=None,
embeddings_freeze=False,
hidden_size=512,
attention_dim=512,
dropout=0.1):
super(SATDecoder,
self).__init__(word_map, embed_dim, encoder_output_dim,
pretrained_embeddings, embeddings_freeze)
self.hidden_size = hidden_size
self.attention_dim = attention_dim
self.dropout_ratio = dropout
# LSTM layers
self.decode_step = LSTMCell(
input_size=encoder_output_dim +
self.embed_dim, # if layer == 0 else hidden_size,
hidden_size=hidden_size,
bias=True)
#self.layers = nn.ModuleList([
# LSTMCell(input_size=encoder_output_dim + embed_dim if layer == 0 else hidden_size,
# hidden_size=hidden_size, bias=True)
# for layer in range(num_layers)
#])
# Attention Layer
self.attention = AttentionLayer(encoder_output_dim, hidden_size,
attention_dim)
# Linear layers to find initial states of LSTMs
self.init_h = nn.Linear(encoder_output_dim, hidden_size)
self.init_c = nn.Linear(encoder_output_dim, hidden_size)
# Gating scalars and sigmoid layer (cf. section 4.2.1 of the paper)
self.f_beta = nn.Linear(hidden_size, encoder_output_dim)
self.sigmoid = nn.Sigmoid()
# Dropout layer
self.dropout = nn.Dropout(p=dropout)
# Linear layers for output generation
self.linear_o = nn.Linear(embed_dim, self.vocab_size)
self.linear_h = nn.Linear(hidden_size, self.embed_dim)
self.linear_z = nn.Linear(encoder_output_dim, self.embed_dim)
def __init__(self, attn_lstm_dim, viz_attn_dim, hidden_size):
super(LanguageLSTM, self).__init__()
self.lstm_cell = LSTMCell(input_size=attn_lstm_dim + viz_attn_dim,
hidden_size=hidden_size,
bias=True)
def __init__(self, embed_dim, lang_lstm_dim, encoder_dim, hidden_size):
super(AttentionLSTM, self).__init__()
self.lstm_cell = LSTMCell(input_size=lang_lstm_dim + encoder_dim +
embed_dim,
hidden_size=hidden_size,
bias=True)