def __init__(self, vocab_size, layers=1, input_dim=None, lstm_dim=None, mlp_hidden_dim=None, trg_embed_dim=None, dropout=None, rnn_spec="lstm", residual_to_output=False): lstm_dim = lstm_dim or model_globals.get("default_layer_dim") mlp_hidden_dim = mlp_hidden_dim or model_globals.get("default_layer_dim") trg_embed_dim = trg_embed_dim or model_globals.get("default_layer_dim") input_dim = input_dim or model_globals.get("default_layer_dim") self.fwd_lstm = RnnDecoder.rnn_from_spec(rnn_spec, layers, trg_embed_dim, lstm_dim, model_globals.dynet_param_collection.param_col, residual_to_output) self.mlp = MLP(input_dim + lstm_dim, mlp_hidden_dim, vocab_size, model_globals.dynet_param_collection.param_col) self.dropout = dropout or model_globals.get("dropout") self.state = None
def __init__(self, input_dim=None, state_dim=None, hidden_dim=None): input_dim = input_dim or model_globals.get("default_layer_dim") state_dim = state_dim or model_globals.get("default_layer_dim") hidden_dim = hidden_dim or model_globals.get("default_layer_dim") self.input_dim = input_dim self.state_dim = state_dim self.hidden_dim = hidden_dim param_collection = model_globals.dynet_param_collection.param_col self.pW = param_collection.add_parameters((hidden_dim, input_dim)) self.pV = param_collection.add_parameters((hidden_dim, state_dim)) self.pb = param_collection.add_parameters(hidden_dim) self.pU = param_collection.add_parameters((1, hidden_dim)) self.curr_sent = None
def __init__(self, input_dim=512, layers=1, hidden_dim=None, downsampling_method="skip", reduce_factor=2, dropout=None): hidden_dim = hidden_dim or model_globals.get("default_layer_dim") dropout = dropout or model_globals.get("dropout") self.dropout = dropout self.builder = pyramidal.PyramidalRNNBuilder( layers, input_dim, hidden_dim, model_globals.dynet_param_collection.param_col, dy.VanillaLSTMBuilder, downsampling_method, reduce_factor)
def init_builder(self, input_dim, layers, hidden_dim=None, chn_dim=3, num_filters=32, filter_size_time=3, filter_size_freq=3, stride=(2, 2), dropout=None): model = model_globals.dynet_param_collection.param_col hidden_dim = hidden_dim or model_globals.get("default_layer_dim") dropout = dropout or model_globals.get("dropout") self.dropout = dropout self.builder = conv_encoder.ConvBiRNNBuilder( layers, input_dim, hidden_dim, model, dy.VanillaLSTMBuilder, chn_dim, num_filters, filter_size_time, filter_size_freq, stride)
def __init__(self, input_dim=512, layers=1, hidden_dim=None, residual_to_output=False, dropout=None, bidirectional=True): model = model_globals.dynet_param_collection.param_col hidden_dim = hidden_dim or model_globals.get("default_layer_dim") dropout = dropout or model_globals.get("dropout") self.dropout = dropout if bidirectional: self.builder = residual.ResidualBiRNNBuilder( layers, input_dim, hidden_dim, model, dy.VanillaLSTMBuilder, residual_to_output) else: self.builder = residual.ResidualRNNBuilder(layers, input_dim, hidden_dim, model, dy.VanillaLSTMBuilder, residual_to_output)
def __init__(self, input_dim=None, layers=1, hidden_dim=None, dropout=None, bidirectional=True): model = model_globals.dynet_param_collection.param_col input_dim = input_dim or model_globals.get("default_layer_dim") hidden_dim = hidden_dim or model_globals.get("default_layer_dim") dropout = dropout or model_globals.get("dropout") self.input_dim = input_dim self.layers = layers self.hidden_dim = hidden_dim self.dropout = dropout if bidirectional: self.builder = dy.BiRNNBuilder(layers, input_dim, hidden_dim, model, dy.VanillaLSTMBuilder) else: self.builder = dy.VanillaLSTMBuilder(layers, input_dim, hidden_dim, model)
def __init__(self, vocab_size, emb_dim=None): self.vocab_size = vocab_size if emb_dim is None: emb_dim = model_globals.get("default_layer_dim") self.emb_dim = emb_dim self.embeddings = model_globals.dynet_param_collection.param_col.add_lookup_parameters( (vocab_size, emb_dim))