Python blocks.bricks.recurrent.LSTM Examples

The following are 3 code examples of blocks.bricks.recurrent.LSTM(). You can vote up the ones you like or vote down the ones you don't like, and go to the original project or source file by following the links above each example. You may also want to check out all available functions/classes of the module blocks.bricks.recurrent , or try the search function .
Example #1
Source File: bricks.py    From image-captioning-for-mortals with BSD 3-Clause "New" or "Revised" License 6 votes vote down vote up 
def __init__(self, image_feature_dim, embedding_dim, **kwargs):
        super(Encoder, self).__init__(**kwargs)

        self.image_embedding = Linear(
              input_dim=image_feature_dim
            , output_dim=embedding_dim
            , name="image_embedding"
            )

        self.to_inputs = Linear(
              input_dim=embedding_dim
            , output_dim=embedding_dim*4 # times 4 cuz vstack(input, forget, cell, hidden)
            , name="to_inputs"
            )

        self.transition = LSTM(
            dim=embedding_dim, name="transition")

        self.children = [ self.image_embedding
                        , self.to_inputs
                        , self.transition
                        ]
Example #2
Source File: model.py    From blocks-char-rnn with MIT License 5 votes vote down vote up 
def lstm_layer(dim, h, n):
    linear = Linear(input_dim=dim, output_dim=dim * 4, name='linear' + str(n))
    lstm = LSTM(dim=dim, name='lstm' + str(n))
    initialize([linear, lstm])
    return lstm.apply(linear.apply(h))
Example #3
Source File: attentive_reader.py    From DeepMind-Teaching-Machines-to-Read-and-Comprehend with MIT License 5 votes vote down vote up 
def make_bidir_lstm_stack(seq, seq_dim, mask, sizes, skip=True, name=''):
    bricks = []

    curr_dim = [seq_dim]
    curr_hidden = [seq]

    hidden_list = []
    for k, dim in enumerate(sizes):
        fwd_lstm_ins = [Linear(input_dim=d, output_dim=4*dim, name='%s_fwd_lstm_in_%d_%d'%(name,k,l)) for l, d in enumerate(curr_dim)]
        fwd_lstm = LSTM(dim=dim, activation=Tanh(), name='%s_fwd_lstm_%d'%(name,k))

        bwd_lstm_ins = [Linear(input_dim=d, output_dim=4*dim, name='%s_bwd_lstm_in_%d_%d'%(name,k,l)) for l, d in enumerate(curr_dim)]
        bwd_lstm = LSTM(dim=dim, activation=Tanh(), name='%s_bwd_lstm_%d'%(name,k))

        bricks = bricks + [fwd_lstm, bwd_lstm] + fwd_lstm_ins + bwd_lstm_ins

        fwd_tmp = sum(x.apply(v) for x, v in zip(fwd_lstm_ins, curr_hidden))
        bwd_tmp = sum(x.apply(v) for x, v in zip(bwd_lstm_ins, curr_hidden))
        fwd_hidden, _ = fwd_lstm.apply(fwd_tmp, mask=mask)
        bwd_hidden, _ = bwd_lstm.apply(bwd_tmp[::-1], mask=mask[::-1])
        hidden_list = hidden_list + [fwd_hidden, bwd_hidden]
        if skip:
            curr_hidden = [seq, fwd_hidden, bwd_hidden[::-1]]
            curr_dim = [seq_dim, dim, dim]
        else:
            curr_hidden = [fwd_hidden, bwd_hidden[::-1]]
            curr_dim = [dim, dim]

    return bricks, hidden_list