two_tier.py 文件源码

def frame_level_rnn(input_sequences, h0, reset):
    """
    input_sequences.shape: (batch size, n frames * FRAME_SIZE)
    h0.shape:              (batch size, N_GRUS, DIM)
    reset.shape:           ()
    output.shape:          (batch size, n frames * FRAME_SIZE, DIM)
    """

    learned_h0 = lib.param(
        'FrameLevel.h0',
        numpy.zeros((N_GRUS, DIM), dtype=theano.config.floatX)
    )
    learned_h0 = T.alloc(learned_h0, h0.shape[0], N_GRUS, DIM)
    learned_h0 = T.patternbroadcast(learned_h0, [False] * learned_h0.ndim)
    h0 = theano.ifelse.ifelse(reset, learned_h0, h0)

    frames = input_sequences.reshape((
        input_sequences.shape[0],
        input_sequences.shape[1] / FRAME_SIZE,
        FRAME_SIZE
    ))

    # Rescale frames from ints in [0, Q_LEVELS) to floats in [-2, 2]
    # (a reasonable range to pass as inputs to the RNN)
    frames = (frames.astype('float32') / lib.floatX(Q_LEVELS/2)) - lib.floatX(1)
    frames *= lib.floatX(2)

    gru0 = lib.ops.LowMemGRU('FrameLevel.GRU0', FRAME_SIZE, DIM, frames, h0=h0[:, 0])
    grus = [gru0]
    for i in xrange(1, N_GRUS):
        gru = lib.ops.LowMemGRU('FrameLevel.GRU'+str(i), DIM, DIM, grus[-1], h0=h0[:, i])
        grus.append(gru)

    output = lib.ops.Linear(
        'FrameLevel.Output', 
        DIM,
        FRAME_SIZE * DIM,
        grus[-1],
        initialization='he'
    )
    output = output.reshape((output.shape[0], output.shape[1] * FRAME_SIZE, DIM))

    last_hidden = T.stack([gru[:,-1] for gru in grus], axis=1)

    return (output, last_hidden)