gram.py 文件源码

def init_params(options):
    params = OrderedDict()

    np.random.seed(0)
    inputDimSize = options['inputDimSize']
    numAncestors = options['numAncestors']
    embDimSize = options['embDimSize']
    hiddenDimSize = options['hiddenDimSize'] #hidden layer does not need an extra space
    attentionDimSize = options['attentionDimSize']
    numClass = options['numClass']

    params['W_emb'] = get_random_weight(inputDimSize+numAncestors, embDimSize)
    if len(options['embFile']) > 0:
        params['W_emb'] = load_embedding(options)
        options['embDimSize'] = params['W_emb'].shape[1]
        embDimSize = options['embDimSize']

    params['W_attention'] = get_random_weight(embDimSize*2, attentionDimSize)
    params['b_attention'] = np.zeros(attentionDimSize).astype(config.floatX)
    params['v_attention'] = np.random.uniform(-0.1, 0.1, attentionDimSize).astype(config.floatX)

    params['W_gru'] = get_random_weight(embDimSize, 3*hiddenDimSize)
    params['U_gru'] = get_random_weight(hiddenDimSize, 3*hiddenDimSize)
    params['b_gru'] = np.zeros(3 * hiddenDimSize).astype(config.floatX)

    params['W_output'] = get_random_weight(hiddenDimSize, numClass)
    params['b_output'] = np.zeros(numClass).astype(config.floatX)

    return params