lstm_layers.py 文件源码

def param_init_encoder(options, params, prefix='lstm_encoder'):

    n_x = options['n_x']
    n_h = options['n_h']

    W = np.concatenate([uniform_weight(n_x,n_h),
                        uniform_weight(n_x,n_h),
                        uniform_weight(n_x,n_h),
                        uniform_weight(n_x,n_h)], axis=1)
    params[_p(prefix, 'W')] = W

    U = np.concatenate([ortho_weight(n_h),
                        ortho_weight(n_h),
                        ortho_weight(n_h),
                        ortho_weight(n_h)], axis=1)
    params[_p(prefix, 'U')] = U

    params[_p(prefix,'b')] = zero_bias(4*n_h)

    # It is observed that setting a high initial forget gate bias for LSTMs can 
    # give slighly better results (Le et al., 2015). Hence, the initial forget
    # gate bias is set to 3.
    params[_p(prefix, 'b')][n_h:2*n_h] = 3*np.ones((n_h,)).astype(theano.config.floatX)

    return params