文本特征向量抽取及PCA降维

1）加载ckpt模型
2）确定输出tensor名称，在bert中，cls的名称为：bert/pooler/dense/Tanh（而不是SoftMax）
3）存储为pb model

主代码：

def extract_bert_vector():
    """ 抽取bert 768 特征向量
    :return:
    """
    OUTPUT_GRAPH = 'pb_model/bert_encoder.pb'
    output_node = ["bert/pooler/dense/Tanh"]
    ckpt_model = r'output'
    bert_config_file = r'chinese_L-12_H-768_A-12/bert_config.json'
    max_seq_length = 200

    gpu_config = tf.ConfigProto()
    gpu_config.gpu_options.allow_growth = True
    sess = tf.Session(config=gpu_config)
    graph = tf.get_default_graph()
    with open(r'data/file_dict.json', 'r') as fr:
        label_list = json.load(fr)
    with graph.as_default():
        print("going to restore checkpoint")
        input_ids_p = tf.placeholder(tf.int32, [None, max_seq_length], name="input_ids")
        input_mask_p = tf.placeholder(tf.int32, [None, max_seq_length], name="input_mask")
        bert_config = modeling.BertConfig.from_json_file(bert_config_file)
        (loss, per_example_loss, logits, probabilities) = create_model(
            bert_config=bert_config, is_training=False, input_ids=input_ids_p, input_mask=input_mask_p,
            segment_ids=None, labels=None, num_labels=len(label_list), use_one_hot_embeddings=False)
        saver = tf.train.Saver()
        saver.restore(sess, tf.train.latest_checkpoint(ckpt_model))
        graph = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, output_node)
        with tf.gfile.GFile(OUTPUT_GRAPH, "wb") as f:
            f.write(graph.SerializeToString())
        print('extract vector pb model saved!')

 768维度明显过高，采用白化处理，将768->256

代码如下：

def compute_kernel_bias(vecs, n_components=256):
    """计算kernel和bias
    vecs.shape = [num_samples, embedding_size]，
    最后的变换：y = (x + bias).dot(kernel)
    """
    mu = vecs.mean(axis=0, keepdims=True)
    cov = np.cov(vecs.T)
    u, s, vh = scipy.linalg.svd(cov)
    W = np.dot(u, np.diag(1 / np.sqrt(s)))
    return W[:, :n_components], -mu


def transform_and_normalize(vecs, kernel=None, bias=None):
    """ 最终向量标准化
    """
    if not (kernel is None or bias is None):
        vecs = (vecs + bias).dot(kernel)
    return vecs / (vecs**2).sum(axis=1, keepdims=True)**0.5

v_data = np.array(v_data)    
kernel,bias=compute_kernel_bias(v_data,256)
v_data=transform_and_normalize(v_data, kernel=kernel, bias=bias)

该功能引用于：使用微调后的Bert模型做编码器进行文本特征向量抽取及特征降维 - 今夜无风 - 博客园 (cnblogs.com)你可能不需要BERT-flow：一个线性变换媲美BERT-flow - 科学空间|Scientific Spaces