# 5. 训练模型

# 定义模型步长
num_steps = 100001

with tf.compat.v1.Session(graph=graph) as session:
    init.run()

    average_loss = 0
    for step in range(num_skips):
        batch_inputs ,batch_labels = generate_batch(batch_size,num_skips,skip_window)
        feed_dict = {train_inputs:batch_inputs,train_labels:batch_labels}

        _,loss_val = session.run([optimizer,loss],feed_dict=feed_dict)
        average_loss += loss_val

        if step % 2000 == 0:
            if step > 0:
                average_loss /= 2000
            print("average loss at step:",step,":",average_loss)
            average_loss = 0

        if step % 10000 == 0:
            sim = similary.eval()
            valid_word = reverse_dictionary[valid_examples]
            top_k = 8
            nearest = (-sim[i,:]).argsort()[:top_k]
            log_str = "Nearest to %s"% valid_word
            for k in range(top_k):
                close_word = reverse_dictionary[nearest[k]]
                log_str = "%s %s,"%(log_str,close_word)
            print(log_str)
    final_embeddings = normnalized_embeddings.eval()

# 6. 输出向量
with open('output/word2vect.text',"w",encoding="utf-8") as fw2v:
    fw2v.write(str(vocabulary_size) + " " + str(embedding_size) + "\n")
    for i in range(final_embeddings.shape[0]):
        sword = reverse_dictionary[i]
        svector = ""

        for j in range(final_embeddings.shape[1]):
            svector = svector + " " + str(final_embeddings[i,j])
        fw2v.write(sword,svector + "\n")