深度学习用的有一年多了,最近开始NLP自然处理方面的研发。刚好趁着这个机会写一系列NLP机器翻译深度学习实战课程。
本系列课程将从原理讲解与数据处理深入到如何动手实践与应用部署,将包括以下内容:(更新ing)
- NLP机器翻译深度学习实战课程·零(基础概念)
- NLP机器翻译深度学习实战课程·壹(RNN base)
- NLP机器翻译深度学习实战课程·贰(RNN+Attention base)
- NLP机器翻译深度学习实战课程·叁(CNN base)
- NLP机器翻译深度学习实战课程·肆(Self-Attention base)
- NLP机器翻译深度学习实战课程·伍(应用部署)
本系列教程参考博客:https://me.csdn.net/chinatelecom08
0. 项目背景
在上个文章中,我们已经简单介绍了NLP机器翻译,这次我们将用实战的方式讲解基于RNN的翻译模型。
0.1 基于RNN的seq2seq架构翻译模型介绍
基于RNN的seq2seq架构包含encoder和decoder,decoder部分又分train和inference两个过程,具体结构如下面两图所示:
可以看出结构很简单(相较于CNN与Attention base),下面我们就通过代码的形式实现,来进一步探究理解模型内在原理。
1. 数据准备
1.1 下载数据
此网站http://www.manythings.org/anki/上有许多翻译数据,包含多种语言,这里此教程选择的是中文到英语的数据集。
训练下载地址:http://www.manythings.org/anki/cmn-eng.zip
解压cmn-eng.zip,可以找到cmn.txt文件,内容如下:
# ========读取原始数据======== with open('cmn.txt', 'r', encoding='utf-8') as f: data = f.read() data = data.split('\n') data = data[:100] print(data[-5:])
['Tom died.\t汤姆去世了。', 'Tom quit.\t汤姆不干了。', 'Tom swam.\t汤姆游泳了。', 'Trust me.\t相信我。', 'Try hard.\t努力。']
可以发现,每对翻译数据在同一行,左边是英文,右边是中文使用 \t 作为英语与中文的分界。
1.2 数据预处理
使用网络训练,需要我们把数据处理成网络可以接收的格式。
针对这个数据,具体来说就是需要把字符转换为数字(句子数字化),句子长度归一化处理。
句子数字化
可以参考我的这博客:『深度应用』NLP命名实体识别(NER)开源实战教程,数据预处理的实现。
分别对英语与汉字做处理。
英文处理
因为英语每个单词都是用空格分开的(除了缩写词,这里缩写词当做一个词处理),还有就是标点符号和单词没有分开,也需要特殊处理一下
这里我用的是一个简单方法处理下,实现在标点前加空格:
def split_dot(strs,dots=", . ! ?"): for d in dots.split(" "): #print(d) strs = strs.replace(d," "+d) #print(strs) return(strs)
使用这个方法来把词个字典化:
ef get_eng_dicts(datas): w_all_dict = {} for sample in datas: for token in sample.split(" "): if token not in w_all_dict.keys(): w_all_dict[token] = 1 else: w_all_dict[token] += 1 sort_w_list = sorted(w_all_dict.items(), key=lambda d: d[1], reverse=True) w_keys = [x for x,_ in sort_w_list[:7000-2]] w_keys.insert(0,"<PAD>") w_keys.insert(0,"<UNK>") w_dict = { x:i for i,x in enumerate(w_keys) } i_dict = { i:x for i,x in enumerate(w_keys) } return w_dict,i_dict
中文处理
在处理中文时可以发现,有繁体也有简体,所以最好转换为统一形式:(参考地址)
# 安装 pip install opencc-python-reimplemented # t2s - 繁体转简体(Traditional Chinese to Simplified Chinese) # s2t - 简体转繁体(Simplified Chinese to Traditional Chinese) # mix2t - 混合转繁体(Mixed to Traditional Chinese) # mix2s - 混合转简体(Mixed to Simplified Chinese)
使用方法,把繁体转换为简体:
import opencc cc = opencc.OpenCC('t2s') s = cc.convert('這是什麼啊?') print(s) #这是什么啊?
再使用jieba分词的方法来从句子中分出词来:
def get_chn_dicts(datas): w_all_dict = {} for sample in datas: for token in jieba.cut(sample): if token not in w_all_dict.keys(): w_all_dict[token] = 1 else: w_all_dict[token] += 1 sort_w_list = sorted(w_all_dict.items(), key=lambda d: d[1], reverse=True) w_keys = [x for x,_ in sort_w_list[:10000-4]] w_keys.insert(0,"<EOS>") w_keys.insert(0,"<GO>") w_keys.insert(0,"<PAD>") w_keys.insert(0,"<UNK>") w_dict = { x:i for i,x in enumerate(w_keys) } i_dict = { i:x for i,x in enumerate(w_keys) } return w_dict,i_dict
下面进行padding
def get_val(keys,dicts): if keys in dicts.keys(): val = dicts[keys] else: keys = "<UNK>" val = dicts[keys] return(val) def padding(lists,lens=LENS): list_ret = [] for l in lists: while(len(l)<lens): l.append(1) if len(l)>lens: l = l[:lens] list_ret.append(l) return(list_ret)
最后统一运行处理一下:
if __name__ == "__main__": df = read2df("cmn-eng/cmn.txt") eng_dict,id2eng = get_eng_dicts(df["eng"]) chn_dict,id2chn = get_chn_dicts(df["chn"]) print(list(eng_dict.keys())[:20]) print(list(chn_dict.keys())[:20]) enc_in = [[get_val(e,eng_dict) for e in eng.split(" ")] for eng in df["eng"]] dec_in = [[get_val("<GO>",chn_dict)]+[get_val(e,chn_dict) for e in jieba.cut(eng)]+[get_val("<EOS>",chn_dict)] for eng in df["chn"]] dec_out = [[get_val(e,chn_dict) for e in jieba.cut(eng)]+[get_val("<EOS>",chn_dict)] for eng in df["chn"]] enc_in_ar = np.array(padding(enc_in,32)) dec_in_ar = np.array(padding(dec_in,30)) dec_out_ar = np.array(padding(dec_out,30))
输出结果如下:
(TF_GPU) D:\Files\Prjs\Pythons\Kerases\MNT_RNN>C:/Datas/Apps/RJ/Miniconda3/envs/TF_GPU/python.exe d:/Files/Prjs/Pythons/Kerases/MNT_RNN/mian.py Using TensorFlow backend. eng chn 0 Hi . 嗨。 1 Hi . 你好。 2 Run . 你用跑的。 3 Wait ! 等等! 4 Hello ! 你好。 save csv Building prefix dict from the default dictionary ... Loading model from cache C:\Users\xiaos\AppData\Local\Temp\jieba.cache Loading model cost 0.788 seconds. Prefix dict has been built succesfully. ['<UNK>', '<PAD>', '.', 'I', 'to', 'the', 'you', 'a', '?', 'is', 'Tom', 'He', 'in', 'of', 'me', ',', 'was', 'for', 'have', 'The'] ['<UNK>', '<PAD>', '<GO>', '<EOS>', '。', '我', '的', '了', '你', '他', '?', '在', '汤姆', '是', '她', '吗', '我们', ',', '不', '很']
2. 构建模型与训练
2.1 构建模型与超参数
用的是双层LSTM网络
# =======预定义模型参数======== EN_VOCAB_SIZE = 7000 CH_VOCAB_SIZE = 10000 HIDDEN_SIZE = 256 LEARNING_RATE = 0.001 BATCH_SIZE = 50 EPOCHS = 100 # ======================================keras model================================== from keras.models import Model from keras.layers import Input, LSTM, Dense, Embedding,CuDNNLSTM from keras.optimizers import Adam import numpy as np def get_model(): # ==============encoder============= encoder_inputs = Input(shape=(None,)) emb_inp = Embedding(output_dim=128, input_dim=EN_VOCAB_SIZE)(encoder_inputs) encoder_h1, encoder_state_h1, encoder_state_c1 = CuDNNLSTM(HIDDEN_SIZE, return_sequences=True, return_state=True)(emb_inp) encoder_h2, encoder_state_h2, encoder_state_c2 = CuDNNLSTM(HIDDEN_SIZE, return_state=True)(encoder_h1) # ==============decoder============= decoder_inputs = Input(shape=(None,)) emb_target = Embedding(output_dim=128, input_dim=CH_VOCAB_SIZE)(decoder_inputs) lstm1 = CuDNNLSTM(HIDDEN_SIZE, return_sequences=True, return_state=True) lstm2 = CuDNNLSTM(HIDDEN_SIZE, return_sequences=True, return_state=True) decoder_dense = Dense(CH_VOCAB_SIZE, activation='softmax') decoder_h1, _, _ = lstm1(emb_target, initial_state=[encoder_state_h1, encoder_state_c1]) decoder_h2, _, _ = lstm2(decoder_h1, initial_state=[encoder_state_h2, encoder_state_c2]) decoder_outputs = decoder_dense(decoder_h2) model = Model([encoder_inputs, decoder_inputs], decoder_outputs) # encoder模型和训练相同 encoder_model = Model(encoder_inputs, [encoder_state_h1, encoder_state_c1, encoder_state_h2, encoder_state_c2]) # 预测模型中的decoder的初始化状态需要传入新的状态 decoder_state_input_h1 = Input(shape=(HIDDEN_SIZE,)) decoder_state_input_c1 = Input(shape=(HIDDEN_SIZE,)) decoder_state_input_h2 = Input(shape=(HIDDEN_SIZE,)) decoder_state_input_c2 = Input(shape=(HIDDEN_SIZE,)) # 使用传入的值来初始化当前模型的输入状态 decoder_h1, state_h1, state_c1 = lstm1(emb_target, initial_state=[decoder_state_input_h1, decoder_state_input_c1]) decoder_h2, state_h2, state_c2 = lstm2(decoder_h1, initial_state=[decoder_state_input_h2, decoder_state_input_c2]) decoder_outputs = decoder_dense(decoder_h2) decoder_model = Model([decoder_inputs, decoder_state_input_h1, decoder_state_input_c1, decoder_state_input_h2, decoder_state_input_c2], [decoder_outputs, state_h1, state_c1, state_h2, state_c2]) return(model,encoder_model,decoder_model)
2.2 模型配置与训练
自定义了一个acc,便于显示效果,keras内置的acc无法使用
import keras.backend as K from keras.models import load_model def my_acc(y_true, y_pred): acc = K.cast(K.equal(K.max(y_true,axis=-1),K.cast(K.argmax(y_pred,axis=-1),K.floatx())),K.floatx()) return acc Train = True if __name__ == "__main__": df = read2df("cmn-eng/cmn.txt") eng_dict,id2eng = get_eng_dicts(df["eng"]) chn_dict,id2chn = get_chn_dicts(df["chn"]) print(list(eng_dict.keys())[:20]) print(list(chn_dict.keys())[:20]) enc_in = [[get_val(e,eng_dict) for e in eng.split(" ")] for eng in df["eng"]] dec_in = [[get_val("<GO>",chn_dict)]+[get_val(e,chn_dict) for e in jieba.cut(eng)]+[get_val("<EOS>",chn_dict)] for eng in df["chn"]] dec_out = [[get_val(e,chn_dict) for e in jieba.cut(eng)]+[get_val("<EOS>",chn_dict)] for eng in df["chn"]] enc_in_ar = np.array(padding(enc_in,32)) dec_in_ar = np.array(padding(dec_in,30)) dec_out_ar = np.array(padding(dec_out,30)) #dec_out_ar = covt2oh(dec_out_ar) if Train: model,encoder_model,decoder_model = get_model() model.load_weights('e2c1.h5') opt = Adam(lr=LEARNING_RATE, beta_1=0.9, beta_2=0.99, epsilon=1e-08) model.compile(optimizer=opt, loss='sparse_categorical_crossentropy',metrics=[my_acc]) model.summary() print(dec_out_ar.shape) model.fit([enc_in_ar, dec_in_ar], np.expand_dims(dec_out_ar,-1), batch_size=50, epochs=64, initial_epoch=0, validation_split=0.1) model.save('e2c1.h5') encoder_model.save("enc1.h5") decoder_model.save("dec1.h5")
64Epoch训练结果如下:
__________________________________________________________________________________________________ Layer (type) Output Shape Param # Connected to ================================================================================================== input_1 (InputLayer) (None, None) 0 __________________________________________________________________________________________________ input_2 (InputLayer) (None, None) 0 __________________________________________________________________________________________________ embedding_1 (Embedding) (None, None, 128) 896000 input_1[0][0] __________________________________________________________________________________________________ embedding_2 (Embedding) (None, None, 128) 1280000 input_2[0][0] __________________________________________________________________________________________________ cu_dnnlstm_1 (CuDNNLSTM) [(None, None, 256), 395264 embedding_1[0][0] __________________________________________________________________________________________________ cu_dnnlstm_3 (CuDNNLSTM) [(None, None, 256), 395264 embedding_2[0][0] cu_dnnlstm_1[0][1] cu_dnnlstm_1[0][2] __________________________________________________________________________________________________ cu_dnnlstm_2 (CuDNNLSTM) [(None, 256), (None, 526336 cu_dnnlstm_1[0][0] __________________________________________________________________________________________________ cu_dnnlstm_4 (CuDNNLSTM) [(None, None, 256), 526336 cu_dnnlstm_3[0][0] cu_dnnlstm_2[0][1] cu_dnnlstm_2[0][2] __________________________________________________________________________________________________ dense_1 (Dense) (None, None, 10000) 2570000 cu_dnnlstm_4[0][0] ================================================================================================== Non-trainable params: 0 __________________________________________________________________________________________________ ... ... 19004/19004 [==============================] - 98s 5ms/step - loss: 0.1371 - my_acc: 0.9832 - val_loss: 2.7299 - val_my_acc: 0.7412 Epoch 58/64 19004/19004 [==============================] - 96s 5ms/step - loss: 0.1234 - my_acc: 0.9851 - val_loss: 2.7378 - val_my_acc: 0.7410 Epoch 59/64 19004/19004 [==============================] - 96s 5ms/step - loss: 0.1132 - my_acc: 0.9867 - val_loss: 2.7477 - val_my_acc: 0.7419 Epoch 60/64 19004/19004 [==============================] - 96s 5ms/step - loss: 0.1050 - my_acc: 0.9879 - val_loss: 2.7660 - val_my_acc: 0.7426 Epoch 61/64 19004/19004 [==============================] - 96s 5ms/step - loss: 0.0983 - my_acc: 0.9893 - val_loss: 2.7569 - val_my_acc: 0.7408 Epoch 62/64 19004/19004 [==============================] - 96s 5ms/step - loss: 0.0933 - my_acc: 0.9903 - val_loss: 2.7775 - val_my_acc: 0.7414 Epoch 63/64 19004/19004 [==============================] - 96s 5ms/step - loss: 0.0885 - my_acc: 0.9911 - val_loss: 2.7885 - val_my_acc: 0.7420 Epoch 64/64 19004/19004 [==============================] - 96s 5ms/step - loss: 0.0845 - my_acc: 0.9920 - val_loss: 2.7914 - val_my_acc: 0.7423
3. 模型应用与预测
从训练集选取部分数据进行测试
Train = False if __name__ == "__main__": df = read2df("cmn-eng/cmn.txt") eng_dict,id2eng = get_eng_dicts(df["eng"]) chn_dict,id2chn = get_chn_dicts(df["chn"]) print(list(eng_dict.keys())[:20]) print(list(chn_dict.keys())[:20]) enc_in = [[get_val(e,eng_dict) for e in eng.split(" ")] for eng in df["eng"]] dec_in = [[get_val("<GO>",chn_dict)]+[get_val(e,chn_dict) for e in jieba.cut(eng)]+[get_val("<EOS>",chn_dict)] for eng in df["chn"]] dec_out = [[get_val(e,chn_dict) for e in jieba.cut(eng)]+[get_val("<EOS>",chn_dict)] for eng in df["chn"]] enc_in_ar = np.array(padding(enc_in,32)) dec_in_ar = np.array(padding(dec_in,30)) dec_out_ar = np.array(padding(dec_out,30)) #dec_out_ar = covt2oh(dec_out_ar) if Train: pass else: encoder_model,decoder_model = load_model("enc1.h5",custom_objects={"my_acc":my_acc}),load_model("dec1.h5",custom_objects={"my_acc":my_acc}) for k in range(16000-20,16000): test_data = enc_in_ar[k:k+1] h1, c1, h2, c2 = encoder_model.predict(test_data) target_seq = np.zeros((1,1)) outputs = [] target_seq[0, len(outputs)] = chn_dict["<GO>"] while True: output_tokens, h1, c1, h2, c2 = decoder_model.predict([target_seq, h1, c1, h2, c2]) sampled_token_index = np.argmax(output_tokens[0, -1, :]) #print(sampled_token_index) outputs.append(sampled_token_index) #target_seq = np.zeros((1, 30)) target_seq[0, 0] = sampled_token_index #print(target_seq) if sampled_token_index == chn_dict["<EOS>"] or len(outputs) > 28: break print("> "+df["eng"][k]) print("< "+' '.join([id2chn[i] for i in outputs[:-1]])) print()
> I can understand you to some extent . < 在 某种程度 上 我 能 了解 你 。 > I can't recall the last time we met . < 我 想不起来 我们 上次 见面 的 情况 了 。 > I can't remember which is my racket . < 我 不 记得 哪个 是 我 的 球拍 。 > I can't stand that noise any longer . < 我 不能 再 忍受 那 噪音 了 。 > I can't stand this noise any longer . < 我 无法 再 忍受 这个 噪音 了 。 > I caught the man stealing the money . < 我 抓 到 了 这个 男人 正在 偷钱 。 > I could not afford to buy a bicycle . < 我 买不起 自行车 。 > I couldn't answer all the questions . < 我 不能 回答 所有 的 问题 。 > I couldn't think of anything to say . < 我 想不到 要说 什么 话 。 > I cry every time I watch this movie . < 我 每次 看 这部 电影 都 会 哭 。 > I did not participate in the dialog . < 我 没有 参与 对话 。 > I didn't really feel like going out . < 我 不是 很想 出去 。 > I don't care a bit about the future . < 我 不在乎 将来 。