『NLP打卡营学习笔记』实践课2：文本语义相似度计算

# 正式开始实验之前首先通过如下命令安装最新版本的 paddlenlp!python -m pip install --upgrade paddlenlp==2.0.2 -i https://pypi.org/simple

import timeimport osimport numpy as npimport paddleimport paddle.nn.functional as Ffrom paddlenlp.datasets import load_datasetimport paddlenlp# 一键加载 Lcqmc 的训练集、验证集train_ds, dev_ds = load_dataset("lcqmc", splits=["train", "dev"])

# paddlenlp 会自动下载 lcqmc 数据集解压到 "${HOME}/.paddlenlp/datasets/LCQMC/lcqmc/lcqmc/" 目录下! ls ${HOME}/.paddlenlp/datasets/LCQMC/lcqmc/lcqmcprint(paddlenlp.__version__)

# 输出训练集的前 20 条样本for idx, example in enumerate(train_ds):    if idx <= 20:        print(example)

# 因为是基于预训练模型 ERNIE-Gram 来进行，所以需要首先加载 ERNIE-Gram 的 tokenizer，# 后续样本转换函数基于 tokenizer 对文本进行切分tokenizer = paddlenlp.transformers.ErnieGramTokenizer.from_pretrained('ernie-gram-zh')

# 将 1 条明文数据的 query、title 拼接起来，根据预训练模型的 tokenizer 将明文转换为 ID 数据# 返回 input_ids 和 token_type_idsdef convert_example(example, tokenizer, max_seq_length=512, is_test=False):

    query, title = example["query"], example["title"]

    encoded_inputs = tokenizer(
        text=query, text_pair=title, max_seq_len=max_seq_length)

    input_ids = encoded_inputs["input_ids"]
    token_type_ids = encoded_inputs["token_type_ids"]    if not is_test:
        label = np.array([example["label"]], dtype="int64")        return input_ids, token_type_ids, label    # 在预测或者评估阶段，不返回 label 字段
    else:        return input_ids, token_type_ids

### 对训练集的第 1 条数据进行转换input_ids, token_type_ids, label = convert_example(train_ds[0], tokenizer)

print(input_ids)

print(token_type_ids)

print(label)

# 为了后续方便使用，我们给 convert_example 赋予一些默认参数from functools import partial# 训练集和验证集的样本转换函数trans_func = partial(
    convert_example,
    tokenizer=tokenizer,
    max_seq_length=512)

from paddlenlp.data import Stack, Pad, Tuplea = [1, 2, 3, 4]
b = [3, 4, 5, 6]
c = [5, 6, 7, 8]
result = Stack()([a, b, c])print("Stacked Data: \n", result)print()

a = [1, 2, 3, 4]
b = [5, 6, 7]
c = [8, 9]
result = Pad(pad_val=0)([a, b, c])print("Padded Data: \n", result)print()

data = [
        [[1, 2, 3, 4], [1]],
        [[5, 6, 7], [0]],
        [[8, 9], [1]],
       ]
batchify_fn = Tuple(Pad(pad_val=0), Stack())
ids, labels = batchify_fn(data)print("ids: \n", ids)print()print("labels: \n", labels)print()

# 我们的训练数据会返回 input_ids, token_type_ids, labels 3 个字段# 因此针对这 3 个字段需要分别定义 3 个组 batch 操作batchify_fn = lambda samples, fn=Tuple(
    Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input_ids
    Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # token_type_ids
    Stack(dtype="int64")  # label): [data for data in fn(samples)]

# 定义分布式 Sampler: 自动对训练数据进行切分，支持多卡并行训练batch_sampler = paddle.io.DistributedBatchSampler(train_ds, batch_size=32, shuffle=True)# 基于 train_ds 定义 train_data_loader# 因为我们使用了分布式的 DistributedBatchSampler, train_data_loader 会自动对训练数据进行切分train_data_loader = paddle.io.DataLoader(
        dataset=train_ds.map(trans_func),
        batch_sampler=batch_sampler,
        collate_fn=batchify_fn,
        return_list=True)# 针对验证集数据加载，我们使用单卡进行评估，所以采用 paddle.io.BatchSampler 即可# 定义 dev_data_loaderbatch_sampler = paddle.io.BatchSampler(dev_ds, batch_size=32, shuffle=False)
dev_data_loader = paddle.io.DataLoader(
        dataset=dev_ds.map(trans_func),
        batch_sampler=batch_sampler,
        collate_fn=batchify_fn,
        return_list=True)

import paddle.nn as nn# 我们基于 ERNIE-Gram 模型结构搭建 Point-wise 语义匹配网络# 所以此处先定义 ERNIE-Gram 的 pretrained_modelpretrained_model = paddlenlp.transformers.ErnieGramModel.from_pretrained('ernie-gram-zh')#pretrained_model = paddlenlp.transformers.ErnieModel.from_pretrained('ernie-1.0')class PointwiseMatching(nn.Layer):
   
    # 此处的 pretained_model 在本例中会被 ERNIE-Gram 预训练模型初始化
    def __init__(self, pretrained_model, dropout=None):
        super().__init__()
        self.ptm = pretrained_model
        self.dropout = nn.Dropout(dropout if dropout is not None else 0.1)        # 语义匹配任务: 相似、不相似 2 分类任务
        self.classifier = nn.Linear(self.ptm.config["hidden_size"], 2)    def forward(self,
                input_ids,
                token_type_ids=None,
                position_ids=None,
                attention_mask=None):

        # 此处的 Input_ids 由两条文本的 token ids 拼接而成
        # token_type_ids 表示两段文本的类型编码
        # 返回的 cls_embedding 就表示这两段文本经过模型的计算之后而得到的语义表示向量
        _, cls_embedding = self.ptm(input_ids, token_type_ids, position_ids,
                                    attention_mask)

        cls_embedding = self.dropout(cls_embedding)        # 基于文本对的语义表示向量进行 2 分类任务
        logits = self.classifier(cls_embedding)
        probs = F.softmax(logits)        return probs# 定义 Point-wise 语义匹配网络model = PointwiseMatching(pretrained_model)

from paddlenlp.transformers import LinearDecayWithWarmup

epochs = 3num_training_steps = len(train_data_loader) * epochs# 定义 learning_rate_scheduler，负责在训练过程中对 lr 进行调度lr_scheduler = LinearDecayWithWarmup(5E-5, num_training_steps, 0.0)# Generate parameter names needed to perform weight decay.# All bias and LayerNorm parameters are excluded.decay_params = [
    p.name for n, p in model.named_parameters()    if not any(nd in n for nd in ["bias", "norm"])
]# 定义 Optimizeroptimizer = paddle.optimizer.AdamW(
    learning_rate=lr_scheduler,
    parameters=model.parameters(),
    weight_decay=0.0,
    apply_decay_param_fun=lambda x: x in decay_params)# 采用交叉熵 损失函数criterion = paddle.nn.loss.CrossEntropyLoss()# 评估的时候采用准确率指标metric = paddle.metric.Accuracy()

# 因为训练过程中同时要在验证集进行模型评估，因此我们先定义评估函数@paddle.no_grad()def evaluate(model, criterion, metric, data_loader, phase="dev"):
    model.eval()
    metric.reset()
    losses = []    for batch in data_loader:
        input_ids, token_type_ids, labels = batch
        probs = model(input_ids=input_ids, token_type_ids=token_type_ids)
        loss = criterion(probs, labels)
        losses.append(loss.numpy())
        correct = metric.compute(probs, labels)
        metric.update(correct)
        accu = metric.accumulate()    print("eval {} loss: {:.5}, accu: {:.5}".format(phase,
                                                    np.mean(losses), accu))
    model.train()
    metric.reset()

# 接下来，开始正式训练模型global_step = 0tic_train = time.time()for epoch in range(1, epochs + 1):    for step, batch in enumerate(train_data_loader, start=1):

        input_ids, token_type_ids, labels = batch
        probs = model(input_ids=input_ids, token_type_ids=token_type_ids)
        loss = criterion(probs, labels)
        correct = metric.compute(probs, labels)
        metric.update(correct)
        acc = metric.accumulate()

        global_step += 1
        
        # 每间隔 10 step 输出训练指标
        if global_step % 10 == 0:            print(                "global step %d, epoch: %d, batch: %d, loss: %.5f, accu: %.5f, speed: %.2f step/s"
                % (global_step, epoch, step, loss, acc,                    10 / (time.time() - tic_train)))
            tic_train = time.time()
        loss.backward()
        optimizer.step()
        lr_scheduler.step()
        optimizer.clear_grad()        # 每间隔 100 step 在验证集和测试集上进行评估
        if global_step % 100 == 0:
            evaluate(model, criterion, metric, dev_data_loader, "dev")            
# 训练结束后，存储模型参数save_dir = os.path.join("checkpoint", "model_%d" % global_step)
os.makedirs(save_dir)

save_param_path = os.path.join(save_dir, 'model_state.pdparams')
paddle.save(model.state_dict(), save_param_path)
tokenizer.save_pretrained(save_dir)

global step 5310, epoch: 3, batch: 1578, loss: 0.31671, accu: 0.95000, speed: 0.63 step/sglobal step 5320, epoch: 3, batch: 1588, loss: 0.36240, accu: 0.94063, speed: 6.98 step/sglobal step 5330, epoch: 3, batch: 1598, loss: 0.41451, accu: 0.93854, speed: 7.40 step/sglobal step 5340, epoch: 3, batch: 1608, loss: 0.31327, accu: 0.94063, speed: 7.01 step/sglobal step 5350, epoch: 3, batch: 1618, loss: 0.40664, accu: 0.93563, speed: 7.83 step/sglobal step 5360, epoch: 3, batch: 1628, loss: 0.33064, accu: 0.93958, speed: 7.34 step/sglobal step 5370, epoch: 3, batch: 1638, loss: 0.38411, accu: 0.93795, speed: 7.72 step/sglobal step 5380, epoch: 3, batch: 1648, loss: 0.35376, accu: 0.93906, speed: 7.92 step/sglobal step 5390, epoch: 3, batch: 1658, loss: 0.39706, accu: 0.93924, speed: 7.47 step/sglobal step 5400, epoch: 3, batch: 1668, loss: 0.41198, accu: 0.93781, speed: 7.41 step/seval dev loss: 0.4177, accu: 0.89082global step 5410, epoch: 3, batch: 1678, loss: 0.34453, accu: 0.93125, speed: 0.63 step/sglobal step 5420, epoch: 3, batch: 1688, loss: 0.34569, accu: 0.93906, speed: 7.75 step/sglobal step 5430, epoch: 3, batch: 1698, loss: 0.39160, accu: 0.92917, speed: 7.54 step/sglobal step 5440, epoch: 3, batch: 1708, loss: 0.46002, accu: 0.93125, speed: 7.05 step/sglobal step 5450, epoch: 3, batch: 1718, loss: 0.32302, accu: 0.93188, speed: 7.14 step/sglobal step 5460, epoch: 3, batch: 1728, loss: 0.40802, accu: 0.93281, speed: 7.22 step/sglobal step 5470, epoch: 3, batch: 1738, loss: 0.34607, accu: 0.93348, speed: 7.44 step/sglobal step 5480, epoch: 3, batch: 1748, loss: 0.34709, accu: 0.93398, speed: 7.38 step/sglobal step 5490, epoch: 3, batch: 1758, loss: 0.31814, accu: 0.93437, speed: 7.39 step/sglobal step 5500, epoch: 3, batch: 1768, loss: 0.42689, accu: 0.93125, speed: 7.74 step/seval dev loss: 0.41789, accu: 0.88968

# 下载我们基于 Lcqmc 事先训练好的语义匹配模型并解压! wget https://paddlenlp.bj.bcebos.com/models/text_matching/ernie_gram_zh_pointwise_matching_model.tar
! tar -xvf ernie_gram_zh_pointwise_matching_model.tar

# 测试数据由 2 列文本构成 tab 分隔# Lcqmc 默认下载到如下路径! head -n10 "${HOME}/.paddlenlp/datasets/LCQMC/lcqmc/lcqmc/test.tsv"

def predict(model, data_loader):
    
    batch_probs = []    # 预测阶段打开 eval 模式，模型中的 dropout 等操作会关掉
    model.eval()    with paddle.no_grad():        for batch_data in data_loader:
            input_ids, token_type_ids = batch_data
            input_ids = paddle.to_tensor(input_ids)
            token_type_ids = paddle.to_tensor(token_type_ids)            
            # 获取每个样本的预测概率: [batch_size, 2] 的矩阵
            batch_prob = model(
                input_ids=input_ids, token_type_ids=token_type_ids).numpy()

            batch_probs.append(batch_prob)
        batch_probs = np.concatenate(batch_probs, axis=0)        return batch_probs

# 预测数据的转换函数# predict 数据没有 label, 因此 convert_exmaple 的 is_test 参数设为 Truetrans_func = partial(
    convert_example,
    tokenizer=tokenizer,
    max_seq_length=512,
    is_test=True)# 预测数据的组 batch 操作# predict 数据只返回 input_ids 和 token_type_ids，因此只需要 2 个 Pad 对象作为 batchify_fnbatchify_fn = lambda samples, fn=Tuple(
    Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input_ids
    Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # segment_ids): [data for data in fn(samples)]# 加载预测数据test_ds = load_dataset("lcqmc", splits=["test"])

batch_sampler = paddle.io.BatchSampler(test_ds, batch_size=32, shuffle=False)# 生成预测数据 data_loaderpredict_data_loader =paddle.io.DataLoader(
        dataset=test_ds.map(trans_func),
        batch_sampler=batch_sampler,
        collate_fn=batchify_fn,
        return_list=True)

pretrained_model = paddlenlp.transformers.ErnieGramModel.from_pretrained('ernie-gram-zh')

model = PointwiseMatching(pretrained_model)

# 刚才下载的模型解压之后存储路径为 ./ernie_gram_zh_pointwise_matching_model/model_state.pdparamsstate_dict = paddle.load("./ernie_gram_zh_pointwise_matching_model/model_state.pdparams")# 刚才下载的模型解压之后存储路径为 ./pointwise_matching_model/ernie1.0_base_pointwise_matching.pdparams# state_dict = paddle.load("pointwise_matching_model/ernie1.0_base_pointwise_matching.pdparams")model.set_dict(state_dict)

for idx, batch in enumerate(predict_data_loader):    if idx < 1:        print(batch)

# 执行预测函数y_probs = predict(model, predict_data_loader)# 根据预测概率获取预测 labely_preds = np.argmax(y_probs, axis=1)

# 我们按照千言文本相似度竞赛的提交格式将预测结果存储在 lcqmc.tsv 中，用来后续提交# 同时将预测结果输出到终端，便于大家直观感受模型预测效果test_ds = load_dataset("lcqmc", splits=["test"])with open("lcqmc.tsv", 'w', encoding="utf-8") as f:
    f.write("index\tprediction\n")    
    for idx, y_pred in enumerate(y_preds):
        f.write("{}\t{}\n".format(idx, y_pred))
        text_pair = test_ds[idx]
        text_pair["label"] = y_pred        print(text_pair)

# 打包预测结果!zip submit.zip lcqmc.tsv paws-x.tsv bq_corpus.tsv

from paddlenlp.data import Vocab# 从文件构建vocab1 = Vocab.load_vocabulary(vocab_file_path)# 从字典构建# dic = {'unk':0, 'pad':1, 'bos':2, 'eos':3, ...}vocab2 = Vocab.from_dict(dic)# 从json构建，一般是已构建好的Vocab对象先保存为json_str或json文件后再进行恢复# json_str方式json_str = vocab1.to_json()
vocab3 = Vocab.from_json(json_str)# json文件方式vocab1.to_json(json_file_path)
vocab4 = Vocab.from_json(json_file_path)

from paddlenlp.data import Vocab, JiebaTokenizer# 词表文件路径，运行示例程序可先下载词表文件# wget https://paddlenlp.bj.bcebos.com/data/senta_word_dict.txtvocab_file_path = './senta_word_dict.txt'# 构建词表vocab = Vocab.load_vocabulary(
    vocab_file_path,
    unk_token='[UNK]',
    pad_token='[PAD]')
tokenizer = JiebaTokenizer(vocab)
tokens = tokenizer.cut('我爱你中国') # ['我爱你', '中国']ids = tokenizer.encode('我爱你中国') # [1170578, 575565]

from paddlenlp.data import SamplerHelperfrom paddle.io import Datasetclass MyDataset(Dataset):
    def __init__(self):
        super(MyDataset, self).__init__()
        self.data = [
            [[1, 2, 3, 4], [1]],
            [[5, 6, 7], [0]],
            [[8, 9], [1]],
        ]    def __getitem__(self, index):
        data = self.data[index][0]
        label = self.data[index][1]        return data, label    def __len__(self):
        return len(self.data)

dataset = MyDataset()# SamplerHelper返回的是数据索引的可迭代对象，产生的迭代的索引为：[0, 1, 2]sampler = SamplerHelper(dataset)# `shuffle()`的作用是随机打乱索引顺序，产生的迭代的索引为：[0, 2, 1]sampler = sampler.shuffle()# sort()的作用是按照指定key为排序方式并在buffer_size大小个样本中排序# 示例中以样本第一个字段的长度进行升序排序，产生的迭代的索引为：[2, 0, 1]key = (lambda x, data_source: len(data_source[x][0]))
sampler = sampler.sort(key=key, buffer_size=2)# batch()的作用是按照batch_size组建mini-batch，产生的迭代的索引为：[[2, 0], [1]]sampler = sampler.batch(batch_size=2)# shard()的作用是为多卡训练切分数据集，当前卡产生的迭代的索引为：[[2, 0]]sampler = sampler.shard(num_replicas=2)

from paddlenlp.data import Stack
a = [1, 2, 3, 4]
b = [3, 4, 5, 6]
c = [5, 6, 7, 8]
result = Stack()([a, b, c])"""
[[1, 2, 3, 4],
 [3, 4, 5, 6],
 [5, 6, 7, 8]]
"""

from paddlenlp.data import Pad
a = [1, 2, 3, 4]
b = [5, 6, 7]
c = [8, 9]
result = Pad(pad_val=0)([a, b, c])"""
[[1, 2, 3, 4],
 [5, 6, 7, 0],
 [8, 9, 0, 0]]
"""

from paddlenlp.data import Stack, Pad, Tupledata = [
        [[1, 2, 3, 4], [1]],
        [[5, 6, 7], [0]],
        [[8, 9], [1]],
       ]
batchify_fn = Tuple(Pad(pad_val=0), Stack())
ids, label = batchify_fn(data)"""
ids:
[[1, 2, 3, 4],
 [5, 6, 7, 0],
 [8, 9, 0, 0]]
label: [[1], [0], [1]]
"""

from paddlenlp.data import Stack, Pad, Dictdata = [
        {'labels':[1], 'token_ids':[1, 2, 3, 4]},
        {'labels':[0], 'token_ids':[5, 6, 7]},
        {'labels':[1], 'token_ids':[8, 9]},
       ]
batchify_fn = Dict({'token_ids':Pad(pad_val=0), 'labels':Stack()})
ids, label = batchify_fn(data)"""
ids:
[[1, 2, 3, 4],
 [5, 6, 7, 0],
 [8, 9, 0, 0]]
label: [[1], [0], [1]]
"""

from paddlenlp.data import Vocab, JiebaTokenizer, Stack, Pad, Tuple, SamplerHelperfrom paddlenlp.datasets import load_datasetfrom paddle.io import DataLoader# 词表文件路径，运行示例程序可先下载词表文件# wget https://paddlenlp.bj.bcebos.com/data/senta_word_dict.txtvocab_file_path = './senta_word_dict.txt'# 构建词表vocab = Vocab.load_vocabulary(
    vocab_file_path,
    unk_token='[UNK]',
    pad_token='[PAD]')# 初始化分词器tokenizer = JiebaTokenizer(vocab)def convert_example(example):
    text, label = example['text'], example['label']
    ids = tokenizer.encode(text)
    label = [label]    return ids, label

dataset = load_dataset('chnsenticorp', splits='train')
dataset = dataset.map(convert_example, lazy=True)

pad_id = vocab.token_to_idx[vocab.pad_token]
batchify_fn = Tuple(
    Pad(axis=0, pad_val=pad_id),  # ids
    Stack(dtype='int64')  # label)

batch_sampler = SamplerHelper(dataset).shuffle().batch(batch_size=16)
data_loader = DataLoader(
    dataset,
    batch_sampler=batch_sampler,
    collate_fn=batchify_fn,
    return_list=True)# 测试数据集for batch in data_loader:
    ids, label = batch    print(ids.shape, label.shape)    print(ids)    print(label)    break

<br/>