【NLP+Android】AI 对联 APP

该项目是对联生成系统的升级版本，从简单界面可视化升级为Android APP。项目用Seq2seq框架，先处理对联数据，建立语料库和字典，划分数据集并封装。接着搭建Encoder、AttentionLayer等网络组件，定义损失函数和超参数后训练模型。还涉及模型预测及动态图转静态图，最后进行Android开发，生成opt模型并编写相关代码实现功能。

项目简介

该项目是我之前精品项目使用Seq2seq框架搭建对联生成系统的升级版本。在之前的项目中，我仅实现了简单的界面可视化，现在升级到了Android APP，嘿嘿！

安卓代码及APP下载地址：点我

APP界面如下图所示:

       

In [1]

import ioimport osimport numpy as npimport paddleimport paddlenlpfrom functools import partialfrom paddle.static import InputSpec

   

1.数据处理

In [2]

data_in_path="/home/aistudio/data/data110057/fixed_couplets_in.txt"data_out_path="/home/aistudio/data/data110057/fixed_couplets_out.txt"

   

In [3]

def openfile(src):    with open(src,'r',encoding="utf-8") as source:        lines=source.readlines()    return lines

   

In [4]

data_in=openfile(data_in_path)data_out=openfile(data_out_path)

   

In [5]

print(len(data_in))print(len(data_out))print(data_in[0])print(data_out[0])print(len(data_in[0]))

       

744915744915腾 飞 上 铁 ， 锐 意 改 革 谋 发 展 ， 勇 当 千 里 马 和 谐 南 供 ， 安 全 送 电 保 畅 通 ， 争 做 领 头 羊 37

       

In [6]

def delete_newline_and_space(lista):     newlist=[]    for i in range(len(lista)):        newlist.append([""]+lista[i].strip().split()+[''])    return newlist

   

In [7]

data_in_nospace=delete_newline_and_space(data_in)data_out_nospace=delete_newline_and_space(data_out)print(data_in_nospace[0])print(data_out_nospace[0])

       

['', '腾', '飞', '上', '铁', '，', '锐', '意', '改', '革', '谋', '发', '展', '，', '勇', '当', '千', '里', '马', '']['', '和', '谐', '南', '供', '，', '安', '全', '送', '电', '保', '畅', '通', '，', '争', '做', '领', '头', '羊', '']

       

计算最长的对联长度couplet_maxlen,并将该长度+2作为向量长。不足进行填充。

In [8]

couplet_maxlen=max([len(i) for i in data_in_nospace])couplet_maxlen

       

34

               

1.1 建立语料库、字符转id的字典和id转字符的字典

• 字符主要指的是汉字，当然还有标点

有个问题： 输入输出的语料库是二者分别建立一个，还是二者一起建立一个？

在这里建立一个统一的语料库进行实验。（毕设的时候我是分开建的，不知道哪个做法正确）

In [9]

def bulid_cropus(data_in,data_out):    crpous=[]    for i in data_in:        crpous.extend(i)    for i in data_out:        crpous.extend(i)    return crpous

   

In [10]

def build_dict(corpus,frequency):    # 首先统计不同词（汉字）的频率，使用字典记录    word_freq_dict={}    for ch in corpus:        if ch not in word_freq_dict:            word_freq_dict[ch]=0        word_freq_dict[ch]+=1        # 根据频率对字典进行排序    word_freq_dict=sorted(word_freq_dict.items(),key=lambda x:x[1],reverse=True)            word2id_dict={}    id2word_dict={}        # 按照频率，从高到低，开始遍历每个单词，并赋予第一无二的 id    for word,freq in word_freq_dict:        if freq>frequency:            curr_id=len(word2id_dict)            word2id_dict[word]=curr_id            id2word_dict[curr_id]=word        else:             # else 部分在 使 单词 指向unk,对于汉字，我们不设置unk,令frequency=0            word2id_dict[word]=1    return word2id_dict,id2word_dict

   

In [11]

word_frequency=0word2id_dict,id2word_dict=build_dict(bulid_cropus(data_in_nospace,data_out_nospace),word_frequency)

   

词汇量大小

In [12]

word_size=len(word2id_dict)id_size=len(id2word_dict)print("汉字个数：",word_size,"
 id个数：",id_size)

       

汉字个数： 9017  id个数： 9017

       

In [13]

with open("word2id.txt",'w',encoding='utf-8') as w2i:    for k,v in word2id_dict.items():        w2i.write(str(k)+","+str(v)+'
')with open("id2word.txt",'w',encoding='utf-8') as w2i:    for k,v in id2word_dict.items():        w2i.write(str(k)+","+str(v)+'
')

   

In [14]

print(word2id_dict[''])print(word2id_dict[''])

       

10

       

创建 tensor

In [15]

def getensor(w2i,datalist,maxlength):    in_tensor=[]    for lista in datalist:        in_samll_tensor=[]        for li in lista:            in_samll_tensor.append(w2i[li])#         if len(in_samll_tensor)']]*(maxlength-len(in_samll_tensor))        in_tensor.append(in_samll_tensor)    return in_tensor

   

In [16]

in_tensor=getensor(word2id_dict,data_in_nospace,couplet_maxlen)out_tensor=getensor(word2id_dict,data_out_nospace,couplet_maxlen)

   

转成数字，带上shape属性

In [17]

in_tensor=np.array(in_tensor)out_tensor=np.array(out_tensor)

   

1.2 划分训练集、验证集、测试集 ，按照8:1:1 固定划分

In [18]

train_in_tensor=in_tensor[:595933]val_in_tensor=in_tensor[595933:670424]test_in_tensor=in_tensor[670424:]train_out_tensor=out_tensor[:595933]val_out_tensor=out_tensor[595933:670424]test_out_tensor=out_tensor[670424:]

   

In [19]

print(len(train_in_tensor),len(test_in_tensor),len(val_in_tensor))

       

595933 74491 74491

       

1.3 封装数据集为可直接进行训练的dataset

In [20]

# 1.继承paddle.io.Datasetclass Mydataset(paddle.io.Dataset):        # 2. 构造函数，定义数据集大小    def __init__(self,first,second):        super(Mydataset,self).__init__()        self.first=first        self.second=second            # 3. 实现__getitem__方法，定义指定index时如何获取数据，并返回单条数据（训练数据，对应的标签）    def __getitem__(self,index):        return self.first[index],self.second[index]        # 4. 实现__len__方法，返回数据集总数目    def __len__(self):        return self.first.shape[0]

   

In [21]

train_tensor=Mydataset(train_in_tensor,train_out_tensor)val_tensor=Mydataset(val_in_tensor,val_out_tensor)test_tensor=Mydataset(test_in_tensor,test_out_tensor)

   

数据加载

In [22]

BATCH_SIZE=64padid=word2id_dict['']

   

In [23]

def prepare_input(inputs,padid):    src,src_length=paddlenlp.data.Pad(pad_val=padid,ret_length=True)([inputsub[0] for inputsub in inputs])    trg,trg_length=paddlenlp.data.Pad(pad_val=padid,ret_length=True)([inputsub[1] for inputsub in inputs])    #src=src.astype(paddle.get_default_dtype())    trg_mask =(trg[:,:-1]!=padid).astype(paddle.get_default_dtype())    return src,src_length,trg[:,:-1],trg[:,1:,np.newaxis],trg_mask

   

In [24]

def create_data_loader(dataset):    data_loader=paddle.io.DataLoader(dataset,batch_sampler=None,batch_size=BATCH_SIZE,collate_fn=partial(prepare_input, padid=padid))    return data_loader

   

In [25]

train_loader=create_data_loader(train_tensor)val_loader=create_data_loader(val_tensor)test_loader=create_data_loader(test_tensor)

   

In [26]

# j=0# for i in train_loader:#     print(len(i))#     for ind,each in enumerate(i):#         print(ind,each.shape,each)#         #print(ind,each.shape)#     j+=1#     if j==2:#         break

   

In [ ]

   

In [27]

# for i in train_loader:#     x,x_length,y,_,_= i#     break# # print(x)

   

2.网络搭建

主要参考的是最新的项目，直达：https://aistudio.baidu.com/aistudio/projectdetail/1321118?shared=1

2.1 Encoder

In [28]

class Encoder(paddle.nn.Layer):    def __init__(self,vocab_size,embedding_dim,hidden_size,num_layers):        super(Encoder,self).__init__()                self.embedding=paddle.nn.Embedding(vocab_size,embedding_dim)        self.lstm=paddle.nn.LSTM(input_size=embedding_dim,                                hidden_size=hidden_size,                                num_layers=num_layers,                                dropout=0.2 if num_layers>1 else 0)                # src_length 的形状为[batch_size],作用是控制inputs中的time_step超过[batch_size]的不再更新状态，就是那些填充    def forward(self,src,src_length):        inputs=self.embedding(src)  # [batch_size,time_steps,embedding_dim]        encoder_out,encoder_state=self.lstm(inputs,sequence_length=src_length) # out[batch_szie,time_steps,hidden_size] state:[[num_layers*1,batch_size,hidden_size],[num_layers*1,batch_size,hidden_size]]        # encoder_out,encoder_state=self.lstm(inputs)        return encoder_out,encoder_state

   

In [29]

# encoder=Encoder(word_size,256,128,2)# #paddle.summary(encoder,[(64,18),(64)],dtypes='int64')# out,state=encoder(x,x_length)# print(out.shape)# print(state)

   

2.2 注意力层

In [30]

class AttentionLayer(paddle.nn.Layer):    def __init__(self,hidden_size):        super(AttentionLayer,self).__init__()        self.attn1=paddle.nn.Linear(hidden_size,hidden_size)        self.attn2=paddle.nn.Linear(hidden_size+hidden_size,hidden_size)    def forward(self,decoder_hidden_h,encoder_output,encoder_padding_mask):                encoder_output=self.attn1(encoder_output) # [batch_size,time_steps,hidden_size]                # decodr_hidden_h 的形状 [batch_size,hidden_size],是lstm公式中的ht.        # unsqueeze之后[batch_size,1,hidden_size]        # transpose_y=True,后两维转置 [batch_size,hidden_size,time_steps]        # matmul之后的 形状 [batch_size,1,time_steps]        a=paddle.unsqueeze(decoder_hidden_h,[1])        # print(a.shape)        # print(encoder_output.shape)        attn_scores=paddle.matmul(a,encoder_output,transpose_y=True)                        # 注意力机制中增加掩码操作，在padding 位加上个非常小的数：-1e9        if encoder_padding_mask is not None:            # encoder_padding_mask的形状为[batch_size,1,time_steps]            attn_scores=paddle.add(attn_scores,encoder_padding_mask)        # softmax操作，默认是最后一个维度，axis=-1,形状不变        attn_scores=paddle.nn.functional.softmax(attn_scores)         # [batch_size,1,time_steps]*[batch_size,time_steps,hidden_size]-->[batch_size,1,hidden_size]        # squeeze之后：[batch_size,hidden_size]        attn_out=paddle.squeeze(paddle.matmul(attn_scores,encoder_output),[1])                # concat之后 [batch_size,hidden_size+hidden_size]        attn_out=paddle.concat([attn_out,decoder_hidden_h],1)        # 最终结果[batch_size,hidden_size]        attn_out=self.attn2(attn_out)        return attn_out

   

2.3 解码器单元

In [31]

class DecoderCell(paddle.nn.RNNCellBase):    def __init__(self,num_layers,embedding_dim,hidden_size):        super(DecoderCell,self).__init__()        self.dropout=paddle.nn.Dropout(0.2)        self.lstmcells=paddle.nn.LayerList([paddle.nn.LSTMCell(            input_size=embedding_dim+hidden_size if i==0 else hidden_size,            hidden_size=hidden_size        ) for i in range(num_layers)])        self.attention=AttentionLayer(hidden_size)        def forward(self,decoder_input,decoder_initial_states,encoder_out,encoder_padding_mask=None):        #forward 函数会执行squence_len次 ，每次的decoder_input 为[batch_size,embeddding_dim]        # 状态分解 states [encoder_final_states,decoder_init_states]        # encoder_final_states [num_layes,batch_size,hiden_size] ???        # decoder_init_states [] ???        encoder_final_states,decoder_init_states=decoder_initial_states        #num_layers=len(encoder_final_states[0])        #decoder_init_states=lstm_init_state        # ???        new_lstm_states=[]        # decoder_input: [batch_size,embedding_dim]        # print("decodercell ",decoder_input.shape)        inputs=paddle.concat([decoder_input,decoder_init_states],1)        # print("concant之后",inputs.shape)        for i ,lstm_cell in enumerate(self.lstmcells):            # inputs 的形状为 [batch_size,input_size]  input_size:输入的大小            state_h,new_lstm_state=lstm_cell(inputs,encoder_final_states[i])            inputs=self.dropout(state_h)            new_lstm_states.append(new_lstm_state)                state_h=self.attention(inputs,encoder_out,encoder_padding_mask)        # print(state_h.shape)        return state_h,[new_lstm_states,state_h]

   

2.4 解码器

解码器由embedding+解码器单元+线性输出层组成

In [32]

class Decoder(paddle.nn.Layer):    def __init__(self,vocab_size,embedding_dim,hidden_size,num_layers):        super(Decoder,self).__init__()        self.embedding=paddle.nn.Embedding(vocab_size,embedding_dim)        self.lstm_attention=paddle.nn.RNN(DecoderCell(num_layers,embedding_dim,hidden_size))        self.fianl=paddle.nn.Linear(hidden_size,vocab_size)    def forward(self,trg, decoder_initial_states,encoder_output,encoder_padding_mask):        # trg 的形状为 [batch_size,sequence_length]        # embedding 之后, [batch_size,sequence_length,embedding_dim]        inputs=self.embedding(trg)        # print("embedding 后的 输入维度",inputs.shape)                # decodr_out [batch_szie,hidden_size]        decoder_out,_ = self.lstm_attention(inputs,                                         initial_states=decoder_initial_states,                                         encoder_out=encoder_output,                                         encoder_padding_mask=encoder_padding_mask)        # predict [batch_size,sequence_len,word_size]        predict=self.fianl(decoder_out)        # print("最后的维度",decoder_out.shape)        return predict

   

2.5 组装Seq2Seq

In [33]

class Seq2Seq(paddle.nn.Layer):    def __init__(self, vocab_size,embedding_dim,hidden_size,num_layers,eos_id):                super(Seq2Seq,self).__init__()        self.hidden_size=hidden_size        self.eos_id=eos_id        self.num_layers=num_layers        self.INF= 1e9        self.encoder=Encoder(vocab_size,embedding_dim,hidden_size,num_layers)        self.decoder=Decoder(vocab_size,embedding_dim,hidden_size,num_layers)            def forward(self,src,src_length,trg):        # encoder_output 的形状为[batch_size,sequence_len,hidden_size]        # encoder_final_state ([num_layers*1,batch_size,hidden_size],[num_layers*1,batch_size,hidden_size]])  tuple类型        encoder_output,encoder_final_state=self.encoder(src,src_length)        encoder_final_states=[(encoder_final_state[0][i],encoder_final_state[1][i]) for i in range(self.num_layers)]        #print(encoder_final_states[0])        # [batch_size,hidden_size] 初始化为0        #lstm_init_state= self.decoder.lstm_attention.cell.get_initial_states(batch_ref=encoder_output,shape=[self.hidden_size])               decoder_initial_states=[encoder_final_states,                                self.decoder.lstm_attention.cell.get_initial_states(batch_ref=encoder_output,shape=[self.hidden_size])]        src_mask=(src!=self.eos_id).astype(paddle.get_default_dtype())        encoder_mask=(src_mask-1)*self.INF        encoder_padding_mask=paddle.unsqueeze(encoder_mask,[1])        predict=self.decoder(trg,decoder_initial_states,encoder_output,encoder_padding_mask)        return predict

   

2.6 自定义交叉熵损失函数及超参数

In [34]

class CrossEntropy(paddle.nn.Layer):    def __init__(self):        super(CrossEntropy,self).__init__()    def forward(self,pre,real,trg_mask):        # 返回的数据类型与pre一致，除了axis维度(未指定则为-1)，其他维度也与pre一致        # logits=pre,[batch_size,sequence_len,word_size],猜测会进行argmax操作，[batch_size,sequence_len,1]        # 默认的soft_label为False，lable=real,[bacth_size,sequence_len,1]        cost=paddle.nn.functional.softmax_with_cross_entropy(logits=pre,label=real)                # 删除axis=2 shape上为1的维度        # 返回结果的形状应为 [batch_size,sequence_len]        cost=paddle.squeeze(cost,axis=[2])        # trg_mask 的形状[batch_size,suqence_len]        # * 这个星号应该是对应位置相乘，返回结果的形状 [bathc_szie,sequence_len]        masked_cost=cost*trg_mask        # paddle.mean 对应轴的对应位置求平均， 在这里返回结果为 [sequence_len]        # paddle.sum 求解方法与paddle.mean一致，最终返回的结果应为[1]        return paddle.sum(paddle.mean(masked_cost,axis=[0]))

   

In [35]

epochs=20eos_id=word2id_dict['']num_layers=2dropout_rate=0.2hidden_size=128embedding_dim=256max_grad_norm=5lr=0.001log_freq=200model_path='./train_model/train_model'

   

In [36]

s2s=Seq2Seq(word_size,embedding_dim,hidden_size,num_layers,eos_id)

       

W0609 11:31:02.978509  1421 device_context.cc:404] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 10.1W0609 11:31:02.982554  1421 device_context.cc:422] device: 0, cuDNN Version: 7.6.

       

In [37]

model=paddle.Model(s2s)# model.parameters() 返回一个包含所有模型参数的列表optimizer=paddle.optimizer.Adam(learning_rate=lr,parameters=model.parameters())# 困惑度ppl_metric=paddlenlp.metrics.Perplexity()model.prepare(optimizer,CrossEntropy(),ppl_metric)

   

2.7 训练并保存

#eval_freq 多少个epoch评估一次 #save_freq 多少个epoch保存模型一次

model.fit(train_data=train_loader, eval_data=val_loader, epochs=epochs, eval_freq=1, save_freq=2, save_dir=model_path, log_freq=log_freq, verbose=2, callbacks=[paddle.callbacks.VisualDL('./log')])

#保存用于预测的模型 #model.save("./infer_model/infer_model",False)

3.模型预测

3.1 定义预测模型

In [38]

class Seq2SeqInfer(Seq2Seq):    def __init__(self,word_size,embedding_dim,hidden_size,num_layers,bos_id,eos_id,beam_size,max_out_len=couplet_maxlen):        self.bos_id=bos_id        self.beam_size=beam_size        self.max_out_len=max_out_len        self.num_layers=num_layers        super(Seq2SeqInfer,self).__init__(word_size,embedding_dim,hidden_size,num_layers,eos_id)        self.beam_search_decoder=paddle.nn.BeamSearchDecoder(            self.decoder.lstm_attention.cell,            start_token=bos_id,            end_token=eos_id,            beam_size=beam_size,            embedding_fn=self.decoder.embedding,            output_fn=self.decoder.fianl)        def forward(self,src,src_length):        encoder_output,encoder_states=self.encoder(src,src_length)        encoder_final_state=[(encoder_states[0][i],encoder_states[1][i]) for i in range(self.num_layers)]        # 初始化decoder的隐藏层状态        decoder_initial_states=[encoder_final_state,                                self.decoder.lstm_attention.cell.get_initial_states(batch_ref=encoder_output,shape=[self.hidden_size])]                src_mask=(src!=self.eos_id).astype(paddle.get_default_dtype())        encoder_padding_mask=(src_mask-1.0)*self.INF        encoder_padding_mask=paddle.unsqueeze(encoder_padding_mask,[1])        # 扩展tensor的bacth维度        encoder_out=paddle.nn.BeamSearchDecoder.tile_beam_merge_with_batch(encoder_output,self.beam_size)        encoder_padding_mask=paddle.nn.BeamSearchDecoder.tile_beam_merge_with_batch(encoder_padding_mask,self.beam_size)        seq_output,_=paddle.nn.dynamic_decode( decoder=self.beam_search_decoder,                                               inits= decoder_initial_states,                                               max_step_num= self.max_out_len,                                               encoder_out=encoder_output,                                               encoder_padding_mask=encoder_padding_mask)                return seq_output

   

In [39]

def pre_process(seq,bos_idx,eos_idx):    #print(bos_idx,eos_idx)    # 结束位置    eos_pos=len(seq)-1    for i ,idx in enumerate(seq):        #print(i,idx[0])        if idx==eos_idx: # 遇到结束标志            eos_pos=i            break    seq=[idx[0] for idx in seq[:eos_pos] if (idx !=bos_idx) ]    return seq

   

3.2 预测超参数

In [40]

beam_size=1bos_id=word2id_dict['']eos_id=word2id_dict['']max_out_len=couplet_maxlenprint(bos_id)

       

0

       

3.3 动态图转静态图

In [41]

x=paddle.to_tensor([[0  , 566, 566, 489, 42 , 165, 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  ]],dtype=paddle.int32)y=paddle.to_tensor([7],dtype=paddle.int32)print(x)print(y)

       

Tensor(shape=[1, 34], dtype=int32, place=CUDAPlace(0), stop_gradient=True,       [[0  , 566, 566, 489, 42 , 165, 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  ]])Tensor(shape=[1], dtype=int32, place=CUDAPlace(0), stop_gradient=True,       [7])

       

In [42]

s2si=Seq2SeqInfer(word_size,embedding_dim,hidden_size,num_layers,bos_id,eos_id,beam_size,max_out_len)dd=paddle.load('./trained_model/10.pdparams')s2si.load_dict(dd)net=paddle.jit.to_static(s2si)out=net(x,y)out

       

Tensor(shape=[1, 6, 1], dtype=int64, place=CUDAPlace(0), stop_gradient=False,       [[[4 ],         [24],         [11],         [75],         [7 ],         [1 ]]])

               

In [43]

paddle.jit.save(net, './trained_model/net')

   

4.Android 开发

4.1 opt模型生成及优化

! pip install paddlelite==2.11-rc

! paddle_lite_opt --model_file=./trained_model/net.pdmodel --param_file=./trained_model/net.pdiparams --optimize_out=./trained_model/v1_opt·

4.2 下载预编译的预测库

可以参考最新文档 Java完整示例

我使用的是2.11-rc版本

4.3 主代码展示

package com.baidu.paddle.lite;import android.content.Context;import android.support.v7.app.AppCompatActivity;import android.os.Bundle;import android.util.Log;import android.view.View;import android.widget.Button;import android.widget.EditText;import android.widget.TextView;import java.io.BufferedOutputStream;import java.io.BufferedReader;import java.io.File;import java.io.FileInputStream;import java.io.FileOutputStream;import java.io.FileReader;import java.io.IOException;import java.io.InputStream;import java.io.InputStreamReader;import java.io.OutputStream;import java.util.Date;import java.util.HashMap;import java.util.Map;public class MainActivity extends AppCompatActivity {    public static final String TAG = "MainActivity";    private Button bt;    private EditText et;    public Map word_id_map = new HashMap();    public Map id_word_map = new HashMap();    @Override    protected void onCreate(Bundle savedInstanceState) {        super.onCreate(savedInstanceState);        setContentView(R.layout.activity_main);        //bt=(Button)findViewById(R.id.bt1);//        String path2="assets/word2id.txt";//        String path2="assets/id2word.txt";        String np1 = copyFromAssetsToCache("word2id.txt", this);        String np2 = copyFromAssetsToCache("id2word.txt", this);        try {            word_id_map=readTxtToObject(np1);        } catch (IOException e) {            e.printStackTrace();        }        try {            id_word_map=readTxtToObject(np2);        } catch (IOException e) {            e.printStackTrace();        }    }    public void click(View v){        int id=v.getId();        switch (id) {            case R.id.bt1:                Log.i("指定onClick属性方式","bt1点击事件");                et=(EditText) findViewById(R.id.text_in);                String ss=et.getText().toString();//                Log.i("指定onClick属性方式",ss);//                Log.i("指定onClick属性方式", String.valueOf(ss.length()));//                Log.i("指定onClick属性方式",word_id_map.toString());                int[] buffer1 = new int[]{0  , 1, 1, 1, 1 , 1, 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  ,                        1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1 };                for(int i=0;i readTxtToObject(String ppath) throws IOException {        Map map = new HashMap();        File f=new File(ppath);        BufferedReader reader=new BufferedReader(new FileReader(f));        String lineTxt=null;        while((lineTxt=reader.readLine())!=null){            String[] names = lineTxt.split(" ");            map.put(names[0],  names[1]);        }        return map;    }    public static String getVersionInfo(String modelName, Context context) {        String modelPath = copyFromAssetsToCache(modelName, context);        //Log.d(TAG,modelPath);        System.out.println(modelPath);        MobileConfig config = new MobileConfig();        config.setModelFromFile(modelPath);        PaddlePredictor predictor = PaddlePredictor.createPaddlePredictor(config);        return predictor.getVersion();        //return modelPath;    }    public static String copyFromAssetsToCache(String modelPath, Context context) {        //context.getCacheDir()：获取应用缓存目录        String newPath = context.getCacheDir() + "/" + modelPath;        //创建file对象        File desDir = new File(newPath);        try {                // context.getAssets().open() 打开assets目录下的文件                // Inputstream 字节输入流的最顶层父类                InputStream stream = context.getAssets().open(modelPath);                // 创建BufferedOutputStream字节缓冲输出流                OutputStream output = new BufferedOutputStream(new FileOutputStream(newPath));                byte data[] = new byte[1024];                int count;                while ((count = stream.read(data)) != -1) {                    output.write(data, 0, count);                }                output.flush();//刷新缓冲输出流                output.close();//关闭流                stream.close();        } catch (Exception e) {            throw new RuntimeException(e);        }        return desDir.getPath();    }    public static Tensor runModel(String modelName, long[] dims1, int[] inputBuffer1,Context context) {//    public static Tensor runModel(String modelName, long[] dims1, int[] inputBuffer1,//        long[] dims2, int[] inputBuffer2, Context context) {        String modelPath = copyFromAssetsToCache(modelName, context);        MobileConfig config = new MobileConfig();        config.setModelFromFile(modelPath);        config.setPowerMode(PowerMode.LITE_POWER_HIGH);        config.setThreads(1);        PaddlePredictor predictor = PaddlePredictor.createPaddlePredictor(config);//        System.out.println(predictor);//        System.out.println(predictor.getVersion());        Tensor input1 = predictor.getInput(0);        input1.resize(dims1);        input1.setData(inputBuffer1);        predictor.run();        Tensor output = predictor.getOutput(0);       return output;    }    public static Tensor setInputAndRunNaiveModel(String modelName, Context context) {        long[] dims1 = {1, 34};        int[] inputBuffer1 = new int[]{0  , 566, 566, 489, 42 , 165, 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  ,                1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1  , 1 };        return runModel(modelName, dims1, inputBuffer1, context);    }    public static String getSecond(String first,Context context){        String result="";        return result;    }}

   

4.4 界面布局代码展示