2019年5月28日 星期二

flask-based web interface deployment for pytorch chatbot 

### folder structure and flask setup
> ls 
data/  pytorch_chatbot/  save/  templates/  web.py

> ls templates/
template.html

> conda install Flask

> python web.py
 * Serving Flask app "web" (lazy loading)
 * Environment: production
   WARNING: Do not use the development server in a production environment.
   Use a production WSGI server instead.
 * Debug mode: off
 * Running on http://127.0.0.1:5000/ (Press CTRL+C to quit)


  1.  
  2. <html>
  3. <title>template.html</title>
  4. <body>
  5. <pre>
  6. Test page for pytorch chatbot on seq2seq dataset
  7. <form action='translate' method='post'>
  8. model: <input type='text' name='model' value='{{param["model"]}}' />
  9. epoch: <input type='text' name='epoch' value='{{param["epoch"]}}' />
  10. topn: <input type='text' name='topn' value='{{param["topn"]}}' />
  11. query: <input type='text' name='query' value='{{param["query"]}}'/>
  12. <input type='submit' value='translate' />
  13. </form>
  14. {{param['result']}}
  15. </pre>
  16. </body>
  17. </html>



  1.  
  2. ##########################
  3. # web.py
  4. #    > python web.py
  5. #########################
  6. from flask import Flask, request, render_template
  7.  
  8. import torch
  9. import random
  10. import pytorch_chatbot.main as pcm
  11. import pytorch_chatbot.evaluate as pce
  12. from pytorch_chatbot.train import indexesFromSentence
  13. from pytorch_chatbot.load import loadPrepareData
  14. from pytorch_chatbot.model import nn, EncoderRNN, LuongAttnDecoderRNN
  15.  
  16. import subprocess
  17. import json
  18.  
  19. def predictLoad(corpus, modelFile, n_layers=1, hidden_size=512):
  20.   print('corpus={}\nmodelFile={}'.format(corpus,modelFile))
  21.  
  22.   torch.set_grad_enabled(False)
  23.   voc, pairs = loadPrepareData(corpus)
  24.   embedding = nn.Embedding(voc.n_words, hidden_size)
  25.   encoder = EncoderRNN(voc.n_words, hidden_size, embedding, n_layers)
  26.   attn_model = 'dot'
  27.   decoder = LuongAttnDecoderRNN(attn_model, embedding, hidden_size, voc.n_words, n_layers)
  28.  
  29.   checkpoint = torch.load(modelFile)
  30.   encoder.load_state_dict(checkpoint['en'])
  31.   decoder.load_state_dict(checkpoint['de'])
  32.  
  33.   # train mode set to false, effect only on dropout, batchNorm
  34.   encoder.train(False)
  35.   decoder.train(False)
  36.  
  37.   #try:
  38.   encoder = encoder.to(device)
  39.   decoder = decoder.to(device)
  40.   #except:
  41.   #  print('cannot get encoder/decoder')
  42.   
  43.   return encoder, decoder, voc
  44.  
  45. def predict(encoder, decoder, voc, question, top):
  46.   result_list = []
  47.  
  48.   if(top==1):
  49.     beam_size = 1
  50.     output_words, _ = pce.evaluate(encoder, decoder, voc, question, beam_size)
  51.     answer = ' '.join(output_words)
  52.     answer = answer.replace('<EOS>','')
  53.     result_list.append(answer)
  54.     #print(output_words)
  55.   else:
  56.     beam_size = top
  57.     output_words_list = pce.evaluate(encoder, decoder, voc, question, beam_size)
  58.     count = 0;
  59.     for output_words, score in output_words_list:
  60.       count = count + 1
  61.       if(count <= top):
  62.         output_sentence = ' '.join(output_words)
  63.         output_sentence = output_sentence.replace('<EOS>','')
  64.         result_list.append(output_sentence)
  65.         #print(" {:.3f} < {}".format(score, output_sentence))
  66.   
  67.   return result_list
  68.  
  69. def filter(voc, question):
  70.   words = question.split()
  71.   result = []
  72.   for w in words:
  73.     if(w in voc.word2index):
  74.       result.append(w) 
  75.   return ' '.join(result)
  76.  
  77. # -------------------------------
  78. def sentence_test(voc,en,de,top,sentence):
  79.     source = sentence.rstrip()
  80.     seg_source = source
  81.     fil_source = filter(voc, seg_source)
  82.     target = predict(en, de, voc, fil_source, top)
  83.     result = "\nsource: '%s'\nfilter: '%s'\n" % (seg_source,fil_source)
  84.     
  85.     for answer in target:
  86.       result = result + "\t'%s'\n" % (answer)
  87.     
  88.     result = result + '\n'
  89.     return result
  90.  
  91. def sentence_test_model(seg_corpus_name, iteration, top, sentence):
  92.   n_layers = 1
  93.   hidden_size = 512
  94.  
  95.   modelFile = home_path + 'save/model/' + seg_corpus_name + '/1-1_512/' + str(iteration) + '_backup_bidir_model.tar'
  96.  
  97.   en, de, voc = predictLoad(seg_corpus_name, modelFile, n_layers, hidden_size)
  98.  
  99.   return sentence_test(voc,en,de,top,sentence)
  100.  
  101. def file_test(voc,en,de,top,test_file_name):
  102.   with open(test_file_name,"r") as f:
  103.     jp_data = f.readlines()
  104.  
  105.   for i,source in enumerate(jp_data):
  106.     source = source.rstrip()
  107.     seg_source = source
  108.     fil_source = filter(voc, seg_source)
  109.     target = predict(en, de, voc, fil_source, top)
  110.     print("%d:\nsource: '%s'\nfilter: '%s'" % (i+1,seg_source,fil_source))
  111.     
  112.     for answer in target:
  113.       print("\t'%s'" % (answer))
  114.  
  115. def file_test_model(seg_corpus_name, iteration, top, test_file_name):
  116.   n_layers = 1
  117.   hidden_size = 512
  118.  
  119.   modelFile = home_path + 'save/model/' + seg_corpus_name + '/1-1_512/' + str(iteration) + '_backup_bidir_model.tar'
  120.  
  121.   en, de, voc = predictLoad(seg_corpus_name, modelFile, n_layers, hidden_size)
  122.  
  123.   file_test(voc,en,de,top,test_file_name)
  124.   
  125. def print_voc(voc):
  126.   print('tw+jp voc size=%d' % (len(voc.word2index)))
  127.   print(voc.index2word)
  128.  
  129. def list_models(seg_corpus_name=''):
  130.   if seg_corpus_name=='':
  131.     modelPath = home_path + 'save/model/'
  132.   else:
  133.     modelPath = home_path + 'save/model/' + seg_corpus_name + '/1-1_512'
  134.  
  135.   out_bytes = subprocess.check_output(['ls','-l',modelPath],
  136.                                     stderr=subprocess.STDOUT)
  137.   out_text = out_bytes.decode('utf-8')
  138.   return out_text
  139.  
  140. def load_source(seg_corpus_name):
  141.   path = home_path + 'data/' + seg_corpus_name + '.txt'
  142.   
  143.   with open(path) as inp:
  144.     data = inp.readlines()
  145.  
  146.   print(len(data), len(data[0::2]), len(data[1::2]))
  147.  
  148.   data = { 'source': data[0::2], 'target': data[1::2] }
  149.   return data
  150.  
  151. # --------------------------
  152. app = Flask(__name__)
  153.  
  154. param0 = { 'model': 'translation2019_train_83k',
  155.           'epoch':  6000,
  156.           'topn' : 10,
  157.           'query' : 'what time is it?',
  158.           'result' : 'result area'
  159.         }
  160.  
  161. @app.route('/')
  162. def forms():
  163.   return render_template('translate.html', param=param0)
  164.  
  165. @app.route('/translate/<model>/<int:epoch>/<int:topn>', methods=['GET', 'POST'])
  166. def translate_long(model,epoch,topn):
  167.   if request.method == 'POST':
  168.     query = request.values['query']
  169.   elif request.method == 'GET':
  170.     query = request.args.get('query')
  171.  
  172.   return translate(model,epoch,topn,query)
  173.  
  174.  
  175. @app.route('/translate', methods=['GET', 'POST'])
  176. def translate_short():
  177.   if request.method == 'POST':
  178.     query = request.values['query']
  179.     model = request.values['model']
  180.     epoch = request.values['epoch']
  181.     topn = request.values['topn']
  182.   elif request.method == 'GET':
  183.     query = request.args.get('query')
  184.     model = request.args.get('model')
  185.     epoch = request.args.get('epoch')
  186.     topn = request.args.get('topn')
  187.  
  188.   return translate(model,epoch,topn,query)
  189.  
  190. def translate(model,epoch,topn,query):
  191.   epoch = int(epoch)
  192.   topn = int(topn)
  193.  
  194.   try:
  195.     target = sentence_test_model(model,epoch,topn,query)
  196.   except:
  197.     target = 'internal error, retry a again'
  198.  
  199.   result = 'query="{}"\nresult="{}"\n'.format(query,target)
  200.   
  201.   param2 = { 'model': model,
  202.           'epoch':  epoch,
  203.           'topn' : topn,
  204.           'query' : query,
  205.           'result' : result
  206.         }
  207.   return render_template('translate.html', param=param2)
  208.  
  209. @app.route('/list/<model>')
  210. def list_model(model):
  211.   mlist = list_models(model)
  212.   return '<pre>{}</pre>'.format(mlist)
  213.  
  214. @app.route('/list/')
  215. def list():
  216.   mlist = list_models()
  217.   return '<pre>{}</pre>'.format(mlist)
  218.  
  219. #######################################
  220.  
  221. USE_CUDA = torch.cuda.is_available()
  222. device = torch.device("cuda" if USE_CUDA else "cpu")
  223. home_path = './'
  224.  
  225. if __name__ == '__main__':
  226.   app.run(host='0.0.0.0',port=8080)


註: 本程式使用 GitHub JavaScript code prettifier 工具標示顏色。其方法如下:
   1.參考 [Blogger] 如何在 Blogger 顯示程式碼 - Google Code Prettify
     於【Blogger 版面配置 HTML/JavaScript小工具】安裝如下套件
       <script src="https://cdn.jsdelivr.net/gh/google/code-prettify@master/loader/run_prettify.js"></script>
   2.文章編輯再以HTML模式為程式包上如下標籤。
       <code class="prettyprint lang-html linenums"> ... </code>
       <code class="prettyprint lang-python linenums"> ... </code>
at
Labels: , ,

沒有留言:

張貼留言

訂閱： 張貼留言 (Atom)