我們在【Day17:TOD版的DialoGPT-ConvLab實作】介紹如何建構一個TOD對話系統，那今天我們來續講如何利用ConvLab-3來架設一個互動網站，看起來就像這樣：

互動網站包含：對話視窗(右半)、各個TOD模組的輸出(左半)

如何運行

• ConvLab-3已利用Flask建立一網站應用(程式碼可至convlab/deploy瀏覽)，
  將ConvLab-3 clone下來並安裝後，再輸入以下指令即可運行:

  cd ./ConvLab-3/convlab
  python ./deploy/run.py
  

• 在convlab/deploy/run.py尾部可看到Flask運行應用的app.run()，
  將host='127.0.0.1'改為host='0.0.0.0'，即可用global IP分享網站啦：

  if __name__ == '__main__':
      # gunicorn deploy.run:app --threads 4
      # Note that if you want to run this on a server, use host '0.0.0.0'
      app.run(host='127.0.0.1', port=dep_conf['net']['port'], debug=True)
  

• 你也可以使用gunicorn進行網站部署(WSGI Http Server):

  gunicorn -b 0.0.0.0:8888 deploy.run:app --threads 4
  

相關gunicorn參數設定可以瀏覽官方documentation

網路設定 (net)

在ConvLab-3只提供埠號的網路設定，需在dep_config.json進行修改，以下是可修改項目及敘述：

可以設定成這樣：

"net":
{
  "port": 8787,
  "app_name": "convlab",
  "session_time_out": 300
}

模組設定 (nlu, dst, policy, nlg)

• 我們可以調用ConvLab-3定義&訓練好的模型，也是在dep_config.json進行修改，以下是可修改項目及敘述：

而在ConvLab-3的dep_config.json只提供MultiWOZ的範例，
在此也提供如何使用CrossWOZ的dep_config.json：

{
  "net": {
    "port": 8787,
    "app_name": "convlab",
    "session_time_out": 300
  },
  "nlu": {
    "t5nlu-mul": {
      "class_path": "convlab.base_models.t5.nlu.T5NLU",
      "data_set": "multiwoz",
      "ini_params": {
        "speaker": "user",
        "context_window_size": 0,
        "model_name_or_path": "ConvLab/t5-small-nlu-multiwoz21"
      },
      "model_name": "t5nlu-mul",
      "max_core": 1,
      "preload": true,
      "enable": true
    },
    "t5nlu-cro": {
      "class_path": "convlab.base_models.t5.nlu.T5NLU",
      "data_set": "crosswoz",
      "ini_params": {
        "speaker": "user",
        "context_window_size": 0,
        "model_name_or_path": "ConvLab/mt5-small-nlu-all-crosswoz"
      },
      "model_name": "t5nlu-cro",
      "max_core": 1,
      "preload": true,
      "enable": true
    }
  },
  "dst": {
    "t5dst-mul": {
      "class_path": "convlab.base_models.t5.dst.T5DST",
      "data_set": "multiwoz",
      "ini_params": {
        "dataset_name": "multiwoz21",
        "speaker": "user",
        "context_window_size": 100,
        "model_name_or_path": "ConvLab/t5-small-dst-multiwoz21"
      },
      "model_name": "t5dst-mul",
      "max_core": 1,
      "preload": true,
      "enable": true
    },
    "t5dst-cro": {
      "class_path": "convlab.base_models.t5.dst.T5DST",
      "data_set": "crosswoz",
      "ini_params": {
        "dataset_name": "crosswoz",
        "speaker": "user",
        "context_window_size": 100,
        "model_name_or_path": "ConvLab/mt5-small-dst-crosswoz"
      },
      "model_name": "t5dst-cro",
      "max_core": 1,
      "preload": true,
      "enable": true
    }
  },
  "policy": {
    "ddpt-mul": {
      "class_path": "convlab.policy.vtrace_DPT.VTRACE",
      "data_set": "multiwoz",
      "ini_params": {
        "is_train": false,
        "seed": 0,
        "load_path": "from_pretrained", 
        "dataset_name": "multiwoz21"
      },
      "model_name": "ddpt-mul",
      "max_core": 1,
      "preload": true,
      "enable": true
    },
    "rule-cro": {
      "class_path": "convlab.policy.rule.crosswoz.Simulator",
      "data_set": "crosswoz",
      "ini_params": {},
      "model_name": "rule-cro",
      "max_core": 1,
      "preload": true,
      "enable": true
    }
  },
  "nlg": {
    "t5nlg-mul": {
      "class_path": "convlab.base_models.t5.nlg.T5NLG",
      "data_set": "multiwoz",
      "ini_params": {
        "speaker": "system",
        "context_window_size": 0,
        "model_name_or_path": "ConvLab/t5-small-nlg-multiwoz21"
      },
      "model_name": "t5nlg-mul",
      "max_core": 1,
      "preload": true,
      "enable": true
    },
    "t5nlg-cro": {
      "class_path": "convlab.base_models.t5.nlg.T5NLG",
      "data_set": "crosswoz",
      "ini_params": {
        "speaker": "system",
        "context_window_size": 0,
        "model_name_or_path": "ConvLab/mt5-small-nlg-all-crosswoz"
      },
      "model_name": "t5nlg-cro",
      "max_core": 1,
      "preload": true,
      "enable": true
    }
  }
}

如何使用

在convlab/deploy/run.py中我們可以看到以下程式碼：

@app.route("/dialog")
def dialog():
    return render_template("dialog.html")

這三行代表將HTML檔dialog.html放至dialog的路由(路徑)，我們以下面的URL進入架好的網站:

http://127.0.0.1:8787/dialog
修改Broadcast IP和Port的URL像這樣 - http://0.0.0.0:[port]/dialog

Reference.
ConvLab-3 Visualization Service Deployment