Parlai: Starting a chat_service with Blender 90M model

[DasShubhadeep]

Hi, I would like to integrate blender with my application and send text to it remotely.
I would like to start a WebSocket based server with Blender 90M version but I am not able to understand what changes are required in this config file to load blender 90M model with the server.

https://github.com/facebookresearch/ParlAI/tree/master/parlai/chat_service/tasks/chatbot

Any help?

[GraphGrailAi]

I was able to implement Telegram bot with some hacking, ie capturing stdout from terminal, but this is not recommended behaviour

[DasShubhadeep]

@GraphGrailAi
I see. But how did you integrate or send command to blender model using the bot? Can you share the steps?

[GraphGrailAi]

https://stackoverflow.com/questions/28616018/multiple-inputs-and-outputs-in-python-subprocess-communicate

and more here https://github.com/pexpect/pexpect

[GraphGrailAi]

@stephenroller could you suggest?
I tried these changes:

tasks:
  default:
    onboard_world: MessengerBotChatOnboardWorld
    task_world: MessengerBotChatTaskWorld
    timeout: 1800
    agents_required: 1
task_name: chatbot
world_module: parlai.chat_service.tasks.chatbot.worlds
overworld: MessengerOverworld
max_workers: 30
opt:
  debug: True
  models:
    blender:
      model: local_human
      model_file: models:blender/blender_90M/model 
      override:
        model: local_human
        no_cuda: True
      no_cuda: True
additional_args:
  page_id: 580729918709559 # Configure Your Own Page

It loads, but no actual communication:

python parlai/chat_service/services/browser_chat/run.py --config-path parlai/chat_service/tasks/chatbot/config_blender.yml --port 10001
[ optional arguments: ] 
[  datapath: /home/joo/ParlAI/data ]
[ Chat Services: ] 
[  config_path: parlai/chat_service/tasks/chatbot/config_blender.yml ]
[  is_debug: False ]
[  password: None ]
[ Browser Chat: ] 
[  port: 10001 ]
[ Current ParlAI commit: 855b8bf1abae78d60de6a3bfe37eb4b654bf3f8d ]
[ warning: overriding opt['model'] to local_human (previously: transformer/generator )]
[ warning: overriding opt['no_cuda'] to True (previously: False )]
Enter [DONE] if you want to end the episode, [EXIT] to quit.
[I 200519 18:21:45 web:2246] 101 GET /websocket (127.0.0.1) 0.60ms
[I 200519 18:21:45 sockets:39] Opened new socket from ip: 127.0.0.1
[I 200519 18:21:45 sockets:40] Current subscribers: {'fe32184c-ed96-4994-857f-b5124a0e5568': <parlai.chat_service.services.websocket.sockets.MessageSocketHandler object at 0x7fae7dd4ccc0>}
[I 200519 18:22:00 sockets:59] websocket message from client: {"text": "hi"}
[I 200519 18:22:00 agents:40] Sending new message: {'id': 'Overworld', 'text': 'Welcome to the overworld for the ParlAI messenger chatbot demo. Please type "begin" to start.', 'quick_replies': ['begin']}
[I 200519 18:22:11 sockets:59] websocket message from client: {"text": "begin"}
[I 200519 18:22:11 agents:55] Received new message: {'text': 'begin', 'payload': None, 'sender': {'id': 'fe32184c-ed96-4994-857f-b5124a0e5568'}, 'recipient': {'id': 0}}
2020-05-19 18:22:11: Adding agent fe32184c-ed96-4994-857f-b5124a0e5568 to pool...
onboarding/overworld complete
starting pool
2020-05-19 18:22:12: Removing agent fe32184c-ed96-4994-857f-b5124a0e5568 from pool...
Starting task t_1...
[E 200519 18:22:12 logging:44] World default had error KeyError('legacy_seq2seq',)
World default had error KeyError('legacy_seq2seq',)
NoneType: None
Next task: None
[I 200519 18:22:12 agents:40] Sending new message: {'id': 'Overworld', 'text': 'Welcome to the overworld for the ParlAI messenger chatbot demo. Please type "begin" to start.', 'quick_replies': ['begin']}

[stephenroller]

I don’t think you want to specify local human along with the model file, those are very different settings

[stephenroller]

@klshuster

[GraphGrailAi]

I changed model: transformer/generator - now it seems model loaded
but another error:

Starting task t_1...
[E 200519 20:15:04 logging:44] World default had error KeyError('legacy_seq2seq',)
World default had error KeyError('legacy_seq2seq',)
Traceback (most recent call last):
  File "/root/anaconda3/lib/python3.6/concurrent/futures/thread.py", line 56, in run
    result = self.fn(*self.args, **self.kwargs)
  File "/var/www/ParlAI/parlai/chat_service/core/world_runner.py", line 128, in _world_fn
    return self._run_world(task, world_name, agents)
  File "/var/www/ParlAI/parlai/chat_service/core/world_runner.py", line 95, in _run_world
    world = world_generator(self.opt, agents)
  File "/var/www/ParlAI/parlai/chat_service/tasks/chatbot/worlds.py", line 50, in generate_world
    opt['shared_bot_params'][MessengerBotChatTaskWorld.MODEL_KEY]
KeyError: 'legacy_seq2seq'
Next task: None
[I 200519 20:15:04 agents:40] Sending new message: {'id': 'Overworld', 'text': 'Welcome to the overworld for the ParlAI messenger chatbot demo. Please type "begin" to start.', 'quick_replies': ['begin']}

[stephenroller]

https://github.com/facebookresearch/ParlAI/blob/6c2e66a9f696571fd336ea58365841fd32d15319/parlai/chat_service/tasks/chatbot/worlds.py#L34

Seems to be coming from here, but I don't quite understand why it's set there. Will defer to Kurt.

[GraphGrailAi]

I fixed this line with MODEL_KEY = 'blender' and now i see actual response from model in terminal, but not in browser because of minor error in code ParlAI/parlai/chat_service/tasks/chatbot/worlds.py",
line 82, in parley
response['id'] = ''

Full traceback:
`` [I 200519 22:33:15 agents:55] Received new message: {'text': 'hi', 'payload': None, 'sender': {'id': 'a39b4190-b951-4676-b2f3-a71ff6cee5b5'}, 'recipient': {'id': 0}} ===act==== {'episode_done': False, 'text': 'hi', 'payload': None} ~~~~~~~~~~~ [E 200519 22:33:16 base_events:1285] Task was destroyed but it is pending! task: <Task pending coro=<WebSocketProtocol13.write_message.<locals>.wrapper() running at /root/anaconda3/lib/python3.6/site-packages/tornado/websocket.py:1102>> /root/anaconda3/lib/python3.6/asyncio/base_events.py:509: RuntimeWarning: coroutine 'WebSocketProtocol13.write_message.<locals>.wrapper' was never awaited self._ready.clear() ===response==== {'id': 'TransformerGenerator', 'episode_done': False, 'text': 'hi , how are you today ? i just got back from a long day of work , how about you ?'} ~~~~~~~~~~~ [E 200519 22:33:19 logging:44] World default had error RuntimeError('Message already contains keyid. If this was intentional, please use the functionforce_set(key, value).',) World default had error RuntimeError('Message already contains keyid. If this was intentional, please use the functionforce_set(key, value).',) Traceback (most recent call last): File "/root/anaconda3/lib/python3.6/concurrent/futures/thread.py", line 56, in run result = self.fn(*self.args, **self.kwargs) File "/var/www/ParlAI/parlai/chat_service/core/world_runner.py", line 128, in _world_fn return self._run_world(task, world_name, agents) File "/var/www/ParlAI/parlai/chat_service/core/world_runner.py", line 99, in _run_world ret_val = world.parley() File "/var/www/ParlAI/parlai/chat_service/tasks/chatbot/worlds.py", line 82, in parley response['id'] = '' File "/var/www/ParlAI/parlai/core/message.py", line 26, in __setitem__ 'please use the functionforce_set(key, value).'.format(key) RuntimeError: Message already contains keyid. If this was intentional, please use the functionforce_set(key, value)`.
Next task: None

[GraphGrailAi]

commented line 82 and it works.
seems we need more review: how it affects logging, user sessions, etc behavior.

What i faced is that session is not persistent: using another browser to chat(i.e. another user) - the same dialog history used for response

[klshuster]

MODEL_KEY = 'legacy_seq2seq' is defined there as this a demo set of worlds; indeed this is what you'd change to be the key in the config file under models (as you did with 'blender')

line 82 is a bug, thanks for flagging

the session is indeed not persistent in the current websockets implementation, we have not had a chance to get that into ParlAI yet

[GraphGrailAi]

@klshuster
That's why i asked about Telegram implementation in another issue. Do you have a plan of session implementation?

If not i would ask 3 simple things to get Telegram ready:
1) I need access to global function that get user_message as input
2) I need access to global function that pass user_message to model and get response
3) I need access to global function that reinput dialog history to model to serve context.
That is enough for me, but inspecting code is curious/

What i found is that all 3 variable are in file https://github.com/facebookresearch/ParlAI/blob/6c2e66a9f696571fd336ea58365841fd32d15319/parlai/chat_service/tasks/chatbot/worlds.py#L69

ie

        a = self.agent.act()
        if a is not None:
            if '[DONE]' in a['text']:
                self.episodeDone = True
            else:
                print("===act====")
                print(a)
                print("~~~~~~~~~~~")
                self.model.observe(a)
                response = self.model.act()
                print("===response====")
                print(response)
                print("~~~~~~~~~~~")
                response['id'] = ''
                self.agent.observe(response)

So
1) self.agent.act()['text'] - user message
2) self.model.act()['text']- bot message (response from model)
3) self.agent.observe(response) - reinput of history (not sure of type - dict or what)

What do you suggest how do i use these variables in my code? Can i simply import this file https://github.com/facebookresearch/ParlAI/blob/master/parlai/chat_service/tasks/chatbot/worlds.py to my code and invoke this vars for example
MessengerBotChatTaskWorld.parley(agent.act()['text']) ?

[klshuster]

The way these chat service worlds work is that parley is called repeatedly until self.episodeDone is True. If you'd like, you could import that world and subclass it to modify the parley function.

Calling a = self.agent.act() will return the user's message in the form of an action/observation dict, or a Message, from which you can access the user message via the text field.

calling self.model.observe(a) will have the model view the user's message, and then subsequently calling response = self.model.act() will return the model's response.

An important aspect of the Message is the episode_done field; if you set this to False, the model will continue to track history of the conversation. Setting it to be True will clear the model's history after the model responds. So, you can either have the model repeatedly observe new human actions with episode_done: False, or you can manually track the history and repeatedly observe episode_done: True with the text field being the full history

[GraphGrailAi]

@klshuster yes, got that, good.
What is opt structure ?
tried invoking

    obj = MessengerBotChatTaskWorld(opt={'models':'blender'}, agent='ChatbotAgent', bot='blender')
    user_message = obj.parley()

got:
'str' object has no attribute 'observe'

[klshuster]

The opt structure is as specified in the config here

If you have not already, I would advise that you read through the documentation for running a chat service world here. The manager for the chat service takes care of agent creation here and task worlds are instantiated and launched via the world runner

[GraphGrailAi]

Of course, i inspected config already (as you can see i tried to pass dict obj = MessengerBotChatTaskWorld(opt={'models':'blender'})
But still, i got 'str' object has no attribute 'observe' - my guess that it's because str passed to agent='ChatbotAgent' so, what is agent should be.

And yes, i have read the docs https://parl.ai/docs/tutorial_chat_service.html - but that seem a complex multi-step task and even these details does not make it clear

I see you operated these concepts as a core developer, but for me it's hard. What i ask is simple step-by-step import statements and how do i pass parameters from code to be able to invoke the main: a = self.agent.act() self.model.observe(a) self.model.act() or parley()

Now i cannot invoke parley().
Should i also somehow call a https://github.com/facebookresearch/ParlAI/blob/6c2e66a9f696571fd336ea58365841fd32d15319/parlai/chat_service/tasks/chatbot/worlds.py#L43 and if yes what is agents ?

What i do expect is that:
1) "Simply import this worlds.py to your code"
2) "Than call these 2 functions with parameters: ... . That will load model and prepare world."
3) "Now you call parley() this way someobject.parley() and inside you can access .model.observe(a) like this: ..."
That's it

[klshuster]

A simple barebones approach is the following:

1. Construct an opt in a config.yml format (similar to the demo ones), and fill in the options as is done in this main function
2. Create a human agent like so
3. Create a model like so
4. Import a MessengerChatBotTaskWorld into your code and subclass it
5. Call generate_world with your setup opt, and agents where agents = [human_agent] created in step 2. Generate world will create a model from the shared model parameters generated in step 3.
6. Override the parley function to do as you please; call world.parley() to produce responses.

Let me know if these instructions are helpful

[GraphGrailAi]

@stephenroller @klshuster
What i try to achieve is manual session storage when user interact with model.
Now i can use Pexpect (terminal emulator) and run multiple processes with
/root/anaconda3/bin/python3.6 safe_interactive.py -t blended_skill_talk -mf zoo:blender/blender_90M/model
Every terminal has separate user session - what we actually need. But it is memory consuming - every python process takes 1300-2000MB of RAM on server, so several sessions will crash server. How do avoid this, sharing the same model for any user?

[stephenroller]

Read the worlds tutorial to understand how agent sharing works. https://parl.ai/docs/tutorial_worlds.html

[GraphGrailAi]

Read the worlds tutorial to understand how agent sharing works. https://parl.ai/docs/tutorial_worlds.html

done, but this is theoretical material that needs practice in code...

[GraphGrailAi]

Solved sessions myself.
That was a tricky coding: i just implemented a['episode_done'] = True for every message and do a manual history tracking with collecting it in text field.

The main thing i did not figured out at the beginning was that:

An important aspect of the Message is the episode_done field; if you set this to False, the model will continue to track history of the conversation. Setting it to be True will clear the model's history after the model responds. So, you can either have the model repeatedly observe new human actions with episode_done: False, or you can manually track the history and repeatedly observe episode_done: True with the text field being the full history

[MOODesign]

@GraphGrailAi I'm facing the same problem, how did you manage to solve the sessions issue ?
I'm running the 90M blender on a websocket and interacting with it using the localhost,
I've already created the config.yml file as you described and edited the worlds.py file to make it work with 'blender'.
But i'm still getting the same context from different sessions.
so, how did you use the self.episodeDone = True & self.model.observe(a) & response = self.model.act() to make the bot clear history after every iteration of parley() and retrieve the history before responding to the human agent ???

[bakszero]

@GraphGrailAi Got the same confusion as above. Would be great if you could elaborate.

Ref. https://github.com/facebookresearch/ParlAI/issues/2681

[rabidgremlin]

@klshuster

An important aspect of the Message is the episode_done field; if you set this to False, the model will continue to track history of the conversation. Setting it to be True will clear the model's history after the model responds. So, you can either have the model repeatedly observe new human actions with episode_done: False, or you can manually track the history and repeatedly observe episode_done: True with the text field being the full history

What format would the text field be ? The previous human inputs concatenated with \n ? Or something else ?

[stephenroller]

@klshuster

An important aspect of the Message is the episode_done field; if you set this to False, the model will continue to track history of the conversation. Setting it to be True will clear the model's history after the model responds. So, you can either have the model repeatedly observe new human actions with episode_done: False, or you can manually track the history and repeatedly observe episode_done: True with the text field being the full history

What format would the text field be ? The previous human inputs concatenated with \n ? Or something else ?

Human and bot inputs concatenated with \n

[viprocket1]

A simple barebones approach is the following:

1. Construct an opt in a config.yml format (similar to the demo ones), and fill in the options as is done in this main function
2. Create a human agent like so
3. Create a model like so
4. Import a MessengerChatBotTaskWorld into your code and subclass it
5. Call generate_world with your setup opt, and agents where agents = [human_agent] created in step 2. Generate world will create a model from the shared model parameters generated in step 3.
6. Override the parley function to do as you please; call world.parley() to produce responses.

Let me know if these instructions are helpful

can you elaborate this ? i am also trying to implement this in Telegram as well . I tried following these steps but its hard to understand.
How should i create human agent ? if like this . What is 'task_id' and other parameters

1. Create a human agent like so

Thank you for your time .

[josharnoldjosh]

I can explain what worked for me, but I think there is an easier way:

• start with the interactive.py script
• On line 88 (or close to this line), you should see human_agent = LocalHumanAgent(opt)
• You basically want to subclass LocalHumanAgent and initialize your agent there. Below is an example of mine:

class EvalAgent(LocalHumanAgent):

    def observe(self, msg):
        self.supplied_context.append(msg['text'])
        print(
            display_messages(
                [msg],
                ignore_fields=self.opt.get('display_ignore_fields', ''),
                prettify=self.opt.get('display_prettify', False),
            )        
        )

    def act(self):        
        reply = Message()
        reply['id'] = self.getID()        
        reply_text = "Hello world"
        print(f"[EvalAgent] {reply_text}")
        reply_text = reply_text.replace('\\n', '\n')
        reply['episode_done'] = False     
        # reply.force_set('episode_done', True)  # enabling this ends an episode if you want to reset chat history
        return reply

Then on line 88 replace human_agent = LocalHumanAgent(opt) with human_agent = EvalAgent(opt)

• To start the model, run Interactive.main(model_file='zoo:blender/blender_90M/model', t="convai2")
• The t="convai2" will enable personas, simply run Interactive.main(model_file='zoo:blender/blender_90M/model') for no persona
• To end the session, in your while loop with parley(), just do human_agent.finished = True
• Also. agent.self_observe({'text':utt}) will allow you to manually set the history of the blender bot

Hope this helps. Let me know if you have any questions that I might be able to answer too!

[klshuster]

@viprocket1 For implementing in Telegram, i'd recommend checking out the chat service tutorial and the chat service readme.

Regarding a human agent - task_id is just a unique identifier for the "task" where your human is. The name is a bit of an artifact; think of it more as e.g. "world_id" or "session_id" (a "task" is basically a thread running an interaction between a model and a human)

[viprocket1]

thank you @klshuster and @josharnoldjosh for responding . I will try both the approaches and get back to you .

[github-actions[bot]]

This issue has not had activity in 30 days. Please feel free to reopen if you have more issues. You may apply the "never-stale" tag to prevent this from happening.

[chen-yifu]

Sorry, has anyone been able to figure out how to allow multiple agents and separate history/context? Would really appreciate it if could see the code. Thanks in advance.

[klshuster]

the canonical example that works in ParlAI is to use the Facebook Messenger chat service implementation. I would recommend reading through that code to understand how it works