A conversational agent with many personalities (PersonaGPT)

PersonaGPT is an open-domain conversational agent designed to do 2 tasks:

It builds on the DialoGPT-medium pretrained model based on the GPT-2 architecture. This model is trained on the Persona-Chat dataset, with added special tokens to better distinguish between conversational history and personality traits for dyadic conversations. Furthermore, some active learning was used to train the model to do controlled decoding using turn-level goals.

Full Repo

Preprocessing, training and implementation details can be found in the personaGPT repo .

How to Use

from transformers import GPT2Tokenizer, GPT2LMHeadModel
import torch
tokenizer = AutoTokenizer.from_pretrained("af1tang/personaGPT")
model = AutoModelForCausalLM.from_pretrained("af1tang/personaGPT")
if torch.cuda.is_available():
    model = model.cuda()
## utility functions ##
flatten = lambda l: [item for sublist in l for item in sublist]

def to_data(x):
    if torch.cuda.is_available():
        x = x.cpu()
    return x.data.numpy()

def to_var(x):
    if not torch.is_tensor(x):
        x = torch.Tensor(x)
    if torch.cuda.is_available():
        x = x.cuda()
    return x

def display_dialog_history(dialog_hx):
    for j, line in enumerate(dialog_hx):
        msg = tokenizer.decode(line)
        if j %2 == 0:
            print(">> User: "+ msg)
        else:
            print("Bot: "+msg)
            print()

def generate_next(bot_input_ids, do_sample=True, top_k=10, top_p=.92,
                  max_length=1000, pad_token=tokenizer.eos_token_id):
    full_msg = model.generate(bot_input_ids, do_sample=True,
                                              top_k=top_k, top_p=top_p, 
                                              max_length=max_length, pad_token_id=tokenizer.eos_token_id)
    msg = to_data(full_msg.detach()[0])[bot_input_ids.shape[-1]:]
    return msg

# get personality facts for conversation
personas = []
for i in range(3):
    response = input(">> Fact %d: "%(i+1))+ tokenizer.eos_token
    personas.append(response)
personas = tokenizer.encode(''.join(['<|p2|>'] + personas + ['<|sep|>'] + ['<|start|>']))

# converse for 8 turns
dialog_hx = []
for step in range(8):
    # encode the user input
    user_inp = tokenizer.encode(input(">> User: ") + tokenizer.eos_token)
    # append to the chat history
    dialog_hx.append(user_inp)
        
    # generated a response while limiting the total chat history to 1000 tokens, 
    bot_input_ids = to_var([personas + flatten(dialog_hx)]).long()
    msg = generate_next(bot_input_ids)
    dialog_hx.append(msg)
    print("Bot: {}".format(tokenizer.decode(msg, skip_special_tokens=True)))

Example of personalized decoding:

## available actions ##
action_space = [ 'ask about kids.', "ask about pets.", 'talk about work.', 
               'ask about marital status.', 'talk about travel.', 'ask about age and gender.',
        'ask about hobbies.', 'ask about favorite food.', 'talk about movies.', 
        'talk about music.', 'talk about politics.']
# converse for 8 turns
dialog_hx = []
for step in range(8):
    # choose an action
    act = None
    while act not in action_space:
        display_dialog_history(dialog_hx)
        print()
        print(" actions: ")
        for k,v in enumerate(action_space): print(k,v)
        try:
            act = action_space[int(input(" input [0-10]: " ))]
        except:
            act = None
    print()
    # format into prefix code
    action_prefix = tokenizer.encode(''.join(['<|act|> '] + [act] + ['<|p1|>'] + [] + ['<|sep|>'] + ['<|start|>']))
    bot_input_ids = to_var([action_prefix + flatten(dialog_hx)]).long()
    
    # generate query conditioned on action
    msg = generate_next(bot_input_ids, top_k=top_k, top_p=top_p, max_length=max_length)
    dialog_hx.append(msg)
    
    # generate bot response
    bot_input_ids = to_var([personas+ flatten(dialog_hx)]).long()
    msg = generate_next(bot_input_ids, top_k=top_k, top_p=top_p, max_length=max_length)
    dialog_hx.append(msg)
display_dialog_history(dialog_hx)

Example of controlled response generation: