I tried running the large model (in a colab notebook) using the approach described in the model card from the huggingface library:
from transformers import AutoModelWithLMHead, AutoTokenizer
import torch
tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-large")
model = AutoModelWithLMHead.from_pretrained("microsoft/DialoGPT-large")
# Let's chat for 5 lines
for step in range(5):
# encode the new user input, add the eos_token and return a tensor in Pytorch
new_user_input_ids = tokenizer.encode(input(">> User:") + tokenizer.eos_token, return_tensors='pt')
# append the new user input tokens to the chat history
bot_input_ids = torch.cat([chat_history_ids, new_user_input_ids], dim=-1) if step > 0 else new_user_input_ids
# generated a response while limiting the total chat history to 1000 tokens,
chat_history_ids = model.generate(bot_input_ids, max_length=1000, pad_token_id=tokenizer.eos_token_id)
# pretty print last ouput tokens from bot
print("DialoGPT: {}".format(tokenizer.decode(chat_history_ids[:, bot_input_ids.shape[-1]:][0], skip_special_tokens=True)))Unfortunately, the output seems to quickly repeat the same sentence over and over again. In example 1 and 2, it repeats the same sentence from the beginning. In the 3rd case, the model starts fine, but as the conversation progresses, it starts repeating the same thing (or parts of it).
Is that an intended behavior?
Example 1
>> User:Today, I bought some ice cream.
DialoGPT: I'm so sorry.
>> User:Why are you sorry?
DialoGPT: I'm sorry for your loss.
>> User:What loss? I didn't lose anything.
DialoGPT: I'm sorry for your loss.
>> User:I'm very confused by what you are saying.
DialoGPT: I'm sorry for your loss.
>> User:Ok, nevermind.
DialoGPT: I'm sorry for your loss.
Example 2
>> User:Can't wait to upvote this 17 different times later this week.
DialoGPT: I'm going to upvote it every time.
>> User:What are you going to upvote every time?
DialoGPT: I'm going to upvote it every time.
>> User:Uh sure, if that's what you want.
DialoGPT: I'm going to upvote it every time.
>> User:I'm very confused by what you are saying.
DialoGPT: I'm going to upvote it every time.
>> User:What's your name?
DialoGPT: I'm going to upvote it every time.
Example 3
>> User:I like to walk my dog in the park.
DialoGPT: I like to walk my dog in the park.
>> User:That's great. What's your dog breed?
DialoGPT: He's a mix of some sort.
>> User:That's funny, you don't even know what mix it is.
DialoGPT: I know what mix it is.
>> User:What is it, then?
DialoGPT: A mix of some sort.
>> User:What sort? "Sort" is not a type of dog.
DialoGPT: I know what mix it is.
>> User:That's not the issue. I'm talking about your dog's breed.
DialoGPT: I know what mix it is.
>> User:Then tell me!
DialoGPT: I know what mix it is.
>> User:What are you waiting for?
DialoGPT: I know what mix?
>> User:Are you serious?
DialoGPT: I know what?
>> User:What?
DialoGPT: I know what?
I tried running the large model (in a colab notebook) using the approach described in the model card from the huggingface library:
Unfortunately, the output seems to quickly repeat the same sentence over and over again. In example 1 and 2, it repeats the same sentence from the beginning. In the 3rd case, the model starts fine, but as the conversation progresses, it starts repeating the same thing (or parts of it).
Is that an intended behavior?
Example 1
Example 2
Example 3