app.py

import transformers
import gradio as gr
# import warnings
import torch
# warnings.simplefilter('ignore')

device = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = transformers.GPT2Tokenizer.from_pretrained('gpt2')
#add padding token, beginstring and endstring tokens
tokenizer.add_special_tokens(
{
    "pad_token":"<pad>",
    "bos_token":"<startstring>",
    "eos_token":"<endstring>"
})
#add bot token since it is not a special token
tokenizer.add_tokens(["<bot>:"])
print("=====Done 1")
model = transformers.GPT2LMHeadModel.from_pretrained('gpt2')
model.resize_token_embeddings(len(tokenizer))
model.load_state_dict(torch.load('./gpt2talk.pt', map_location=torch.device('cpu')))
print("=====Done 2")
model.eval()
def inference(quiz):
    quiz1 = quiz
    quiz = "<startstring>"+quiz+" <bot>:"
    
    quiztoken = tokenizer(quiz,
                          return_tensors='pt'
                         )
    
    answer = model.generate(**quiztoken, max_length=200, top_k=0.7,top_p=0.1)[0]
    answer = tokenizer.decode(answer, skip_special_tokens=True)
    answer = answer.replace(" <bot>:","").replace(quiz1,"") + '.'
    return answer

def chatbot(input_text):
    response = inference(input_text)
    return response

# Create the Gradio interface
print("=====Done 3")
gr.Interface(
    fn=chatbot,
    inputs='text',
    outputs='text',
    live=False, #set false to avoid caching
    # interpretation="chat",
    title="ChatFinance",
    description="Ask the a question and see its response!",
).launch()
# print("=====Done 4")
# # Launch the Gradio interface
# iface.launch()