import gradio as gr from langchain.chat_models import ChatOpenAI from langchain.chains import ConversationChain from transformers import pipeline model_name="nateraw/bert-base-uncased-emotion" model = pipeline('text-classification', model_name, truncation=True) from transformers import AutoTokenizer, AutoModelWithLMHead model_name = "mrm8488/t5-base-finetuned-emotion" tokenizer = AutoTokenizer.from_pretrained(model_name) model_t5 = AutoModelWithLMHead.from_pretrained(model_name) def get_emotion(text): input_ids = tokenizer.encode(text + '', return_tensors='pt') output = model_t5.generate(input_ids=input_ids, return_dict_in_generate=True, output_scores=True) transition_scores = model_t5.compute_transition_scores(output.sequences, [s.softmax(dim=1) for s in output.scores], normalize_logits=False) dec = [tokenizer.decode(ids) for ids in output.sequences] score = transition_scores.min().item() return f"{dec[0].replace('','').replace('','').strip()} [{score}]" chat = ChatOpenAI() conversation = ConversationChain(llm=chat) #Write a text example of someone angry with gr.Blocks() as demo: label_text = gr.Textbox(label="Sentiment Text", text="") chatbot = gr.Chatbot(scale=2) msg = gr.Textbox() clear = gr.ClearButton([msg, chatbot]) def respond(message, chat_history): bot_message = conversation.run(message) chat_history.append((message, bot_message)) l = model(bot_message)[0] label_value = f"{l['label']} [{l['score']}]" label_value_t5 = get_emotion(bot_message) return "", chat_history, f"Model1: {label_value_t5} - Model2: {label_value}" msg.submit(respond, [msg, chatbot], [msg, chatbot, label_text]) demo.launch()