from transformers import AutoTokenizer, AutoModelForMaskedLM import torch BERTTokenizer = AutoTokenizer.from_pretrained("cl-tohoku/bert-base-japanese") BERTModel = AutoModelForMaskedLM.from_pretrained("cl-tohoku/bert-base-japanese") from transformers import AutoModelForSeq2SeqLM mT5Tokenizer = AutoTokenizer.from_pretrained("google/mt5-base") mT5Model = AutoModelForSeq2SeqLM.from_pretrained("google/mt5-base") from transformers import AutoModelForCausalLM GPT2Tokenizer = AutoTokenizer.from_pretrained("rinna/japanese-gpt2-medium") GPT2Model = AutoModelForCausalLM.from_pretrained("rinna/japanese-gpt2-medium") import gradio as gr votes=[] BERT=None mT5=None GPT2=None def MELCHIOR(sue): #BERT allow=BERTTokenizer("承認").input_ids[1] deny=BERTTokenizer("否決").input_ids[1] output=BERTModel(**BERTTokenizer('MELCHIORは科学者としての人格を持っています。人間とMELCHIORの対話です。人間「'+sue+'。承認 か 否決 か?」'+"MELCHIOR 「[MASK]」",return_tensors="pt")).logits BERTTokenizer.batch_decode(torch.argmax(output,-1)) mask=output[0,-3,:] votes.append(1 if mask[allow]>mask[deny] else -1) return "承認" if mask[allow]>mask[deny] else "否定" def BALTHASAR(sue): #mT5 allow=mT5Tokenizer("承認").input_ids[1] deny=mT5Tokenizer("否決").input_ids[1] encoder_output=mT5Model.encoder(**mT5Tokenizer('BALTHASARは母としての人格としての人格を持っています。人間とBALTHASARの対話です。人間「'+sue+'。承認 か 否決 か?」'+"BALTHASAR 「」",return_tensors="pt")) id=None p_answer=None probs=None i=0 txt="" probs=mT5Model(inputs_embeds=encoder_output.last_hidden_state,decoder_input_ids=mT5Tokenizer(txt,return_tensors="pt").input_ids).logits[0] id=torch.argmax(probs[i+1]) txt=txt+"" i=i+1 probs=mT5Model(inputs_embeds=encoder_output.last_hidden_state,decoder_input_ids=mT5Tokenizer(txt,return_tensors="pt").input_ids).logits[0] id=torch.argmax(probs[i+1]) txt=txt+mT5Tokenizer.decode(id) votes.append(1 if probs[i+1][allow]>probs[i+1][deny] else -1) return "承認" if probs[i+1][allow]>probs[i+1][deny] else "否定" def CASPER(sue): #GPT2 allow=GPT2Tokenizer("承認").input_ids[1] deny=GPT2Tokenizer("否決").input_ids[1] probs=GPT2Model(**GPT2Tokenizer('CASPERは女としての人格を持っています。人間とCASPERの対話です。人間「'+sue+'。承認 か 否決 か?」'+"CASPER 「",return_tensors="pt")).logits[0] i=-1 p_answer=probs id=torch.argmax(probs[i]) votes.append(1 if probs[i][allow]>probs[i][deny] else -1) return "承認" if probs[i][allow]>probs[i][deny] else "否決" def greet(sue): text1="BERT-1"+MELCHIOR(sue) text2="GPT-2"+CASPER(sue) text3="mT5-3"+BALTHASAR(sue) return text1+" "+text2+" "+text3+"\n___\n\n"+("|可決|" if sum(votes[-3:])>0 else "| 否決 |")+"\n___" css="@import url('https://fonts.googleapis.com/css2?family=Shippori+Mincho:wght@800&display=swap'); .gradio-container {background-color: black} .gr-button {background-color: blue;color:black; weight:200%;font-family:'Shippori Mincho', serif;}" css+=".block{color:orange;} ::placeholder {font-size:35%} .gr-box {text-align: center;font-size: 125%;border-color:orange;background-color: #000000;weight:200%;font-family:'Shippori Mincho', serif;}:disabled {color: orange;opacity:1.0;}" with gr.Blocks(css=css) as demo: sue = gr.Textbox(label="NAGI System",placeholder="決議内容を入力") greet_btn = gr.Button("提訴") output = gr.Textbox(label="決議",placeholder="本システムは事前学習モデルのpromptにより行われています.決議結果に対して当サービス開発者は一切の責任を負いません.") greet_btn.click(fn=greet, inputs=sue, outputs=output) demo.launch()