| from transformers import Pipeline | |
| class MyPipeline: | |
| def __init__(self,model,tokenizer): | |
| self.model=model | |
| self.tokenizer=tokenizer | |
| def chat_Format(self,context, quetion, answer): | |
| return "Instruction:/n check answer is true or false of next quetion using context below:\n" + "#context: " + context + f".\n#quetion: " + quetion + f".\n#student answer: " + answer + ".\n#response:" | |
| def __call__(self, context, quetion, answer,generate=1,max_new_tokens=4, num_beams=2, do_sample=False,num_return_sequences=1): | |
| inp=self.chat_Format(context, quetion, answer) | |
| w = self.tokenizer(inp, add_special_tokens=True, | |
| pad_to_max_length=True, | |
| return_attention_mask=True, | |
| return_tensors='pt') | |
| response="" | |
| if(generate): | |
| outputs = self.tokenizer.batch_decode(self.model.generate(input_ids=w['input_ids'].cuda(), attention_mask=w['attention_mask'].cuda(), max_new_tokens=max_new_tokens, num_beams=num_beams, do_sample=do_sample, num_return_sequences=num_return_sequences), skip_special_tokens=True) | |
| response = outputs | |
| s =self.model(input_ids=w['input_ids'].cuda(), attention_mask=w['attention_mask'].cuda())['logits'][0][-1] | |
| s = F.softmax(s, dim=-1) | |
| yes_token_id = self.tokenizer.convert_tokens_to_ids("صØŃÙĬØŃ") | |
| no_token_id = self.tokenizer.convert_tokens_to_ids("خط") | |
| print(yes_token_id,no_token_id) | |
| for i in ["Yes", "yes", "True", "true","صحيح"]: | |
| s[yes_token_id] += s[self.tokenizer.convert_tokens_to_ids(i)] | |
| for i in ["No", "no", "False", "false","خطأ"]: | |
| s[no_token_id] += s[self.tokenizer.convert_tokens_to_ids(i)] | |
| true = (s[yes_token_id] / (s[no_token_id] + s[yes_token_id])).item() | |
| return {"response": response, "true": true} | |