|
|
import torch |
|
|
import torch.nn as nn |
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, LogitsProcessor |
|
|
from configs.hyperparametric import Reward_config,Tree_config |
|
|
from utils.warp import Warp |
|
|
from utils.model_generate import extra_span_from_tokens |
|
|
|
|
|
|
|
|
config = Reward_config().to_dict() |
|
|
|
|
|
def reinforced_loss(W,V,I_a,I_s,p,lower=1e-3,upper=1-1e-3): |
|
|
prob = I_a / I_s * p |
|
|
|
|
|
W = torch.where(W > lower, W, lower) |
|
|
W = torch.where(W < upper, W, upper) |
|
|
|
|
|
return - torch.log(torch.mean(W * prob)) |
|
|
|
|
|
|
|
|
class PRGTrainer(Trainer): |
|
|
def __init__(self,tokenizer,logits_processor=None,**kwargs): |
|
|
super(PRGTrainer,self).__init__(**kwargs) |
|
|
self.loss_tokenizer = tokenizer |
|
|
self.logits_processor = logits_processor if logits_processor else LogitsProcessor |
|
|
self.ce = nn.CrossEntropyLoss(ignore_index=-100) |
|
|
|
|
|
self.open_tag = tokenizer.encode('<v>',add_special_tokens=False) |
|
|
self.close_tag = tokenizer.encode('</v>',add_special_tokens=False) |
|
|
|
|
|
self.state = 'ref' |
|
|
|
|
|
def compute_loss(self, model, inputs, return_outputs=False,num_items_in_batch=None): |
|
|
V,I_a,I_s,prob = inputs.pop('reward'),inputs.pop('I-all'),inputs.pop('I-sample'),inputs.pop('prob') |
|
|
input_ids,attention_mask = inputs.pop('input_ids'),inputs.pop('attention_mask') |
|
|
labels,label_mask = inputs.pop('labels'),inputs.pop('label_mask') |
|
|
|
|
|
outputs = model(input_ids=input_ids, |
|
|
attention_mask=attention_mask, |
|
|
return_dict=config['return_dict_in_generate'], |
|
|
) |
|
|
logits = outputs.logits |
|
|
logits = torch.softmax(logits,dim=-1) |
|
|
|
|
|
shift_logits = logits[..., :-1, :].contiguous() |
|
|
shift_labels = labels[..., 1:].contiguous() |
|
|
shift_label_mask = label_mask[..., 1:].contiguous() |
|
|
|
|
|
batch_label_prob = shift_logits * shift_label_mask.unsqueeze(-1) |
|
|
shift_label_mask = (batch_label_prob != 0).any(dim=-1).any(dim=0) |
|
|
shift_logits = shift_logits[:,shift_label_mask,:] |
|
|
shift_labels = shift_labels[:,shift_label_mask] |
|
|
|
|
|
W = torch.max(shift_logits,dim=-1)[0] |
|
|
W = torch.mean(W,dim=-1) |
|
|
|
|
|
loss_ce = self.ce(shift_logits.view(-1, shift_logits.size(-1)),shift_labels.view(-1)) |
|
|
loss_fb = reinforced_loss(W=W,V=V,I_a=I_a,I_s=I_s,p=prob) |
|
|
|
|
|
loss = loss_ce + loss_fb |
|
|
|
|
|
return (loss, outputs) if return_outputs else loss |
