codeclm/trainer/codec_song_pl.py

"""
Main model for using CodecLM. This will combine all the required components
and provide easy access to the generation API.
"""

import typing as tp
import warnings
import sys
import time
import torch
import torch.nn as nn
from torch.nn import functional as F
import torchaudio
import numpy as np
import lightning as pl
from torchmetrics.classification import MulticlassAccuracy
import pdb
from codeclm.models import builders
import math
from torch.optim import Optimizer
from torch.optim.lr_scheduler import _LRScheduler
from peft import LoraConfig, get_peft_model
from datetime import datetime
import os 
os.environ['TOKENIZERS_PARALLELISM'] = "false"


class CodecLM_PL(pl.LightningModule):
    def __init__(self, cfg):
        super().__init__()

        self.cfg = cfg
        
        # 1) Build audio tokenizer (usually None during training)
        self.audio_tokenizer = builders.get_audio_tokenizer_model(self.cfg.audio_tokenizer_checkpoint, self.cfg)
        if self.audio_tokenizer is not None:
            for param in self.audio_tokenizer.parameters():
                param.requires_grad = False
        if "audio_tokenizer_checkpoint_sep" in self.cfg.keys():
            self.seperate_tokenizer = builders.get_audio_tokenizer_model(self.cfg.audio_tokenizer_checkpoint_sep, self.cfg)
            for param in self.seperate_tokenizer.parameters():
                param.requires_grad = False
        else:
            self.seperate_tokenizer = None
        
        # 2) Build LM
        self.audiolm = builders.get_lm_model(self.cfg)
        print(self.audiolm)
        # 输出参数量
        print('Number of parameters: ', sum(p.numel() for p in self.audiolm.parameters()))
        # 3) Load pretrained checkpoint (if any)
        if self.cfg.use_pretrained == 'deepspeed':
            checkpoint = torch.load(self.cfg.pretrained.deepspeed_checkpoint, map_location='cpu')
            missing, unexpected = self.load_state_dict(checkpoint, strict=False)
            print(f'-------------Missing--------------\n{missing}')
            print(f'-------------Unexpected--------------\n{unexpected}')
            print("successfully load deepspeed pretrained model {}".format(self.cfg.pretrained.deepspeed_checkpoint))
            self.missing = missing
        else:
            self.missing = []
        # 如果cfg参数中有lora
        if hasattr(self.cfg, 'lora'):
            perf_config = LoraConfig(
                r = self.cfg.lora.r,
                lora_alpha = self.cfg.lora.lora_alpha,
                target_modules = self.cfg.lora.target_modules,
                lora_dropout = self.cfg.lora.lora_dropout,
                bias = self.cfg.lora.bias,
                task_type = self.cfg.lora.task_type,
            )
            self.audiolm = get_peft_model(self.audiolm, perf_config)
        
        # 4) Build metrics
        self.val_steps = []
        self.train_slide_acc = []
        self.train_steps = []
        self.top1_acc_metric = nn.ModuleList([MulticlassAccuracy(
            self.audiolm.code_size, 
            top_k=1,
            average="micro", multidim_average="global",
            ignore_index=self.cfg.lm.code_size, # ignore EOS token prediction
        ) for _ in range(self.audiolm.code_depth)])
        self.top10_acc_metric = nn.ModuleList([MulticlassAccuracy(
            self.audiolm.code_size,
            top_k=10,
            average="micro", multidim_average="global",
            ignore_index=self.cfg.lm.code_size,
        ) for _ in range(self.audiolm.code_depth)])

        self.epoch = 0
        print("++++++++++++++++ training <song> +++++++++++++++++")

    # TODO: move this part to loader
    def generate_mask_and_end_token(self, x, sequence_lengths, end_id=16384):
        batch_size = sequence_lengths.size(0)
        max_length = x.size(2)

        # pad one frame, if the maximum sequence length is equal to the input length
        if max_length == sequence_lengths.max():
            x = F.pad(x, (0, 1), value=end_id)
        max_length = x.size(2)

        if max_length <= sequence_lengths.max() + 1:
            sequence_lengths = sequence_lengths - (sequence_lengths.max()+1 - max_length)

        # Add end token to x according to the sequence length
        x[torch.arange(batch_size), :, sequence_lengths] = end_id
        sequence_lengths += 1

        mask = torch.arange(max_length).expand(batch_size, max_length) < sequence_lengths.unsqueeze(1)
        mask = mask.to(x.device)
        mask_3d = mask.unsqueeze(1).expand(batch_size, x.size(1), max_length)
        x = torch.where(mask_3d, x, end_id+1)
        return x, mask_3d

    @torch.no_grad()
    def preprocess_batch(self, batch):  # this function is usually called during training
        # 处理 dataloader 返回的数据
        audio, text_lyric, time_stamp, structure_dur, prompt_audio, structure_labels = batch

        dur, valid_st, valid_et = zip(*time_stamp)
        
        if self.audio_tokenizer is not None:
            # only used in inference
            self.audio_tokenizer.eval()
            with torch.no_grad():
                with torch.cuda.amp.autocast(enabled=False):
                    audio_tokens, scale = self.audio_tokenizer.encode(audio)
                audio_tokens = audio_tokens[:,:self.cfg.lm.code_depth,:]
                audio_tokens = audio_tokens.long()
        else:
            audio_tokens = audio.long()
        
        token_dur = (torch.Tensor(dur) * self.cfg.audio_tokenizer_frame_rate).int()
        audio_tokens, audio_padding_mask = self.generate_mask_and_end_token(audio_tokens, token_dur, 
                                                                            end_id=self.audiolm.eos_token_id)
        condition_tensors = self.audiolm.prepare_condition_tensors(batch_size=len(text_lyric),
                                                                   text=text_lyric, audio_qt_emb=prompt_audio)

        return condition_tensors, audio_tokens, audio_padding_mask

    def get_time(self):
        # 获取当前的日期和时间
        now = datetime.now()

        # 使用strftime函数格式化日期和时间
        formatted_now = now.strftime("%Y-%m-%d %H:%M:%S.%f")
        return formatted_now

    def training_step(self, batch, batch_idx):
        # 1) data processing
        condition_tensors, audio_tokens, padding_mask = self.preprocess_batch(batch)
        
        # 2) compute model predictions (model forward)
        model_output = self.audiolm.compute_predictions(audio_tokens, condition_tensors, 
                                                        training_steps=self.global_step)  # this input can be ignored        
        logits = model_output.logits.float()
        mask = padding_mask & model_output.mask
        
        # 3) compute loss (float)
        with torch.cuda.amp.autocast(enabled=False):
            ce, ce_per_codebook = self._compute_cross_entropy(logits, audio_tokens, mask)
        
        total_loss = ce
        if torch.isnan(total_loss):
            print(self.trainer.global_rank, ce, padding_mask, batch[1])
            print('--------------------------------------------------------------')
            return None
            # torchaudio.save("error_rank{}.wav".format(self.trainer.global_rank), batch[0][:,0].cpu(), 24000)
            # import pdb; pdb.set_trace()
        # 4) compute metrics and log
        metrics = {}
        self.log('ce', ce, prog_bar=True) 
        metrics['ppl'] = torch.exp(ce)
        for k, ce_q in enumerate(ce_per_codebook):
            metrics[f'ce_q{k + 1}'] = ce_q
            metrics[f'ppl_q{k + 1}'] = torch.exp(ce_q)

        masked_labels = audio_tokens.masked_fill(~mask, value=self.cfg.lm.code_size)
        metrics['acc'] = []
        for k in range(self.audiolm.code_depth):
            metrics['acc'].append(self.top1_acc_metric[k](logits[:, k].transpose(1,2).detach(), 
                                                          masked_labels[:, k]).item())
        metrics['acc'] = torch.mean(torch.Tensor(metrics['acc'])).item()

        self.train_steps.append({'ce': ce.detach().cpu().item(), 'acc': metrics['acc']})        
        self.log('train_acc', metrics['acc']+1e-8, prog_bar=True)
        self.log('lr', self.trainer.optimizers[0].param_groups[0]['lr'], prog_bar=True)  
        self.log_dict(metrics)

        return total_loss
    
    @torch.no_grad()
    def validation_step(self, batch, batch_idx):
        # 1) data processing
        condition_tensors, audio_tokens, padding_mask = self.preprocess_batch(batch)

        # 2) compute model predictions
        model_output = self.audiolm.compute_predictions(audio_tokens, condition_tensors)  
        logits = model_output.logits
        mask = padding_mask & model_output.mask
        
        # 3) compute loss and metrics
        ce, ce_per_codebook = self._compute_cross_entropy(logits, audio_tokens, mask)
        metrics = {}   
        metrics['val_ce'] = ce
        metrics['val_ppl'] = torch.exp(ce)
        for k, ce_q in enumerate(ce_per_codebook):
            metrics[f'val_ce_q{k + 1}'] = ce_q
            metrics[f'val_ppl_q{k + 1}'] = torch.exp(ce_q)
        masked_labels = audio_tokens.masked_fill(~mask, value=self.cfg.lm.code_size)

        for k in range(self.audiolm.code_depth):
            self.top1_acc_metric[k].update(logits[:, k].transpose(1,2).detach(), masked_labels[:,k]) #* total_length
            self.top10_acc_metric[k].update(logits[:, k].transpose(1,2).detach(), masked_labels[:,k])
        self.val_steps.append(metrics)
            
        metrics['acc'] = []
        metrics['acc_top10'] = []
        for k in range(self.audiolm.code_depth):
            metrics['acc'].append(self.top1_acc_metric[k](logits[:, k].transpose(1,2).detach(), masked_labels[:,k]).item())
            metrics['acc_top10'].append(self.top10_acc_metric[k](logits[:, k].transpose(1,2).detach(), masked_labels[:,k]).item())
        metrics['acc'] = torch.mean(torch.Tensor(metrics['acc']))
        metrics['acc_top10'] = torch.mean(torch.Tensor(metrics['acc_top10'])) 
        
        return metrics['acc']


    def on_validation_epoch_end(self) -> None:        
        final_metrics = {}
        for i in self.val_steps:
            for k in i:
                final_metrics[k] = final_metrics.get(k, []) + [i[k]]
        final_metrics = {k: sum(v) / len(v) for k,v in list(final_metrics.items())}
        self.log_dict(final_metrics)

        q_acc = []
        q_acc10 = []
        for i in range(self.audiolm.code_depth):
            q_acc.append(self.top1_acc_metric[i].compute())
            q_acc10.append(self.top10_acc_metric[i].compute())
            self.log(f"val_Top1Acc_{i}", q_acc[-1])
            self.log(f"val_Top10Acc_{i}", q_acc10[-1])
            self.top1_acc_metric[i].reset()
            self.top10_acc_metric[i].reset()
        
        self.log('val_Top1Acc', sum(q_acc) / self.audiolm.code_depth)
        self.log('val_Top10Acc', sum(q_acc10) / self.audiolm.code_depth)

        return super().on_validation_epoch_end()


    def on_validation_epoch_start(self) -> None:
        self.val_steps = []
        for i in range(self.audiolm.code_depth):
            self.top1_acc_metric[i].reset()
            self.top10_acc_metric[i].reset()

        if len(self.train_steps) > 0:
            train_metrics = {}
            for i in self.train_steps:
                for k in i:
                    train_metrics[k] = train_metrics.get(k, []) + [i[k]]
            train_metrics = {k: sum(v) / len(v) for k,v in list(train_metrics.items())}
            self.log('train_summary_Top1Acc', train_metrics['acc'])
            self.log('train_summary_ce', train_metrics['ce'])
            self.train_steps = []

        return super().on_validation_epoch_start()


    # 定义优化器
    def configure_optimizers(self):
        total_updates = self.cfg.optim.epochs * self.cfg.optim.updates_per_epoch
        optim_dict = {}

        param_groups = []
        missing_params = []
        other_params = []
        cnt = 0
        # 去掉开头的‘audiolm.'
        print('before missing len', len(self.missing))
        self.missing = [name.replace('audiolm.', '') for name in self.missing]
        print('after missing len', len(self.missing))
        for name, param in self.audiolm.named_parameters():
            if name in self.missing:
                cnt += 1
                print(name)
                missing_params.append(param)
            else:
                other_params.append(param)
        print(cnt)
        assert cnt == len(self.missing)
        param_groups.append({'params': other_params, 'lr': self.cfg.optim.old_lr})
        param_groups.append({
            'params': missing_params,
            'lr': self.cfg.optim.new_lr  # 为missing参数设置10倍的学习率,你可以调整这个倍数
        })

        if self.cfg.optim.optimizer == "adamw":
            optim_dict['optimizer'] = torch.optim.AdamW(
                param_groups,  # 使用参数分组替代原来的 self.audiolm.parameters()
                betas=tuple(self.cfg.optim.adam.betas),
                weight_decay=self.cfg.optim.adam.weight_decay,
                eps=self.cfg.optim.adam.eps,
            )
        else:
            raise NotImplementedError

        if self.cfg.schedule is None:
            pass
        elif self.cfg.schedule.lr_scheduler == "cosine":
            scheduler = CosineLRScheduler(optim_dict['optimizer'], 
                                          total_steps=total_updates, 
                                          warmup_steps=self.cfg.schedule.cosine.warmup,
                                          lr_min_ratio=self.cfg.schedule.cosine.lr_min_ratio,
                                          cycle_length=self.cfg.schedule.cosine.cycle_length,
                                          )
            optim_dict['lr_scheduler'] = {"scheduler": scheduler, "interval": "step"}
        else:
            raise NotImplementedError
        
        return optim_dict

    
    def _compute_cross_entropy(
        self, logits: torch.Tensor, targets: torch.Tensor, mask: torch.Tensor
    ) -> tp.Tuple[torch.Tensor, tp.List[torch.Tensor]]:
        """Compute cross entropy between multi-codebook targets and model's logits.
        The cross entropy is computed per codebook to provide codebook-level cross entropy.
        Valid timesteps for each of the codebook are pulled from the mask, where invalid
        timesteps are set to 0.

        Args:
            logits (torch.Tensor): Model's logits of shape [B, K, T, card].
            targets (torch.Tensor): Target codes, of shape [B, K, T].
            mask (torch.Tensor): Mask for valid target codes, of shape [B, K, T].
        Returns:
            ce (torch.Tensor): Cross entropy averaged over the codebooks
            ce_per_codebook (list of torch.Tensor): Cross entropy per codebook (detached).
        """
        # import pdb; pdb.set_trace()
        B, K, T = targets.shape
        assert logits.shape[:-1] == targets.shape
        assert mask.shape == targets.shape
        ce = torch.zeros([], device=targets.device)
        ce_per_codebook: tp.List[torch.Tensor] = []
        for k in range(K):
            logits_k = logits[:, k, ...].contiguous().view(-1, logits.size(-1))  # [B x T, card]
            targets_k = targets[:, k, ...].contiguous().view(-1)  # [B x T]
            mask_k = mask[:, k, ...].contiguous().view(-1)  # [B x T]
            ce_targets = targets_k[mask_k]
            ce_logits = logits_k[mask_k]
            q_ce = F.cross_entropy(ce_logits, ce_targets)
            ce += q_ce
            ce_per_codebook.append(q_ce.detach())
        # average cross entropy across codebooks
        ce = ce / K
        return ce, ce_per_codebook


class CodecLM_PL_FT(pl.LightningModule):
    def __init__(self, cfg):
        super().__init__()

        self.cfg = cfg
        
        # 1) Build audio tokenizer (usually None during training)
        self.audio_tokenizer = builders.get_audio_tokenizer_model(self.cfg)
        if self.audio_tokenizer is not None:
            for param in self.audio_tokenizer.parameters():
                param.requires_grad = False
        
        # 2) Build LM
        self.audiolm = builders.get_lm_model(self.cfg)
        
        # 3) Load pretrained checkpoint (if any)
        if self.cfg.use_pretrained == 'deepspeed':
            checkpoint = torch.load(self.cfg.pretrained.deepspeed_checkpoint,  map_location='cpu')
            missing, unexpected = self.load_state_dict(checkpoint, strict=False)
            print(f'-------------Missing--------------\n{missing}')
            print(f'-------------Unexpected--------------\n{unexpected}')
            print("successfully load deepspeed pretrained model {}".format(self.cfg.pretrained.deepspeed_checkpoint))

        # 4) Build metrics
        self.val_steps = []
        self.train_slide_acc = []
        self.train_steps = []
        self.top1_acc_metric = nn.ModuleList([MulticlassAccuracy(
            self.audiolm.code_size, 
            top_k=1,
            average="micro", multidim_average="global",
            ignore_index=self.cfg.lm.code_size, # ignore EOS token prediction
        ) for _ in range(self.audiolm.code_depth)])
        self.top10_acc_metric = nn.ModuleList([MulticlassAccuracy(
            self.audiolm.code_size,
            top_k=10,
            average="micro", multidim_average="global",
            ignore_index=self.cfg.lm.code_size,
        ) for _ in range(self.audiolm.code_depth)])

        self.epoch = 0
        print("++++++++++++++++ training <song> +++++++++++++++++")

    # TODO: move this part to loader
    def generate_mask_and_end_token(self, x, sequence_lengths, end_id=16384):
        batch_size = sequence_lengths.size(0)
        max_length = x.size(2)

        # pad one frame, if the maximum sequence length is equal to the input length
        if max_length == sequence_lengths.max():
            x = F.pad(x, (0, 1), value=end_id)
        max_length = x.size(2)

        if max_length <= sequence_lengths.max() + 1:
            sequence_lengths = sequence_lengths - (sequence_lengths.max()+1 - max_length)

        # Add end token to x according to the sequence length
        x[torch.arange(batch_size), :, sequence_lengths] = end_id
        sequence_lengths += 1

        mask = torch.arange(max_length).expand(batch_size, max_length) < sequence_lengths.unsqueeze(1)
        mask = mask.to(x.device)
        mask_3d = mask.unsqueeze(1).expand(batch_size, x.size(1), max_length)
        x = torch.where(mask_3d, x, end_id+1)
        return x, mask_3d

    @torch.no_grad()
    def preprocess_batch(self, batch):  # this function is usually called during training
        # 处理 dataloader 返回的数据
        audio, text_lyric, time_stamp, lang_type, prompt_audio = batch        
        dur, valid_st, valid_et = zip(*time_stamp)
        
        if self.audio_tokenizer is not None:
            # only used in inference
            self.audio_tokenizer.eval()
            with torch.no_grad():
                with torch.cuda.amp.autocast(enabled=False):
                    audio_tokens, scale = self.audio_tokenizer.encode(audio)
                audio_tokens = audio_tokens[:,:self.cfg.lm.code_depth,:]
                audio_tokens = audio_tokens.long()
        else:
            audio_tokens = audio.long()
        
        token_dur = (torch.Tensor(dur) * self.cfg.audio_tokenizer_frame_rate).int()

        audio_tokens, audio_padding_mask = self.generate_mask_and_end_token(audio_tokens, token_dur, 
                                                                            end_id=self.audiolm.eos_token_id)
        condition_tensors = self.audiolm.prepare_condition_tensors(batch_size=len(text_lyric),
                                                                   text=text_lyric, audio_qt_emb=prompt_audio)

        return condition_tensors, audio_tokens, audio_padding_mask

    def get_time(self):
        # 获取当前的日期和时间
        now = datetime.now()

        # 使用strftime函数格式化日期和时间
        formatted_now = now.strftime("%Y-%m-%d %H:%M:%S.%f")
        return formatted_now

    def training_step(self, batch, batch_idx):
        # 1) data processing
        condition_tensors, audio_tokens, padding_mask = self.preprocess_batch(batch)
        
        # 2) compute model predictions (model forward)
        model_output = self.audiolm.compute_predictions(audio_tokens, condition_tensors, 
                                                        training_steps=self.global_step)  # this input can be ignored        
        logits = model_output.logits.float()
        mask = padding_mask & model_output.mask
        
        # 3) compute loss (float)
        with torch.cuda.amp.autocast(enabled=False):
            ce, ce_per_codebook = self._compute_cross_entropy(logits, audio_tokens, mask)
        
        total_loss = ce
        if torch.isnan(total_loss):
            print(self.trainer.global_rank, ce, padding_mask, batch[1])
            # print('------------------------------------------------------------------------')
            torchaudio.save("error_rank{}.wav".format(self.trainer.global_rank), batch[0][:,0].cpu(), 24000)
            import pdb; pdb.set_trace()
            return None

        # 4) compute metrics and log
        metrics = {}
        self.log('ce', ce, prog_bar=True) 
        metrics['ppl'] = torch.exp(ce)
        for k, ce_q in enumerate(ce_per_codebook):
            metrics[f'ce_q{k + 1}'] = ce_q
            metrics[f'ppl_q{k + 1}'] = torch.exp(ce_q)

        masked_labels = audio_tokens.masked_fill(~mask, value=self.cfg.lm.code_size)
        metrics['acc'] = []
        for k in range(self.audiolm.code_depth):
            metrics['acc'].append(self.top1_acc_metric[k](logits[:, k].transpose(1,2).detach(), 
                                                          masked_labels[:, k]).item())
        metrics['acc'] = torch.mean(torch.Tensor(metrics['acc'])).item()

        self.train_steps.append({'ce': ce.detach().cpu().item(), 'acc': metrics['acc']})        
        self.log('train_acc', metrics['acc']+1e-8, prog_bar=True)
        self.log('lr', self.trainer.optimizers[0].param_groups[0]['lr'], prog_bar=True)  
        self.log_dict(metrics)

        return total_loss
    
    @torch.no_grad()
    def validation_step(self, batch, batch_idx):
        # 1) data processing
        condition_tensors, audio_tokens, padding_mask = self.preprocess_batch(batch)

        # 2) compute model predictions
        model_output = self.audiolm.compute_predictions(audio_tokens, condition_tensors)  
        logits = model_output.logits
        mask = padding_mask & model_output.mask
        
        # 3) compute loss and metrics
        ce, ce_per_codebook = self._compute_cross_entropy(logits, audio_tokens, mask)
        metrics = {}   
        metrics['val_ce'] = ce
        metrics['val_ppl'] = torch.exp(ce)
        for k, ce_q in enumerate(ce_per_codebook):
            metrics[f'val_ce_q{k + 1}'] = ce_q
            metrics[f'val_ppl_q{k + 1}'] = torch.exp(ce_q)
        masked_labels = audio_tokens.masked_fill(~mask, value=self.cfg.lm.code_size)

        for k in range(self.audiolm.code_depth):
            self.top1_acc_metric[k].update(logits[:, k].transpose(1,2).detach(), masked_labels[:,k]) #* total_length
            self.top10_acc_metric[k].update(logits[:, k].transpose(1,2).detach(), masked_labels[:,k])
        self.val_steps.append(metrics)
        metrics['acc'] = []
        metrics['acc_top10'] = []
        for k in range(self.audiolm.code_depth):
            metrics['acc'].append(self.top1_acc_metric[k](logits[:, k].transpose(1,2).detach(), masked_labels[:,k]).item())
            metrics['acc_top10'].append(self.top10_acc_metric[k](logits[:, k].transpose(1,2).detach(), masked_labels[:,k]).item())
        metrics['acc'] = torch.mean(torch.Tensor(metrics['acc']))
        metrics['acc_top10'] = torch.mean(torch.Tensor(metrics['acc_top10']))
        
        return metrics['acc']

    def on_validation_epoch_end(self) -> None:        
        final_metrics = {}
        for i in self.val_steps:
            for k in i:
                final_metrics[k] = final_metrics.get(k, []) + [i[k]]
        final_metrics = {k: sum(v) / len(v) for k,v in list(final_metrics.items())}
        self.log_dict(final_metrics)

        q_acc = []
        q_acc10 = []
        for i in range(self.audiolm.code_depth):
            q_acc.append(self.top1_acc_metric[i].compute())
            q_acc10.append(self.top10_acc_metric[i].compute())
            self.log(f"val_Top1Acc_{i}", q_acc[-1])
            self.log(f"val_Top10Acc_{i}", q_acc10[-1])
            self.top1_acc_metric[i].reset()
            self.top10_acc_metric[i].reset()
        
        self.log('val_Top1Acc', sum(q_acc) / self.audiolm.code_depth)
        self.log('val_Top10Acc', sum(q_acc10) / self.audiolm.code_depth)

        return super().on_validation_epoch_end()


    def on_validation_epoch_start(self) -> None:
        self.val_steps = []
        for i in range(self.audiolm.code_depth):
            self.top1_acc_metric[i].reset()
            self.top10_acc_metric[i].reset()

        if len(self.train_steps) > 0:
            train_metrics = {}
            for i in self.train_steps:
                for k in i:
                    train_metrics[k] = train_metrics.get(k, []) + [i[k]]
            train_metrics = {k: sum(v) / len(v) for k,v in list(train_metrics.items())}
            self.log('train_summary_Top1Acc', train_metrics['acc'])
            self.log('train_summary_ce', train_metrics['ce'])
            self.train_steps = []

        return super().on_validation_epoch_start()


    # 定义优化器
    def configure_optimizers(self):
        total_updates = self.cfg.optim.epochs * self.cfg.optim.updates_per_epoch
        optim_dict = {}

        if self.cfg.optim.optimizer == "adamw":
            optim_dict['optimizer'] = torch.optim.AdamW(
                self.audiolm.parameters(),
                lr=self.cfg.optim.lr,
                betas=tuple(self.cfg.optim.adam.betas),
                weight_decay=self.cfg.optim.adam.weight_decay,
                eps=self.cfg.optim.adam.eps,
            )
        else:
            raise NotImplementedError

        if self.cfg.schedule is None:
            pass
        elif self.cfg.schedule.lr_scheduler == "cosine":
            scheduler = CosineLRScheduler(optim_dict['optimizer'], 
                                          total_steps=total_updates, 
                                          warmup_steps=self.cfg.schedule.cosine.warmup,
                                          lr_min_ratio=self.cfg.schedule.cosine.lr_min_ratio,
                                          cycle_length=self.cfg.schedule.cosine.cycle_length,
                                          )
            optim_dict['lr_scheduler'] = {"scheduler": scheduler, "interval": "step"}
        else:
            raise NotImplementedError
        
        return optim_dict

    
    def _compute_cross_entropy(
        self, logits: torch.Tensor, targets: torch.Tensor, mask: torch.Tensor
    ) -> tp.Tuple[torch.Tensor, tp.List[torch.Tensor]]:
        """Compute cross entropy between multi-codebook targets and model's logits.
        The cross entropy is computed per codebook to provide codebook-level cross entropy.
        Valid timesteps for each of the codebook are pulled from the mask, where invalid
        timesteps are set to 0.

        Args:
            logits (torch.Tensor): Model's logits of shape [B, K, T, card].
            targets (torch.Tensor): Target codes, of shape [B, K, T].
            mask (torch.Tensor): Mask for valid target codes, of shape [B, K, T].
        Returns:
            ce (torch.Tensor): Cross entropy averaged over the codebooks
            ce_per_codebook (list of torch.Tensor): Cross entropy per codebook (detached).
        """
        # import pdb; pdb.set_trace()
        B, K, T = targets.shape
        assert logits.shape[:-1] == targets.shape
        assert mask.shape == targets.shape
        ce = torch.zeros([], device=targets.device)
        ce_per_codebook: tp.List[torch.Tensor] = []
        for k in range(K):
            logits_k = logits[:, k, ...].contiguous().view(-1, logits.size(-1))  # [B x T, card]
            targets_k = targets[:, k, ...].contiguous().view(-1)  # [B x T]
            mask_k = mask[:, k, ...].contiguous().view(-1)  # [B x T]
            ce_targets = targets_k[mask_k]
            ce_logits = logits_k[mask_k]
            q_ce = F.cross_entropy(ce_logits, ce_targets)
            ce += q_ce
            ce_per_codebook.append(q_ce.detach())
        # average cross entropy across codebooks
        ce = ce / K
        return ce, ce_per_codebook

class CosineLRScheduler(_LRScheduler):# 
    """Cosine LR scheduler.

    Args:
        optimizer (Optimizer): Torch optimizer.
        warmup_steps (int): Number of warmup steps.
        total_steps (int): Total number of steps.
        lr_min_ratio (float): Minimum learning rate.
        cycle_length (float): Cycle length.
    """
    def __init__(self, optimizer: Optimizer, total_steps: int, warmup_steps: int,
                 lr_min_ratio: float = 0.0, cycle_length: float = 1.0):
        self.warmup_steps = warmup_steps
        assert self.warmup_steps >= 0
        self.total_steps = total_steps
        assert self.total_steps >= 0
        self.lr_min_ratio = lr_min_ratio
        self.cycle_length = cycle_length
        super().__init__(optimizer)

    def _get_sched_lr(self, lr: float, step: int):
        if step < self.warmup_steps:
            lr_ratio = step / self.warmup_steps
            lr = lr_ratio * lr
        elif step <= self.total_steps:
            s = (step - self.warmup_steps) / (self.total_steps - self.warmup_steps)
            lr_ratio = self.lr_min_ratio + 0.5 * (1 - self.lr_min_ratio) * \
                (1. + math.cos(math.pi * s / self.cycle_length))
            lr = lr_ratio * lr
        else:
            lr_ratio = self.lr_min_ratio
            lr = lr_ratio * lr
        return lr

    def get_lr(self):
        return [self._get_sched_lr(lr, self.last_epoch) for lr in self.base_lrs]