import os
import sys
import logging

logger = logging.getLogger(__name__)

now_dir = os.getcwd()
sys.path.append(os.path.join(now_dir))

import datetime
from tqdm import tqdm  # Added import

from infer.lib.train import utils

hps = utils.get_hparams()
os.environ["CUDA_VISIBLE_DEVICES"] = hps.gpus.replace("-", ",")
n_gpus = len(hps.gpus.split("-"))
from random import randint, shuffle

import torch

try:
    import intel_extension_for_pytorch as ipex  # pylint: disable=import-error, unused-import

    if torch.xpu.is_available():
        from infer.modules.ipex import ipex_init
        from infer.modules.ipex.gradscaler import gradscaler_init
        from torch.xpu.amp import autocast

        GradScaler = gradscaler_init()
        ipex_init()
    else:
        from torch.cuda.amp import GradScaler, autocast
except Exception:
    from torch.cuda.amp import GradScaler, autocast

torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = False
from time import sleep
from time import time as ttime

import torch.distributed as dist
import torch.multiprocessing as mp
from torch.nn import functional as F
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

from infer.lib.infer_pack import commons
from infer.lib.train.data_utils import (
    DistributedBucketSampler,
    TextAudioCollate,
    TextAudioCollateMultiNSFsid,
    TextAudioLoader,
    TextAudioLoaderMultiNSFsid,
)

if hps.version == "v1":
    from infer.lib.infer_pack.models import MultiPeriodDiscriminator
    from infer.lib.infer_pack.models import SynthesizerTrnMs256NSFsid as RVC_Model_f0
    from infer.lib.infer_pack.models import (
        SynthesizerTrnMs256NSFsid_nono as RVC_Model_nof0,
    )
else:
    from infer.lib.infer_pack.models import (
        SynthesizerTrnMs768NSFsid as RVC_Model_f0,
        SynthesizerTrnMs768NSFsid_nono as RVC_Model_nof0,
        MultiPeriodDiscriminatorV2 as MultiPeriodDiscriminator,
    )

from infer.lib.train.losses import (
    discriminator_loss,
    feature_loss,
    generator_loss,
    kl_loss,
)
from infer.lib.train.mel_processing import mel_spectrogram_torch, spec_to_mel_torch
from infer.lib.train.process_ckpt import savee

global_step = 0


class EpochRecorder:
    def __init__(self):
        self.last_time = ttime()

    def record(self):
        now_time = ttime()
        elapsed_time = now_time - self.last_time
        self.last_time = now_time
        elapsed_time_str = str(datetime.timedelta(seconds=elapsed_time))
        current_time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
        return f"[{current_time}] | ({elapsed_time_str})"


def main():
    n_gpus = torch.cuda.device_count()

    if torch.cuda.is_available() == False and torch.backends.mps.is_available() == True:
        n_gpus = 1
    if n_gpus < 1:
        # patch to unblock people without gpus. there is probably a better way.
        print("NO GPU DETECTED: falling back to CPU - this may take a while")
        n_gpus = 1
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = str(randint(20000, 55555))
    children = []
    logger = utils.get_logger(hps.model_dir)
    logger.info(f"Starting training with {n_gpus} GPU(s)")
    for i in range(n_gpus):
        subproc = mp.Process(
            target=run,
            args=(i, n_gpus, hps, logger),
        )
        children.append(subproc)
        subproc.start()

    for i in range(n_gpus):
        children[i].join()


def run(rank, n_gpus, hps, logger: logging.Logger):
    global global_step
    if rank == 0:
        logger.info(f"Process {rank}/{n_gpus-1} started")
        logger.info(hps)
        writer = SummaryWriter(log_dir=hps.model_dir)
        writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))

    dist.init_process_group(
        backend="gloo", init_method="env://", world_size=n_gpus, rank=rank
    )
    torch.manual_seed(hps.train.seed)
    if torch.cuda.is_available():
        torch.cuda.set_device(rank)

    if hps.if_f0 == 1:
        train_dataset = TextAudioLoaderMultiNSFsid(hps.data.training_files, hps.data)
    else:
        train_dataset = TextAudioLoader(hps.data.training_files, hps.data)
    train_sampler = DistributedBucketSampler(
        train_dataset,
        hps.train.batch_size * n_gpus,
        [100, 200, 300, 400, 500, 600, 700, 800, 900],
        num_replicas=n_gpus,
        rank=rank,
        shuffle=True,
    )
    
    if hps.if_f0 == 1:
        collate_fn = TextAudioCollateMultiNSFsid()
    else:
        collate_fn = TextAudioCollate()
    
    train_loader = DataLoader(
        train_dataset,
        num_workers=4,
        shuffle=False,
        pin_memory=True,
        collate_fn=collate_fn,
        batch_sampler=train_sampler,
        persistent_workers=True,
        prefetch_factor=8,
    )
    
    if rank == 0:
        logger.info(f"Training dataset size: {len(train_dataset)}")
        logger.info(f"Number of batches per epoch: {len(train_loader)}")

    if hps.if_f0 == 1:
        net_g = RVC_Model_f0(
            hps.data.filter_length // 2 + 1,
            hps.train.segment_size // hps.data.hop_length,
            **hps.model,
            is_half=hps.train.fp16_run,
            sr=hps.sample_rate,
        )
    else:
        net_g = RVC_Model_nof0(
            hps.data.filter_length // 2 + 1,
            hps.train.segment_size // hps.data.hop_length,
            **hps.model,
            is_half=hps.train.fp16_run,
        )
    
    if torch.cuda.is_available():
        net_g = net_g.cuda(rank)
    
    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm)
    if torch.cuda.is_available():
        net_d = net_d.cuda(rank)
    
    optim_g = torch.optim.AdamW(
        net_g.parameters(),
        hps.train.learning_rate,
        betas=hps.train.betas,
        eps=hps.train.eps,
    )
    optim_d = torch.optim.AdamW(
        net_d.parameters(),
        hps.train.learning_rate,
        betas=hps.train.betas,
        eps=hps.train.eps,
    )
    
    if hasattr(torch, "xpu") and torch.xpu.is_available():
        pass
    elif torch.cuda.is_available():
        net_g = DDP(net_g, device_ids=[rank])
        net_d = DDP(net_d, device_ids=[rank])
    else:
        net_g = DDP(net_g)
        net_d = DDP(net_d)

    try:
        _, _, _, epoch_str = utils.load_checkpoint(
            utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d
        )
        if rank == 0:
            logger.info("Loaded discriminator checkpoint")
        
        _, _, _, epoch_str = utils.load_checkpoint(
            utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g
        )
        global_step = (epoch_str - 1) * len(train_loader)
        if rank == 0:
            logger.info(f"Resuming from epoch {epoch_str}, global step {global_step}")
    except:
        epoch_str = 1
        global_step = 0
        if hps.pretrainG != "":
            if rank == 0:
                logger.info(f"Loading pretrained generator from {hps.pretrainG}")
            if hasattr(net_g, "module"):
                net_g.module.load_state_dict(
                    torch.load(hps.pretrainG, map_location="cpu")["model"]
                )
            else:
                net_g.load_state_dict(
                    torch.load(hps.pretrainG, map_location="cpu")["model"]
                )
        if hps.pretrainD != "":
            if rank == 0:
                logger.info(f"Loading pretrained discriminator from {hps.pretrainD}")
            if hasattr(net_d, "module"):
                net_d.module.load_state_dict(
                    torch.load(hps.pretrainD, map_location="cpu")["model"]
                )
            else:
                net_d.load_state_dict(
                    torch.load(hps.pretrainD, map_location="cpu")["model"]
                )

    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
        optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
    )
    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(
        optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
    )

    scaler = GradScaler(enabled=hps.train.fp16_run)

    cache = []
    
    if rank == 0:
        logger.info(f"Starting training from epoch {epoch_str} to {hps.train.epochs}")
        logger.info(f"Total epochs to train: {hps.train.epochs - epoch_str + 1}")
    
    for epoch in range(epoch_str, hps.train.epochs + 1):
        if rank == 0:
            train_and_evaluate(
                rank,
                epoch,
                hps,
                [net_g, net_d],
                [optim_g, optim_d],
                [scheduler_g, scheduler_d],
                scaler,
                [train_loader, None],
                logger,
                [writer, writer_eval],
                cache,
            )
        else:
            train_and_evaluate(
                rank,
                epoch,
                hps,
                [net_g, net_d],
                [optim_g, optim_d],
                [scheduler_g, scheduler_d],
                scaler,
                [train_loader, None],
                None,
                None,
                cache,
            )
        scheduler_g.step()
        scheduler_d.step()


def train_and_evaluate(
    rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers, cache
):
    net_g, net_d = nets
    optim_g, optim_d = optims
    train_loader, eval_loader = loaders
    if writers is not None:
        writer, writer_eval = writers

    train_loader.batch_sampler.set_epoch(epoch)
    global global_step

    net_g.train()
    net_d.train()

    # Prepare data iterator
    if hps.if_cache_data_in_gpu == True:
        if cache == []:
            if rank == 0:
                logger.info("Caching data in GPU...")
                cache_progress = tqdm(total=len(train_loader), 
                                    desc="Caching", 
                                    position=0, 
                                    leave=True,
                                    disable=(rank != 0))
            
            for batch_idx, info in enumerate(train_loader):
                if hps.if_f0 == 1:
                    (
                        phone,
                        phone_lengths,
                        pitch,
                        pitchf,
                        spec,
                        spec_lengths,
                        wave,
                        wave_lengths,
                        sid,
                    ) = info
                else:
                    (
                        phone,
                        phone_lengths,
                        spec,
                        spec_lengths,
                        wave,
                        wave_lengths,
                        sid,
                    ) = info
                
                if torch.cuda.is_available():
                    phone = phone.cuda(rank, non_blocking=True)
                    phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
                    if hps.if_f0 == 1:
                        pitch = pitch.cuda(rank, non_blocking=True)
                        pitchf = pitchf.cuda(rank, non_blocking=True)
                    sid = sid.cuda(rank, non_blocking=True)
                    spec = spec.cuda(rank, non_blocking=True)
                    spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
                    wave = wave.cuda(rank, non_blocking=True)
                
                if hps.if_f0 == 1:
                    cache.append(
                        (
                            batch_idx,
                            (
                                phone,
                                phone_lengths,
                                pitch,
                                pitchf,
                                spec,
                                spec_lengths,
                                wave,
                                wave_lengths,
                                sid,
                            ),
                        )
                    )
                else:
                    cache.append(
                        (
                            batch_idx,
                            (
                                phone,
                                phone_lengths,
                                spec,
                                spec_lengths,
                                wave,
                                wave_lengths,
                                sid,
                            ),
                        )
                    )
                
                if rank == 0:
                    cache_progress.update(1)
            
            if rank == 0:
                cache_progress.close()
                logger.info(f"Cached {len(cache)} batches in GPU")
        
        shuffle(cache)
        data_iterator = cache
    else:
        data_iterator = enumerate(train_loader)

    # Initialize tqdm progress bar for training
    if rank == 0:
        epoch_progress = tqdm(total=len(train_loader), 
                            desc=f"Epoch {epoch}/{hps.train.epochs}", 
                            position=0, 
                            leave=True,
                            bar_format='{l_bar}{bar:20}{r_bar}{bar:-20b}')
    
    epoch_recorder = EpochRecorder()
    
    for batch_idx, info in data_iterator:
        # Unpack data
        if hps.if_f0 == 1:
            (
                phone,
                phone_lengths,
                pitch,
                pitchf,
                spec,
                spec_lengths,
                wave,
                wave_lengths,
                sid,
            ) = info
        else:
            phone, phone_lengths, spec, spec_lengths, wave, wave_lengths, sid = info
        
        if (hps.if_cache_data_in_gpu == False) and torch.cuda.is_available():
            phone = phone.cuda(rank, non_blocking=True)
            phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
            if hps.if_f0 == 1:
                pitch = pitch.cuda(rank, non_blocking=True)
                pitchf = pitchf.cuda(rank, non_blocking=True)
            sid = sid.cuda(rank, non_blocking=True)
            spec = spec.cuda(rank, non_blocking=True)
            spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
            wave = wave.cuda(rank, non_blocking=True)

        # Forward pass
        with autocast(enabled=hps.train.fp16_run):
            if hps.if_f0 == 1:
                (
                    y_hat,
                    ids_slice,
                    x_mask,
                    z_mask,
                    (z, z_p, m_p, logs_p, m_q, logs_q),
                ) = net_g(phone, phone_lengths, pitch, pitchf, spec, spec_lengths, sid)
            else:
                (
                    y_hat,
                    ids_slice,
                    x_mask,
                    z_mask,
                    (z, z_p, m_p, logs_p, m_q, logs_q),
                ) = net_g(phone, phone_lengths, spec, spec_lengths, sid)
            
            mel = spec_to_mel_torch(
                spec,
                hps.data.filter_length,
                hps.data.n_mel_channels,
                hps.data.sampling_rate,
                hps.data.mel_fmin,
                hps.data.mel_fmax,
            )
            y_mel = commons.slice_segments(
                mel, ids_slice, hps.train.segment_size // hps.data.hop_length
            )
            
            with autocast(enabled=False):
                y_hat_mel = mel_spectrogram_torch(
                    y_hat.float().squeeze(1),
                    hps.data.filter_length,
                    hps.data.n_mel_channels,
                    hps.data.sampling_rate,
                    hps.data.hop_length,
                    hps.data.win_length,
                    hps.data.mel_fmin,
                    hps.data.mel_fmax,
                )
            
            if hps.train.fp16_run == True:
                y_hat_mel = y_hat_mel.half()
            
            wave = commons.slice_segments(
                wave, ids_slice * hps.data.hop_length, hps.train.segment_size
            )

            # Discriminator forward
            y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, y_hat.detach())
            with autocast(enabled=False):
                loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(
                    y_d_hat_r, y_d_hat_g
                )
        
        # Discriminator backward
        optim_d.zero_grad()
        scaler.scale(loss_disc).backward()
        scaler.unscale_(optim_d)
        grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
        scaler.step(optim_d)

        # Generator forward
        with autocast(enabled=hps.train.fp16_run):
            y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(wave, y_hat)
            with autocast(enabled=False):
                loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
                loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
                loss_fm = feature_loss(fmap_r, fmap_g)
                loss_gen, losses_gen = generator_loss(y_d_hat_g)
                loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl
        
        # Generator backward
        optim_g.zero_grad()
        scaler.scale(loss_gen_all).backward()
        scaler.unscale_(optim_g)
        grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
        scaler.step(optim_g)
        scaler.update()

        # Update progress bar and logging
        if rank == 0:
            if epoch_progress is not None:
                epoch_progress.update(1)
                
                # Update progress bar description with current losses
                if batch_idx % hps.train.log_interval == 0:
                    postfix_dict = {
                        'G': f'{loss_gen_all:.3f}',
                        'D': f'{loss_disc:.3f}',
                        'Mel': f'{loss_mel:.3f}',
                        'KL': f'{loss_kl:.3f}',
                        'Step': global_step
                    }
                    epoch_progress.set_postfix(postfix_dict)
            
            if global_step % hps.train.log_interval == 0:
                lr = optim_g.param_groups[0]["lr"]
                
                logger.info(f"\nEpoch: {epoch} [{batch_idx}/{len(train_loader)}]")
                logger.info(f"Global Step: {global_step}")
                logger.info(f"Learning Rate: {lr:.6f}")
                logger.info(f"Generator Loss: {loss_gen_all:.3f} (FM: {loss_fm:.3f}, Mel: {loss_mel:.3f}, KL: {loss_kl:.3f})")
                logger.info(f"Discriminator Loss: {loss_disc:.3f}")
                logger.info(f"Grad Norm - G: {grad_norm_g:.3f}, D: {grad_norm_d:.3f}")
                
                # Tensorboard logging
                scalar_dict = {
                    "loss/g/total": loss_gen_all,
                    "loss/d/total": loss_disc,
                    "learning_rate": lr,
                    "grad_norm_d": grad_norm_d,
                    "grad_norm_g": grad_norm_g,
                    "loss/g/fm": loss_fm,
                    "loss/g/mel": loss_mel,
                    "loss/g/kl": loss_kl,
                }
                scalar_dict.update(
                    {"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}
                )
                scalar_dict.update(
                    {"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}
                )
                scalar_dict.update(
                    {"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}
                )
                
                image_dict = {
                    "slice/mel_org": utils.plot_spectrogram_to_numpy(
                        y_mel[0].data.cpu().numpy()
                    ),
                    "slice/mel_gen": utils.plot_spectrogram_to_numpy(
                        y_hat_mel[0].data.cpu().numpy()
                    ),
                    "all/mel": utils.plot_spectrogram_to_numpy(
                        mel[0].data.cpu().numpy()
                    ),
                }
                
                utils.summarize(
                    writer=writer,
                    global_step=global_step,
                    images=image_dict,
                    scalars=scalar_dict,
                )
        
        global_step += 1
    
    # Close progress bar
    if rank == 0 and epoch_progress is not None:
        epoch_progress.close()
    
    # Save checkpoints
    if epoch % hps.save_every_epoch == 0 and rank == 0:
        if hps.if_latest == 0:
            save_path_g = os.path.join(hps.model_dir, f"G_{global_step}.pth")
            save_path_d = os.path.join(hps.model_dir, f"D_{global_step}.pth")
            utils.save_checkpoint(
                net_g,
                optim_g,
                hps.train.learning_rate,
                epoch,
                save_path_g,
            )
            utils.save_checkpoint(
                net_d,
                optim_d,
                hps.train.learning_rate,
                epoch,
                save_path_d,
            )
            logger.info(f"Saved checkpoints: {save_path_g}, {save_path_d}")
        else:
            save_path_g = os.path.join(hps.model_dir, "G_2333333.pth")
            save_path_d = os.path.join(hps.model_dir, "D_2333333.pth")
            utils.save_checkpoint(
                net_g,
                optim_g,
                hps.train.learning_rate,
                epoch,
                save_path_g,
            )
            utils.save_checkpoint(
                net_d,
                optim_d,
                hps.train.learning_rate,
                epoch,
                save_path_d,
            )
            logger.info(f"Saved latest checkpoints: {save_path_g}, {save_path_d}")
        
        if rank == 0 and hps.save_every_weights == "1":
            if hasattr(net_g, "module"):
                ckpt = net_g.module.state_dict()
            else:
                ckpt = net_g.state_dict()
            
            model_name = hps.name + f"_e{epoch}_s{global_step}"
            save_result = savee(
                ckpt,
                hps.sample_rate,
                hps.if_f0,
                model_name,
                epoch,
                hps.version,
                hps,
            )
            logger.info(f"Saved weights checkpoint: {model_name}: {save_result}")

    # Log epoch completion
    if rank == 0:
        logger.info(f"Completed Epoch {epoch} {epoch_recorder.record()}")
        logger.info(f"Global Step: {global_step}")
    
    # End training if completed
    if epoch >= hps.total_epoch and rank == 0:
        logger.info("Training completed!")
        
        if hasattr(net_g, "module"):
            ckpt = net_g.module.state_dict()
        else:
            ckpt = net_g.state_dict()
        
        final_save = savee(
            ckpt, hps.sample_rate, hps.if_f0, hps.name, epoch, hps.version, hps
        )
        logger.info(f"Saved final model: {final_save}")
        
        sleep(2)  # Give time for final logging
        os._exit(0)


if __name__ == "__main__":
    torch.multiprocessing.set_start_method("spawn")
    main()