Upload 2 files

data_utils.py

@@ -0,0 +1,392 @@
+import time
+import os
+import random
+import numpy as np
+import torch
+import torch.utils.data
+
+import commons 
+from mel_processing import spectrogram_torch
+from utils import load_wav_to_torch, load_filepaths_and_text
+from text import text_to_sequence, cleaned_text_to_sequence
+
+
+class TextAudioLoader(torch.utils.data.Dataset):
+    """
+        1) loads audio, text pairs
+        2) normalizes text and converts them to sequences of integers
+        3) computes spectrograms from audio files.
+    """
+    def __init__(self, audiopaths_and_text, hparams):
+        self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
+        self.text_cleaners  = hparams.text_cleaners
+        self.max_wav_value  = hparams.max_wav_value
+        self.sampling_rate  = hparams.sampling_rate
+        self.filter_length  = hparams.filter_length 
+        self.hop_length     = hparams.hop_length 
+        self.win_length     = hparams.win_length
+        self.sampling_rate  = hparams.sampling_rate 
+
+        self.cleaned_text = getattr(hparams, "cleaned_text", False)
+
+        self.add_blank = hparams.add_blank
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 190)
+
+        random.seed(1234)
+        random.shuffle(self.audiopaths_and_text)
+        self._filter()
+
+
+    def _filter(self):
+        """
+        Filter text & store spec lengths
+        """
+        # Store spectrogram lengths for Bucketing
+        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+        # spec_length = wav_length // hop_length
+
+        audiopaths_and_text_new = []
+        lengths = []
+        for audiopath, text in self.audiopaths_and_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_and_text_new.append([audiopath, text])
+                lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
+        self.audiopaths_and_text = audiopaths_and_text_new
+        self.lengths = lengths
+
+    def get_audio_text_pair(self, audiopath_and_text):
+        # separate filename and text
+        audiopath, text = audiopath_and_text[0], audiopath_and_text[1]
+        text = self.get_text(text)
+        spec, wav = self.get_audio(audiopath)
+        return (text, spec, wav)
+
+    def get_audio(self, filename):
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError("{} {} SR doesn't match target {} SR".format(
+                sampling_rate, self.sampling_rate))
+        audio_norm = audio / self.max_wav_value
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pt")
+        if os.path.exists(spec_filename):
+            spec = torch.load(spec_filename)
+        else:
+            spec = spectrogram_torch(audio_norm, self.filter_length,
+                self.sampling_rate, self.hop_length, self.win_length,
+                center=False)
+            spec = torch.squeeze(spec, 0)
+            torch.save(spec, spec_filename)
+        return spec, audio_norm
+
+    def get_text(self, text):
+        if self.cleaned_text:
+            text_norm = cleaned_text_to_sequence(text)
+        else:
+            text_norm = text_to_sequence(text, self.text_cleaners)
+        if self.add_blank:
+            text_norm = commons.intersperse(text_norm, 0)
+        text_norm = torch.LongTensor(text_norm)
+        return text_norm
+
+    def __getitem__(self, index):
+        return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class TextAudioCollate():
+    """ Zero-pads model inputs and targets
+    """
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text and aduio
+        PARAMS
+        ------
+        batch: [text_normalized, spec_normalized, wav_normalized]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        _, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([x[1].size(1) for x in batch]),
+            dim=0, descending=True)
+
+        max_text_len = max([len(x[0]) for x in batch])
+        max_spec_len = max([x[1].size(1) for x in batch])
+        max_wav_len = max([x[2].size(1) for x in batch])
+
+        text_lengths = torch.LongTensor(len(batch))
+        spec_lengths = torch.LongTensor(len(batch))
+        wav_lengths = torch.LongTensor(len(batch))
+
+        text_padded = torch.LongTensor(len(batch), max_text_len)
+        spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
+        wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
+        text_padded.zero_()
+        spec_padded.zero_()
+        wav_padded.zero_()
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            text = row[0]
+            text_padded[i, :text.size(0)] = text
+            text_lengths[i] = text.size(0)
+
+            spec = row[1]
+            spec_padded[i, :, :spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wav = row[2]
+            wav_padded[i, :, :wav.size(1)] = wav
+            wav_lengths[i] = wav.size(1)
+
+        if self.return_ids:
+            return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, ids_sorted_decreasing
+        return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths
+
+
+"""Multi speaker version"""
+class TextAudioSpeakerLoader(torch.utils.data.Dataset):
+    """
+        1) loads audio, speaker_id, text pairs
+        2) normalizes text and converts them to sequences of integers
+        3) computes spectrograms from audio files.
+    """
+    def __init__(self, audiopaths_sid_text, hparams):
+        self.audiopaths_sid_text = load_filepaths_and_text(audiopaths_sid_text)
+        self.text_cleaners = hparams.text_cleaners
+        self.max_wav_value = hparams.max_wav_value
+        self.sampling_rate = hparams.sampling_rate
+        self.filter_length  = hparams.filter_length
+        self.hop_length     = hparams.hop_length
+        self.win_length     = hparams.win_length
+        self.sampling_rate  = hparams.sampling_rate
+
+        self.cleaned_text = getattr(hparams, "cleaned_text", False)
+
+        self.add_blank = hparams.add_blank
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 190)
+
+        random.seed(1234)
+        random.shuffle(self.audiopaths_sid_text)
+        self._filter()
+
+    def _filter(self):
+        """
+        Filter text & store spec lengths
+        """
+        # Store spectrogram lengths for Bucketing
+        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+        # spec_length = wav_length // hop_length
+
+        audiopaths_sid_text_new = []
+        lengths = []
+        for audiopath, sid, text in self.audiopaths_sid_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_sid_text_new.append([audiopath, sid, text])
+                lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
+        self.audiopaths_sid_text = audiopaths_sid_text_new
+        self.lengths = lengths
+
+    def get_audio_text_speaker_pair(self, audiopath_sid_text):
+        # separate filename, speaker_id and text
+        audiopath, sid, text = audiopath_sid_text[0], audiopath_sid_text[1], audiopath_sid_text[2]
+        text = self.get_text(text)
+        spec, wav = self.get_audio(audiopath)
+        sid = self.get_sid(sid)
+        return (text, spec, wav, sid)
+
+    def get_audio(self, filename):
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError("{} {} SR doesn't match target {} SR".format(
+                sampling_rate, self.sampling_rate))
+        audio_norm = audio / self.max_wav_value
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pt")
+        if os.path.exists(spec_filename):
+            spec = torch.load(spec_filename)
+        else:
+            spec = spectrogram_torch(audio_norm, self.filter_length,
+                self.sampling_rate, self.hop_length, self.win_length,
+                center=False)
+            spec = torch.squeeze(spec, 0)
+            torch.save(spec, spec_filename)
+        return spec, audio_norm
+
+    def get_text(self, text):
+        if self.cleaned_text:
+            text_norm = cleaned_text_to_sequence(text)
+        else:
+            text_norm = text_to_sequence(text, self.text_cleaners)
+        if self.add_blank:
+            text_norm = commons.intersperse(text_norm, 0)
+        text_norm = torch.LongTensor(text_norm)
+        return text_norm
+
+    def get_sid(self, sid):
+        sid = torch.LongTensor([int(sid)])
+        return sid
+
+    def __getitem__(self, index):
+        return self.get_audio_text_speaker_pair(self.audiopaths_sid_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_sid_text)
+
+
+class TextAudioSpeakerCollate():
+    """ Zero-pads model inputs and targets
+    """
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text, audio and speaker identities
+        PARAMS
+        ------
+        batch: [text_normalized, spec_normalized, wav_normalized, sid]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        _, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([x[1].size(1) for x in batch]),
+            dim=0, descending=True)
+
+        max_text_len = max([len(x[0]) for x in batch])
+        max_spec_len = max([x[1].size(1) for x in batch])
+        max_wav_len = max([x[2].size(1) for x in batch])
+
+        text_lengths = torch.LongTensor(len(batch))
+        spec_lengths = torch.LongTensor(len(batch))
+        wav_lengths = torch.LongTensor(len(batch))
+        sid = torch.LongTensor(len(batch))
+
+        text_padded = torch.LongTensor(len(batch), max_text_len)
+        spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
+        wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
+        text_padded.zero_()
+        spec_padded.zero_()
+        wav_padded.zero_()
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            text = row[0]
+            text_padded[i, :text.size(0)] = text
+            text_lengths[i] = text.size(0)
+
+            spec = row[1]
+            spec_padded[i, :, :spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wav = row[2]
+            wav_padded[i, :, :wav.size(1)] = wav
+            wav_lengths[i] = wav.size(1)
+
+            sid[i] = row[3]
+
+        if self.return_ids:
+            return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid, ids_sorted_decreasing
+        return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid
+
+
+class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
+    """
+    Maintain similar input lengths in a batch.
+    Length groups are specified by boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
+  
+    It removes samples which are not included in the boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
+    """
+    def __init__(self, dataset, batch_size, boundaries, num_replicas=None, rank=None, shuffle=True):
+        super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
+        self.lengths = dataset.lengths
+        self.batch_size = batch_size
+        self.boundaries = boundaries
+  
+        self.buckets, self.num_samples_per_bucket = self._create_buckets()
+        self.total_size = sum(self.num_samples_per_bucket)
+        self.num_samples = self.total_size // self.num_replicas
+  
+    def _create_buckets(self):
+        buckets = [[] for _ in range(len(self.boundaries) - 1)]
+        for i in range(len(self.lengths)):
+            length = self.lengths[i]
+            idx_bucket = self._bisect(length)
+            if idx_bucket != -1:
+                buckets[idx_bucket].append(i)
+  
+        for i in range(len(buckets) - 1, 0, -1):
+            if len(buckets[i]) == 0:
+                buckets.pop(i)
+                self.boundaries.pop(i+1)
+  
+        num_samples_per_bucket = []
+        for i in range(len(buckets)):
+            len_bucket = len(buckets[i])
+            total_batch_size = self.num_replicas * self.batch_size
+            rem = (total_batch_size - (len_bucket % total_batch_size)) % total_batch_size
+            num_samples_per_bucket.append(len_bucket + rem)
+        return buckets, num_samples_per_bucket
+  
+    def __iter__(self):
+      # deterministically shuffle based on epoch
+      g = torch.Generator()
+      g.manual_seed(self.epoch)
+  
+      indices = []
+      if self.shuffle:
+          for bucket in self.buckets:
+              indices.append(torch.randperm(len(bucket), generator=g).tolist())
+      else:
+          for bucket in self.buckets:
+              indices.append(list(range(len(bucket))))
+  
+      batches = []
+      for i in range(len(self.buckets)):
+          bucket = self.buckets[i]
+          len_bucket = len(bucket)
+          ids_bucket = indices[i]
+          num_samples_bucket = self.num_samples_per_bucket[i]
+  
+          # add extra samples to make it evenly divisible
+          rem = num_samples_bucket - len_bucket
+          ids_bucket = ids_bucket + ids_bucket * (rem // len_bucket) + ids_bucket[:(rem % len_bucket)]
+  
+          # subsample
+          ids_bucket = ids_bucket[self.rank::self.num_replicas]
+  
+          # batching
+          for j in range(len(ids_bucket) // self.batch_size):
+              batch = [bucket[idx] for idx in ids_bucket[j*self.batch_size:(j+1)*self.batch_size]]
+              batches.append(batch)
+  
+      if self.shuffle:
+          batch_ids = torch.randperm(len(batches), generator=g).tolist()
+          batches = [batches[i] for i in batch_ids]
+      self.batches = batches
+  
+      assert len(self.batches) * self.batch_size == self.num_samples
+      return iter(self.batches)
+  
+    def _bisect(self, x, lo=0, hi=None):
+      if hi is None:
+          hi = len(self.boundaries) - 1
+  
+      if hi > lo:
+          mid = (hi + lo) // 2
+          if self.boundaries[mid] < x and x <= self.boundaries[mid+1]:
+              return mid
+          elif x <= self.boundaries[mid]:
+              return self._bisect(x, lo, mid)
+          else:
+              return self._bisect(x, mid + 1, hi)
+      else:
+          return -1
+
+    def __len__(self):
+        return self.num_samples // self.batch_size

train.py

@@ -0,0 +1,290 @@
+import os
+import json
+import argparse
+import itertools
+import math
+import torch
+from torch import nn, optim
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+from torch.utils.tensorboard import SummaryWriter
+import torch.multiprocessing as mp
+import torch.distributed as dist
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.cuda.amp import autocast, GradScaler
+
+import commons
+import utils
+from data_utils import (
+  TextAudioLoader,
+  TextAudioCollate,
+  DistributedBucketSampler
+)
+from models import (
+  SynthesizerTrn,
+  MultiPeriodDiscriminator,
+)
+from losses import (
+  generator_loss,
+  discriminator_loss,
+  feature_loss,
+  kl_loss
+)
+from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
+from text.symbols import symbols
+
+
+torch.backends.cudnn.benchmark = True
+global_step = 0
+
+
+def main():
+  """Assume Single Node Multi GPUs Training Only"""
+  assert torch.cuda.is_available(), "CPU training is not allowed."
+
+  n_gpus = torch.cuda.device_count()
+  os.environ['MASTER_ADDR'] = 'localhost'
+  os.environ['MASTER_PORT'] = '80000'
+
+  hps = utils.get_hparams()
+  mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,))
+
+
+def run(rank, n_gpus, hps):
+  global global_step
+  if rank == 0:
+    logger = utils.get_logger(hps.model_dir)
+    logger.info(hps)
+    utils.check_git_hash(hps.model_dir)
+    writer = SummaryWriter(log_dir=hps.model_dir)
+    writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
+
+  dist.init_process_group(backend='nccl', init_method='env://', world_size=n_gpus, rank=rank)
+  torch.manual_seed(hps.train.seed)
+  torch.cuda.set_device(rank)
+
+  train_dataset = TextAudioLoader(hps.data.training_files, hps.data)
+  train_sampler = DistributedBucketSampler(
+      train_dataset,
+      hps.train.batch_size,
+      [32,300,400,500,600,700,800,900,1000],
+      num_replicas=n_gpus,
+      rank=rank,
+      shuffle=True)
+  collate_fn = TextAudioCollate()
+  train_loader = DataLoader(train_dataset, num_workers=8, shuffle=False, pin_memory=True,
+      collate_fn=collate_fn, batch_sampler=train_sampler)
+  if rank == 0:
+    eval_dataset = TextAudioLoader(hps.data.validation_files, hps.data)
+    eval_loader = DataLoader(eval_dataset, num_workers=8, shuffle=False,
+        batch_size=hps.train.batch_size, pin_memory=True,
+        drop_last=False, collate_fn=collate_fn)
+
+  net_g = SynthesizerTrn(
+      len(symbols),
+      hps.data.filter_length // 2 + 1,
+      hps.train.segment_size // hps.data.hop_length,
+      **hps.model).cuda(rank)
+  net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).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)
+  net_g = DDP(net_g, device_ids=[rank])
+  net_d = DDP(net_d, device_ids=[rank])
+
+  try:
+    _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g)
+    _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d)
+    global_step = (epoch_str - 1) * len(train_loader)
+  except:
+    epoch_str = 1
+    global_step = 0
+
+  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)
+
+  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, eval_loader], logger, [writer, writer_eval])
+    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)
+    scheduler_g.step()
+    scheduler_d.step()
+
+
+def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers):
+  net_g, net_d = nets
+  optim_g, optim_d = optims
+  scheduler_g, scheduler_d = schedulers
+  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()
+  for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths) in enumerate(train_loader):
+    x, x_lengths = x.cuda(rank, non_blocking=True), x_lengths.cuda(rank, non_blocking=True)
+    spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(rank, non_blocking=True)
+    y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(rank, non_blocking=True)
+
+    with autocast(enabled=hps.train.fp16_run):
+      y_hat, l_length, attn, ids_slice, x_mask, z_mask,\
+      (z, z_p, m_p, logs_p, m_q, logs_q) = net_g(x, x_lengths, spec, spec_lengths)
+
+      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)
+      y_hat_mel = mel_spectrogram_torch(
+          y_hat.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
+      )
+
+      y = commons.slice_segments(y, ids_slice * hps.data.hop_length, hps.train.segment_size) # slice 
+
+      # Discriminator
+      y_d_hat_r, y_d_hat_g, _, _ = net_d(y, 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)
+        loss_disc_all = loss_disc
+    optim_d.zero_grad()
+    scaler.scale(loss_disc_all).backward()
+    scaler.unscale_(optim_d)
+    grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
+    scaler.step(optim_d)
+
+    with autocast(enabled=hps.train.fp16_run):
+      # Generator
+      y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
+      with autocast(enabled=False):
+        loss_dur = torch.sum(l_length.float())
+        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_dur + loss_kl
+    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()
+
+    if rank==0:
+      if global_step % hps.train.log_interval == 0:
+        lr = optim_g.param_groups[0]['lr']
+        losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_dur, loss_kl]
+        logger.info('Train Epoch: {} [{:.0f}%]'.format(
+          epoch,
+          100. * batch_idx / len(train_loader)))
+        logger.info([x.item() for x in losses] + [global_step, lr])
+        
+        scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc_all, "learning_rate": lr, "grad_norm_d": grad_norm_d, "grad_norm_g": grad_norm_g}
+        scalar_dict.update({"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/dur": loss_dur, "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()),
+            "all/attn": utils.plot_alignment_to_numpy(attn[0,0].data.cpu().numpy())
+        }
+        utils.summarize(
+          writer=writer,
+          global_step=global_step, 
+          images=image_dict,
+          scalars=scalar_dict)
+
+      if global_step % hps.train.eval_interval == 0:
+        evaluate(hps, net_g, eval_loader, writer_eval)
+        utils.save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "G_{}.pth".format(global_step)))
+        utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "D_{}.pth".format(global_step)))
+    global_step += 1
+  
+  if rank == 0:
+    logger.info('====> Epoch: {}'.format(epoch))
+
+ 
+def evaluate(hps, generator, eval_loader, writer_eval):
+    generator.eval()
+    with torch.no_grad():
+      for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths) in enumerate(eval_loader):
+        x, x_lengths = x.cuda(0), x_lengths.cuda(0)
+        spec, spec_lengths = spec.cuda(0), spec_lengths.cuda(0)
+        y, y_lengths = y.cuda(0), y_lengths.cuda(0)
+
+        # remove else
+        x = x[:1]
+        x_lengths = x_lengths[:1]
+        spec = spec[:1]
+        spec_lengths = spec_lengths[:1]
+        y = y[:1]
+        y_lengths = y_lengths[:1]
+        break
+      y_hat, attn, mask, *_ = generator.module.infer(x, x_lengths, max_len=1000)
+      y_hat_lengths = mask.sum([1,2]).long() * hps.data.hop_length
+
+      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_hat_mel = mel_spectrogram_torch(
+        y_hat.squeeze(1).float(),
+        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
+      )
+    image_dict = {
+      "gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy())
+    }
+    audio_dict = {
+      "gen/audio": y_hat[0,:,:y_hat_lengths[0]]
+    }
+    if global_step == 0:
+      image_dict.update({"gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())})
+      audio_dict.update({"gt/audio": y[0,:,:y_lengths[0]]})
+
+    utils.summarize(
+      writer=writer_eval,
+      global_step=global_step, 
+      images=image_dict,
+      audios=audio_dict,
+      audio_sampling_rate=hps.data.sampling_rate
+    )
+    generator.train()
+
+                           
+if __name__ == "__main__":
+  main()