Fix imports

TTS/tts/layers/glow_tts/encoder.py

@@ -9,7 +9,7 @@ from TTS.tts.layers.generic.time_depth_sep_conv import TimeDepthSeparableConvBlo
 from TTS.tts.layers.glow_tts.duration_predictor import DurationPredictor
 from TTS.tts.layers.glow_tts.glow import ResidualConv1dLayerNormBlock
 from TTS.tts.layers.glow_tts.transformer import RelativePositionTransformer
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 
 
 class Encoder(nn.Module):

TTS/tts/layers/glow_tts/monotonic_align/__init__.py

@@ -2,7 +2,7 @@ import numpy as np
 import torch
 from torch.nn import functional as F
 
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 
 try:
     # TODO: fix pypi cython installation problem.

TTS/tts/layers/losses.py

@@ -6,7 +6,7 @@ from coqpit import Coqpit
 from torch import nn
 from torch.nn import functional
 
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 from TTS.tts.utils.ssim import ssim
 from TTS.utils.audio import TorchSTFT
 

TTS/tts/layers/vits/networks.py

@@ -5,7 +5,7 @@ from torch import nn
 
 from TTS.tts.layers.glow_tts.glow import WN
 from TTS.tts.layers.glow_tts.transformer import RelativePositionTransformer
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 
 LRELU_SLOPE = 0.1
 

TTS/tts/models/align_tts.py

@@ -12,7 +12,7 @@ from TTS.tts.layers.feed_forward.encoder import Encoder
 from TTS.tts.layers.generic.pos_encoding import PositionalEncoding
 from TTS.tts.layers.glow_tts.monotonic_align import generate_path, maximum_path
 from TTS.tts.models.base_tts import BaseTTS
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 from TTS.tts.utils.visual import plot_alignment, plot_spectrogram
 from TTS.utils.audio import AudioProcessor
 from TTS.utils.io import load_fsspec

TTS/tts/models/base_tacotron.py

@@ -9,7 +9,7 @@ from torch import nn
 
 from TTS.tts.layers.losses import TacotronLoss
 from TTS.tts.models.base_tts import BaseTTS
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 from TTS.tts.utils.speakers import SpeakerManager, get_speaker_manager
 from TTS.tts.utils.text import make_symbols
 from TTS.utils.generic_utils import format_aux_input

TTS/tts/models/base_tts.py

@@ -1,3 +1,4 @@
+from TTS.tts.utils.helpers import segment
 import os
 from typing import Dict, List, Tuple
 

TTS/tts/models/fast_pitch.py

@@ -13,9 +13,10 @@ from TTS.tts.layers.generic.pos_encoding import PositionalEncoding
 from TTS.tts.layers.glow_tts.duration_predictor import DurationPredictor
 from TTS.tts.layers.glow_tts.monotonic_align import generate_path, maximum_path
 from TTS.tts.models.base_tts import BaseTTS
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 from TTS.tts.utils.visual import plot_alignment, plot_pitch, plot_spectrogram
 from TTS.utils.audio import AudioProcessor
+from TTS.tts.utils.helpers import average_over_durations
 
 
 @dataclass
@@ -416,7 +417,7 @@ class FastPitch(BaseTTS):
         """
         o_pitch = self.pitch_predictor(o_en, x_mask)
         if pitch is not None:
-            avg_pitch = average_pitch(pitch, dr)
+            avg_pitch = average_over_durations(pitch, dr)
             o_pitch_emb = self.pitch_emb(avg_pitch)
             return o_pitch_emb, o_pitch, avg_pitch
         o_pitch_emb = self.pitch_emb(o_pitch)
@@ -670,28 +671,3 @@ class FastPitch(BaseTTS):
         """Enable binary alignment loss when needed"""
         if trainer.total_steps_done > self.config.binary_align_loss_start_step:
             self.use_binary_alignment_loss = True
-
-
-def average_pitch(pitch, durs):
-    """Compute the average pitch value for each input character based on the durations.
-
-    Shapes:
-        - pitch: :math:`[B, 1, T_de]`
-        - durs: :math:`[B, T_en]`
-    """
-
-    durs_cums_ends = torch.cumsum(durs, dim=1).long()
-    durs_cums_starts = torch.nn.functional.pad(durs_cums_ends[:, :-1], (1, 0))
-    pitch_nonzero_cums = torch.nn.functional.pad(torch.cumsum(pitch != 0.0, dim=2), (1, 0))
-    pitch_cums = torch.nn.functional.pad(torch.cumsum(pitch, dim=2), (1, 0))
-
-    bs, l = durs_cums_ends.size()
-    n_formants = pitch.size(1)
-    dcs = durs_cums_starts[:, None, :].expand(bs, n_formants, l)
-    dce = durs_cums_ends[:, None, :].expand(bs, n_formants, l)
-
-    pitch_sums = (torch.gather(pitch_cums, 2, dce) - torch.gather(pitch_cums, 2, dcs)).float()
-    pitch_nelems = (torch.gather(pitch_nonzero_cums, 2, dce) - torch.gather(pitch_nonzero_cums, 2, dcs)).float()
-
-    pitch_avg = torch.where(pitch_nelems == 0.0, pitch_nelems, pitch_sums / pitch_nelems)
-    return pitch_avg

TTS/tts/models/glow_tts.py

@@ -9,7 +9,7 @@ from TTS.tts.layers.glow_tts.decoder import Decoder
 from TTS.tts.layers.glow_tts.encoder import Encoder
 from TTS.tts.layers.glow_tts.monotonic_align import generate_path, maximum_path
 from TTS.tts.models.base_tts import BaseTTS
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 from TTS.tts.utils.speakers import get_speaker_manager
 from TTS.tts.utils.synthesis import synthesis
 from TTS.tts.utils.visual import plot_alignment, plot_spectrogram

TTS/tts/models/speedy_speech.py

@@ -10,7 +10,7 @@ from TTS.tts.layers.feed_forward.encoder import Encoder
 from TTS.tts.layers.generic.pos_encoding import PositionalEncoding
 from TTS.tts.layers.glow_tts.monotonic_align import generate_path
 from TTS.tts.models.base_tts import BaseTTS
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 from TTS.tts.utils.measures import alignment_diagonal_score
 from TTS.tts.utils.visual import plot_alignment, plot_spectrogram
 from TTS.utils.audio import AudioProcessor

TTS/tts/models/vits.py

@@ -14,7 +14,7 @@ from TTS.tts.layers.vits.discriminator import VitsDiscriminator
 from TTS.tts.layers.vits.networks import PosteriorEncoder, ResidualCouplingBlocks, TextEncoder
 from TTS.tts.layers.vits.stochastic_duration_predictor import StochasticDurationPredictor
 from TTS.tts.models.base_tts import BaseTTS
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 from TTS.tts.utils.speakers import get_speaker_manager
 from TTS.tts.utils.synthesis import synthesis
 from TTS.tts.utils.visual import plot_alignment
@@ -22,28 +22,7 @@ from TTS.utils.audio import AudioProcessor
 from TTS.utils.trainer_utils import get_optimizer, get_scheduler
 from TTS.vocoder.models.hifigan_generator import HifiganGenerator
 from TTS.vocoder.utils.generic_utils import plot_results
-
-
-def segment(x: torch.tensor, segment_indices: torch.tensor, segment_size=4):
-    """Segment each sample in a batch based on the provided segment indices"""
-    segments = torch.zeros_like(x[:, :, :segment_size])
-    for i in range(x.size(0)):
-        index_start = segment_indices[i]
-        index_end = index_start + segment_size
-        segments[i] = x[i, :, index_start:index_end]
-    return segments
-
-
-def rand_segment(x: torch.tensor, x_lengths: torch.tensor = None, segment_size=4):
-    """Create random segments based on the input lengths."""
-    B, _, T = x.size()
-    if x_lengths is None:
-        x_lengths = T
-    max_idxs = x_lengths - segment_size + 1
-    assert all(max_idxs > 0), " [!] At least one sample is shorter than the segment size."
-    segment_indices = (torch.rand([B]).type_as(x) * max_idxs).long()
-    ret = segment(x, segment_indices, segment_size)
-    return ret, segment_indices
+from TTS.tts.utils.helpers import rand_segments, segment
 
 
 @dataclass
@@ -451,7 +430,7 @@ class Vits(BaseTTS):
         logs_p = torch.einsum("klmn, kjm -> kjn", [attn, logs_p])
 
         # select a random feature segment for the waveform decoder
-        z_slice, slice_ids = rand_segment(z, y_lengths, self.spec_segment_size)
+        z_slice, slice_ids = rand_segments(z, y_lengths, self.spec_segment_size)
         o = self.waveform_decoder(z_slice, g=g)
         outputs.update(
             {

TTS/tts/utils/data.py

@@ -54,33 +54,3 @@ def pad_per_step(inputs, pad_len):
     return np.pad(inputs, [[0, 0], [0, 0], [0, pad_len]], mode="constant", constant_values=0.0)
 
 
-# pylint: disable=attribute-defined-outside-init
-class StandardScaler:
-    def set_stats(self, mean, scale):
-        self.mean_ = mean
-        self.scale_ = scale
-
-    def reset_stats(self):
-        delattr(self, "mean_")
-        delattr(self, "scale_")
-
-    def transform(self, X):
-        X = np.asarray(X)
-        X -= self.mean_
-        X /= self.scale_
-        return X
-
-    def inverse_transform(self, X):
-        X = np.asarray(X)
-        X *= self.scale_
-        X += self.mean_
-        return X
-
-
-# from https://gist.github.com/jihunchoi/f1434a77df9db1bb337417854b398df1
-def sequence_mask(sequence_length, max_len=None):
-    if max_len is None:
-        max_len = sequence_length.data.max()
-    seq_range = torch.arange(max_len, dtype=sequence_length.dtype, device=sequence_length.device)
-    # B x T_max
-    return seq_range.unsqueeze(0) < sequence_length.unsqueeze(1)

TTS/tts/utils/visual.py

@@ -101,6 +101,7 @@ def visualize(
     figsize=(8, 24),
     output_fig=False,
 ):
+    """Intended to be used in Notebooks."""
 
     if decoder_output is not None:
         num_plot = 4

TTS/utils/audio.py

@@ -9,7 +9,7 @@ import soundfile as sf
 import torch
 from torch import nn
 
-from TTS.tts.utils.data import StandardScaler
+from TTS.tts.utils.helpers import StandardScaler
 
 
 class TorchSTFT(nn.Module):  # pylint: disable=abstract-method

docs/source/faq.md

@@ -7,7 +7,7 @@ We tried to collect common issues and questions we receive about 🐸TTS. It is
 - If you feel like it's a bug to be fixed, then prefer Github issues with the same level of scrutiny.
 
 ## What are the requirements of a good 🐸TTS dataset?
-* https://github.com/coqui-ai/TTS/wiki/What-makes-a-good-TTS-dataset
+* {ref}`See this page <what_makes_a_good_dataset>`
 
 ## How should I choose the right model?
 - First, train Tacotron. It is smaller and faster to experiment with. If it performs poorly, try Tacotron2.

docs/source/training_a_model.md

@@ -54,7 +54,7 @@
 
  4. Run the training.
 
-    You need to call the python training script.
+    You need to run the training script.
 
     ```bash
     $ CUDA_VISIBLE_DEVICES="0" python train_glowtts.py
@@ -63,7 +63,7 @@
     Notice that you set the GPU you want to use on your system by setting `CUDA_VISIBLE_DEVICES` environment variable.
     To see available GPUs on your system, you can use `nvidia-smi` command on the terminal.
 
-    If you like to run a multi-gpu training
+    If you like to run a multi-gpu training using DDP back-end,
 
     ```bash
     $ CUDA_VISIBLE_DEVICES="0, 1, 2" python TTS/bin/distribute.py --script <path_to_your_script>/train_glowtts.py

docs/source/what_makes_a_good_dataset.md

@@ -1,3 +1,4 @@
+(what_makes_a_good_dataset)=
 # What makes a good TTS dataset
 
 ## What Makes a Good Dataset

tests/tts_tests/test_feed_forward_layers.py

@@ -2,7 +2,7 @@ import torch
 
 from TTS.tts.layers.feed_forward.decoder import Decoder
 from TTS.tts.layers.feed_forward.encoder import Encoder
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 

tests/tts_tests/test_speedy_speech_layers.py

@@ -3,7 +3,7 @@ import torch
 from TTS.tts.configs import SpeedySpeechConfig
 from TTS.tts.layers.feed_forward.duration_predictor import DurationPredictor
 from TTS.tts.models.speedy_speech import SpeedySpeech, SpeedySpeechArgs
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 
 use_cuda = torch.cuda.is_available()
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

tests/tts_tests/test_tacotron_layers.py

@@ -4,7 +4,7 @@ import torch as T
 
 from TTS.tts.layers.losses import L1LossMasked, SSIMLoss
 from TTS.tts.layers.tacotron.tacotron import CBHG, Decoder, Encoder, Prenet
-from TTS.tts.utils.data import sequence_mask
+from TTS.tts.utils.helpers import sequence_mask
 
 # pylint: disable=unused-variable