Init deepvoice3 commit

.gitingore

+*.pyc
+*.tar.*

LICENSE

+   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+   
+
+Part of code was copied or adpated from https://github.com/r9y9/deepvoice3_pytorch/
+Copyright (c) 2017: Ryuichi Yamamoto, whose applies.
+

README.md

+# Parakeet
+
+Parakeet aims to provide a flexible, efficient and state-of-the-art text-to-speech toolkit for the open-source community. It is built on Paddle Fluid dynamic graph, with the support of many influential TTS models proposed by [Baidu Research](http://research.baidu.com) and other academic institutions.  
+
+## Installation
+
+### Install paddlepaddle
+
+For faster training speed and better support, it is recommended that you install the lasted develop version of paddlepaddle. Please refer to the [quick installation guide](https://paddlepaddle.org.cn/install/quick).
+
+### Other Requirements
+
+Install other requirements with pip.
+
+
+```bash
+pip install -r requirements.txt
+```
+
+## Supported models
+
+- [Deep Voice 3: Scaling Text-to-Speech with Convolutional Sequence Learning](./deepvoice3)

data/data.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import random
+
+import io
+import platform
+from os.path import dirname, join
+
+from nnmnkwii.datasets import FileSourceDataset, FileDataSource
+from os.path import join, expanduser
+import random
+
+# import global hyper parameters
+from hparams import hparams
+from modules import frontend
+import builder
+
+_frontend = getattr(frontend, hparams.frontend)
+
+
+def _pad(seq, max_len, constant_values=0):
+    return np.pad(seq, (0, max_len - len(seq)),
+                  mode="constant",
+                  constant_values=constant_values)
+
+
+def _pad_2d(x, max_len, b_pad=0):
+    x = np.pad(x, [(b_pad, max_len - len(x) - b_pad), (0, 0)],
+               mode="constant",
+               constant_values=0)
+    return x
+
+
+class TextDataSource(FileDataSource):
+    def __init__(self, data_root, speaker_id=None):
+        self.data_root = data_root
+        self.speaker_ids = None
+        self.multi_speaker = False
+        # If not None, filter by speaker_id
+        self.speaker_id = speaker_id
+
+    def collect_files(self):
+        meta = join(self.data_root, "train.txt")
+        with io.open(meta, "rt", encoding="utf-8") as f:
+            lines = f.readlines()
+        l = lines[0].split("|")
+        assert len(l) == 4 or len(l) == 5
+        self.multi_speaker = len(l) == 5
+        texts = list(map(lambda l: l.split("|")[3], lines))
+        if self.multi_speaker:
+            speaker_ids = list(map(lambda l: int(l.split("|")[-1]), lines))
+            # Filter by speaker_id
+            # using multi-speaker dataset as a single speaker dataset
+            if self.speaker_id is not None:
+                indices = np.array(speaker_ids) == self.speaker_id
+                texts = list(np.array(texts)[indices])
+                self.multi_speaker = False
+                return texts
+
+            return texts, speaker_ids
+        else:
+            return texts
+
+    def collect_features(self, *args):
+        if self.multi_speaker:
+            text, speaker_id = args
+        else:
+            text = args[0]
+        global _frontend
+        if _frontend is None:
+            _frontend = getattr(frontend, hparams.frontend)
+        seq = _frontend.text_to_sequence(
+            text, p=hparams.replace_pronunciation_prob)
+
+        if platform.system() == "Windows":
+            if hasattr(hparams, "gc_probability"):
+                _frontend = None  # memory leaking prevention in Windows
+                if np.random.rand() < hparams.gc_probability:
+                    gc.collect()  # garbage collection enforced
+                    print("GC done")
+
+        if self.multi_speaker:
+            return np.asarray(seq, dtype=np.int32), int(speaker_id)
+        else:
+            return np.asarray(seq, dtype=np.int32)
+
+
+class _NPYDataSource(FileDataSource):
+    def __init__(self, data_root, col, speaker_id=None):
+        self.data_root = data_root
+        self.col = col
+        self.frame_lengths = []
+        self.speaker_id = speaker_id
+
+    def collect_files(self):
+        meta = join(self.data_root, "train.txt")
+        with io.open(meta, "rt", encoding="utf-8") as f:
+            lines = f.readlines()
+        l = lines[0].split("|")
+        assert len(l) == 4 or len(l) == 5
+        multi_speaker = len(l) == 5
+        self.frame_lengths = list(map(lambda l: int(l.split("|")[2]), lines))
+
+        paths = list(map(lambda l: l.split("|")[self.col], lines))
+        paths = list(map(lambda f: join(self.data_root, f), paths))
+
+        if multi_speaker and self.speaker_id is not None:
+            speaker_ids = list(map(lambda l: int(l.split("|")[-1]), lines))
+            # Filter by speaker_id
+            # using multi-speaker dataset as a single speaker dataset
+            indices = np.array(speaker_ids) == self.speaker_id
+            paths = list(np.array(paths)[indices])
+            self.frame_lengths = list(np.array(self.frame_lengths)[indices])
+            # aha, need to cast numpy.int64 to int
+            self.frame_lengths = list(map(int, self.frame_lengths))
+
+        return paths
+
+    def collect_features(self, path):
+        return np.load(path)
+
+
+class MelSpecDataSource(_NPYDataSource):
+    def __init__(self, data_root, speaker_id=None):
+        super(MelSpecDataSource, self).__init__(data_root, 1, speaker_id)
+
+
+class LinearSpecDataSource(_NPYDataSource):
+    def __init__(self, data_root, speaker_id=None):
+        super(LinearSpecDataSource, self).__init__(data_root, 0, speaker_id)
+
+
+class PartialyRandomizedSimilarTimeLengthSampler(object):
+    """Partially randmoized sampler
+
+    1. Sort by lengths
+    2. Pick a small patch and randomize it
+    3. Permutate mini-batchs
+    """
+
+    def __init__(self,
+                 lengths,
+                 batch_size=16,
+                 batch_group_size=None,
+                 permutate=True):
+        self.sorted_indices = np.argsort(lengths)
+        self.lengths = np.array(lengths)[self.sorted_indices]
+        self.batch_size = batch_size
+        if batch_group_size is None:
+            batch_group_size = min(batch_size * 32, len(self.lengths))
+            if batch_group_size % batch_size != 0:
+                batch_group_size -= batch_group_size % batch_size
+
+        self.batch_group_size = batch_group_size
+        assert batch_group_size % batch_size == 0
+        self.permutate = permutate
+
+    def __iter__(self):
+        indices = self.sorted_indices.copy()
+        batch_group_size = self.batch_group_size
+        s, e = 0, 0
+        for i in range(len(indices) // batch_group_size):
+            s = i * batch_group_size
+            e = s + batch_group_size
+            random.shuffle(indices[s:e])
+
+        # Permutate batches
+        if self.permutate:
+            perm = np.arange(len(indices[:e]) // self.batch_size)
+            random.shuffle(perm)
+            indices[:e] = indices[:e].reshape(
+                -1, self.batch_size)[perm, :].reshape(-1)
+
+        # Handle last elements
+        s += batch_group_size
+        if s < len(indices):
+            random.shuffle(indices[s:])
+
+        return iter(indices)
+
+    def __len__(self):
+        return len(self.sorted_indices)
+
+
+class Dataset(object):
+    def __init__(self, X, Mel, Y):
+        self.X = X
+        self.Mel = Mel
+        self.Y = Y
+        # alias
+        self.multi_speaker = X.file_data_source.multi_speaker
+
+    def __getitem__(self, idx):
+        if self.multi_speaker:
+            text, speaker_id = self.X[idx]
+            return text, self.Mel[idx], self.Y[idx], speaker_id
+        else:
+            return self.X[idx], self.Mel[idx], self.Y[idx]
+
+    def __len__(self):
+        return len(self.X)
+
+
+def make_loader(dataset, batch_size, shuffle, sampler, create_batch_fn,
+                trainer_count, local_rank):
+    assert not (
+        shuffle and
+        sampler), "shuffle and sampler should not be valid in the same time."
+    num_samples = len(dataset)
+
+    def wrapper():
+        if sampler is None:
+            ids = range(num_samples)
+            if shuffle:
+                random.shuffle(ids)
+        else:
+            ids = sampler
+        batch, batches = [], []
+        for idx in ids:
+            batch.append(dataset[idx])
+            if len(batch) >= batch_size:
+                batches.append(batch)
+                batch = []
+            if len(batches) >= trainer_count:
+                yield create_batch_fn(batches[local_rank])
+                batches = []
+
+        if len(batch) > 0:
+            batches.append(batch)
+        if len(batches) >= trainer_count:
+            yield create_batch_fn(batches[local_rank])
+
+    return wrapper
+
+
+def create_batch(batch):
+    """Create batch"""
+    r = hparams.outputs_per_step
+    downsample_step = hparams.downsample_step
+    multi_speaker = len(batch[0]) == 4
+
+    # Lengths
+    input_lengths = [len(x[0]) for x in batch]
+    max_input_len = max(input_lengths)
+    input_lengths = np.array(input_lengths, dtype=np.int64)
+
+    target_lengths = [len(x[1]) for x in batch]
+
+    max_target_len = max(target_lengths)
+    target_lengths = np.array(target_lengths, dtype=np.int64)
+
+    if max_target_len % (r * downsample_step) != 0:
+        max_target_len += (r * downsample_step) - max_target_len % (
+            r * downsample_step)
+        assert max_target_len % (r * downsample_step) == 0
+
+    # Set 0 for zero beginning padding
+    # imitates initial decoder states
+    b_pad = r
+    max_target_len += b_pad * downsample_step
+
+    x_batch = np.array(
+        [_pad(x[0], max_input_len) for x in batch], dtype=np.int64)
+    x_batch = np.expand_dims(x_batch, axis=-1)
+
+    mel_batch = np.array(
+        [_pad_2d(
+            x[1], max_target_len, b_pad=b_pad) for x in batch],
+        dtype=np.float32)
+
+    # down sampling is done here
+    if downsample_step > 1:
+        mel_batch = mel_batch[:, 0::downsample_step, :]
+    mel_batch = np.expand_dims(np.transpose(mel_batch, axes=[0, 2, 1]), axis=2)
+
+    y_batch = np.array(
+        [_pad_2d(
+            x[2], max_target_len, b_pad=b_pad) for x in batch],
+        dtype=np.float32)
+    y_batch = np.expand_dims(np.transpose(y_batch, axes=[0, 2, 1]), axis=2)
+
+    # text positions
+    text_positions = np.array(
+        [_pad(np.arange(1, len(x[0]) + 1), max_input_len) for x in batch],
+        dtype=np.int64)
+    text_positions = np.expand_dims(text_positions, axis=-1)
+
+    max_decoder_target_len = max_target_len // r // downsample_step
+
+    # frame positions
+    s, e = 1, max_decoder_target_len + 1
+    frame_positions = np.tile(
+        np.expand_dims(
+            np.arange(
+                s, e, dtype=np.int64), axis=0), (len(batch), 1))
+    frame_positions = np.expand_dims(frame_positions, axis=-1)
+
+    # done flags
+    done = np.array([
+        _pad(
+            np.zeros(
+                len(x[1]) // r // downsample_step - 1, dtype=np.float32),
+            max_decoder_target_len,
+            constant_values=1) for x in batch
+    ])
+    done = np.expand_dims(np.expand_dims(done, axis=1), axis=1)
+
+    if multi_speaker:
+        speaker_ids = np.expand_dims(np.array([x[3] for x in batch]), axis=-1)
+        return (x_batch, input_lengths, mel_batch, y_batch, text_positions,
+                frame_positions, done, target_lengths, speaker_ids)
+    else:
+        speaker_ids = None
+        return (x_batch, input_lengths, mel_batch, y_batch, text_positions,
+                frame_positions, done, target_lengths)

deepvoice3/README.md

+# Deep Voice 3 with Paddle Fluid
+
+[中文版](README_cn.md)
+
+Paddle fluid implementation of DeepVoice 3, a convolutional network based text-to-speech synthesis model. The implementation is based on [Deep Voice 3: Scaling Text-to-Speech with Convolutional Sequence Learning](https://arxiv.org/abs/1710.07654).
+
+We implement Deepvoice3 model in paddle fluid with dynamic graph, which is convenient for flexible network architectures.
+
+## Installation
+
+You additionally need to download punkt and cmudict for nltk, because we tokenize text with `punkt` and convert text into phonemes with `cmudict`.
+
+```python
+import nltk
+nltk.download("punkt")
+nltk.download("cmudict")
+```
+
+## Model Architecture
+
+![DeepVoice3 model architecture](./_images/model_architecture.png)
+
+The model consists of an encoder, a decoder and a converter (and a speaker embedding for multispeaker models). The encoder, together with the decoder forms the seq2seq part of the model, and the converter forms the postnet part.
+
+## Project Structure
+
+```text
+├── audio.py                        # audio processing
+├── compute_timestamp_ratio.py      # script to compute position rate
+├── conversion                      # parameter conversion from pytorch model
+├── requirements.txt                # requirements
+├── hparams.py                      # HParam class for deepvoice3
+├── hparam_tf                       # hyper parameter related stuffs
+├── ljspeech.py                     # functions for ljspeech preprocessing
+├── preprocess.py                   # preprocrssing script
+├── presets                         # preset hyperparameters
+├── deepvoice3_paddle               # DeepVoice3 model implementation
+├── eval_model.py                   # functions for model evaluation
+├── synthesis.py                    # script for speech synthesis
+├── train_model.py                  # functions for model training
+└── train.py                        # script for model training
+```
+
+## Usage
+
+There are many hyperparameters to be tuned depending on the specification of model and dataset you are working on. Hyperparameters that are known to work good are provided in the repository. See `presets` directory for details. Now we only provide preset with LJSpeech dataset (`deepvoice3_ljspeech.json`). Support for more models and datasets is pending.
+
+Note that `preprocess.py`, `train.py` and `synthesis.py` all accept a `--preset` parameter. To ensure consistency, you should use the same preset for preprocessing, training and synthesizing.
+
+Note that you can overwrite preset hyperparameters with command line argument `--hparams`, just pass several key-value pair in  `${key}=${value}` format seperated by comma （`,`）. For example `--hparams="batch_size=8, nepochs=500"` can overwrite default values in the preset json file.
+
+Some hyperparameters are only related to training, like `batch_size`, `checkpoint_interval` and you can use different values for preprocessing and training. But hyperparameters related to data preprocessing, like `num_mels` and `ref_level_db`, should be kept the same for preprocessing and training.
+
+For more details about hyperparameters, see `hparams.py`, which contains the definition of `hparams`. Priority order of hyperparameters is command line option `--hparams` > `--preset` json configuration file > definition of hparams in `hparams.py`.
+
+### Dataset
+
+Download and unzip [LJSpeech](https://keithito.com/LJ-Speech-Dataset/).
+
+```bash
+wget https://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2
+tar xjvf LJSpeech-1.1.tar.bz2
+```
+
+Preprocessing with `preprocess.py`.
+
+```bash
+python preprocess.py \
+    --preset=${preset_json_path} \
+    --hparams="hyper parameters you want to overwrite" \
+    ${name} ${in_dir} ${out_dir}
+```
+
+Now `${name}$` only supports `ljspeech`. Support for other datasets is pending.
+
+Assuming that you use `presers/deepvoice3_ljspeech.json` for LJSpeech and the path of the unziped dataset is `./data/LJSpeech-1.1`, then you can preprocess data with the following command.
+
+```bash
+python preprocess.py \
+    --preset=presets/deepvoice3_ljspeech.json \
+    ljspeech ./data/LJSpeech-1.1/ ./data/ljspeech
+```
+
+When this is done, you will see extracted features in `./data/ljspeech` including:
+
+1. text and corresponding file names for the extracted features in `train.txt`.
+2. mel-spectrogram in `ljspeech-mel-*.npy` .
+3. linear-spectrogram in `ljspeech-spec-*.npy`.
+
+### Train on single GPU
+
+Training the whole model on one single GPU:
+
+```bash
+export PYTHONPATH=../:$PYTHONPATH
+export CUDA_VISIBLE_DEVICES=0
+python train.py --data-root=${data-root} --use-gpu \
+    --preset=${preset_json_path} \
+    --hparams="parameters you may want to override"
+```
+
+For more details about `train.py`, see `python train.py --help`.
+
+#### load checkpoints
+
+We provide a trained model ([dv3.single_frame](https://paddlespeech.bj.bcebos.com/Parakeet/dv3.single_frame.tar.gz)) for downloading, which is trained with the default preset. Unzip the downloaded file with `tar xzvf dv3.single_frame.tar.gz`, you will get `config.json`, `model.pdparams` and `README.md`. `config.json` is the preset json with which the model is trained, `model.pdparams` is the parameter file, and `README.md` is a brief introduction of the model.
+
+You can load saved checkpoint and resume training with `--checkpoint` (You only need to provide the base name of the parameter file, eg. if you want to load `model.pdparams`, just use `--checkpoint=model`). If there is also a file with the same basename and extension name `.pdopt` in the same folder with the model file (i.e. `model.pdopt`, which is the optimizer file), it is also loaded automatically. If you wan to reset optimizer states, pass `--reset-optimizer` in addition.
+
+
+#### train a part of the model
+
+You can also train parts of the model while freezing other parts, by passing `--train-seq2seq-only` or `--train-postnet-only`. When training only parts of the model, other parts should be loaded from saved checkpoint.
+
+To train only the `seq2seq` or `postnet`, you should load from a whole model  with `--checkpoint` and keep the same configurations with which the checkpoint is trained. Note that when training only the `postnet`, you should set `use_decoder_state_for_postnet_input=false`, because when train only the postnet, the postnet takes the ground truth mel-spectrogram as input. Note that the default value for `use_decoder_state_for_postnet_input` is `True`.
+
+example:
+
+```bash
+export CUDA_VISIBLE_DEVICES=0
+python train.py --data-root=${data-root} --use-gpu \
+    --preset=${preset_json_path} \
+    --hparams="parameters you may want to override" \
+    --train-seq2seq-only \
+    --output=${directory_to_save_results}
+```
+
+### Training on multiple GPUs
+
+Training on multiple GPUs with data parallel is enabled. You can run `train.py` with `paddle.distributed.launch` module. Here is the command line usage.
+
+```bash
+python -m paddle.distributed.launch \
+    --started_port ${port_of_the_first_worker} \
+    --selected_gpus ${logical_gpu_ids_to_choose} \
+    --log_dir ${path_of_write_log} \
+    training_script ...
+```
+
+`paddle.distributed.launch` parallelizes training in multiprocessing mode.`--selected_gpus` means the logical ids of the selected GPUs, and `started_port` means the port used by the first worker.  Outputs of each process are saved in `--log_dir.` Then follows the command for training on a single GPU, except that you should pass `--use-data-paralle` in addition.
+
+```bash
+export CUDA_VISIBLE_DEVICES=2,3,4,5    # The IDs of visible physical devices
+python -m paddle.distributed.launch \
+    --selected_gpus=0,1,2,3 --log_dir ${multi_gpu_log_dir} \
+    train.py --data-root=${data-root} \
+    --use-gpu --use-data-parallel \
+    --preset=${preset_json_path} \
+    --hparams="parameters you may want to override"
+```
+
+In the example above, we set only GPU `2, 3, 4, 5` to be visible. Then `--selected_gpus="0, 1, 2, 3"` means the logical ids of the selected gpus, which correponds to GPU `2, 3, 4, 5`.
+
+Model checkpoints (`*.pdparams` for the model and `*.pdopt` for the optimizer) are saved in `${directory_to_save_results}/checkpoints` per 10000 steps by default. Layer-wise averaged attention alignments (.png) are saved in `${directory_to_save_results}/checkpoints/alignment_ave`. And alignments for each attention layer are saved in `${directory_to_save_results}/checkpoints/alignment_layer{attention_layer_num}` per 10000 steps for inspection.
+
+Synthesis results of 6 sentences (hardcoded in `eval_model.py`) are saved in `${directory_to_save_results}/checkpoints/eval`, including  `step{step_num}_text{text_id}_single_alignment.png` for averaged alignments and `step{step_num}_text{text_id}_single_predicted.wav` for the predicted waveforms.
+
+
+### Monitor with Tensorboard
+
+Logs with tensorboard are saved in `${directory_to_save_results}/log/`  directory by default. You can monitor logs by tensorboard.
+
+```bash
+tensorboard --logdir=${log_dir} --host=$HOSTNAME --port=8888
+```
+
+### Synthesize from a checkpoint
+
+Given a list of text, `synthesis.py` synthesize audio signals from a trained model.
+
+```bash
+python synthesis.py --use-gpu --preset=${preset_json_path} \
+    --hparams="parameters you may want to override" \
+      ${checkpoint} ${text_list_file} ${dst_dir}
+```
+
+Example test_list.txt:
+
+```text
+Generative adversarial network or variational auto-encoder.
+Once upon a time there was a dear little girl who was loved by every one who looked at her, but most of all by her grandmother, and there was nothing that she would not have given to the child.
+A text-to-speech synthesis system typically consists of multiple stages, such as a text analysis frontend, an acoustic model and an audio synthesis module.
+```
+
+generated waveform files and alignment files are saved in `${dst_dir}`.
+
+### Compute position ratio
+
+According to [Deep Voice 3: Scaling Text-to-Speech with Convolutional Sequence Learning](https://arxiv.org/abs/1710.07654), the position rate is different for different datasets. There are 2 position rates, one for the query and the other for the key, which are referred to as $\omega_1$ and $\omega_2$ in th paper, and the corresponding names in preset json are `query_position_rate` and `key_position_rate`.
+
+For example, the `query_position_rate` and `key_position_rate` for LJSpeech are `1.0` and `1.385`, respectively. Fix the `query_position_rate` as 1.0, the `key_position_rate` is computed with `compute_timestamp_ratio.py`. Run the command below, where `${data_root}` means the path of the preprocessed dataset.
+
+```bash
+python compute_timestamp_ratio.py --preset=${preset_json_path} \
+    --hparams="parameters you may want to override" ${data_root}
+```
+
+You will get outputs like this.
+
+```text
+100%|██████████████████████████████████████████████████████████| 13047/13047 [00:12<00:00, 1058.19it/s]
+1345587 1863884.0 1.3851828235558161
+```
+
+Then set the `key_position_rate=1.385` and `query_position_rate=1.0` in the preset.
+
+## Acknowledgement
+
+We thankfully included and adapted some files from r9y9's [deepvoice3_pytorch](https://github.com/r9y9/deepvoice3_pytorch).

deepvoice3/README_cn.md

+# Deep Voice 3 with Paddle Fluid
+
+[English](README.md)
+
+Paddle 实现的 Deepvoice3，一个基于卷积神经网络的语音合成 (Text to Speech) 模型。本实现基于 [Deep Voice 3: Scaling Text-to-Speech with Convolutional Sequence Learning](https://arxiv.org/abs/1710.07654) 。
+
+本 Deepvoice3 实现使用 Paddle 动态图模式，这对于灵活的网络结构更为方便。
+
+## 安装
+
+### 安装 paddlepaddle 框架
+
+为了更快的训练速度和更好的支持，我们推荐使用最新的开发版 paddle。用户可以最新编译的开发版 whl 包，也可以选择从源码编译 Paddle。
+
+1. 下载最新编译的开发版 whl 包。可以从  [**多版本 wheel 包列表-dev**](https://www.paddlepaddle.org.cn/documentation/docs/zh/beginners_guide/install/Tables.html#whl-dev) 页面中选择合适的版本。
+
+2. 从源码编译 Paddle. 参考[**从源码编译**](https://www.paddlepaddle.org.cn/documentation/docs/zh/beginners_guide/install/compile/fromsource.html) 页面。注意，如果你需要使用多卡训练，那么编译前需要设置选项 `-DWITH_DISTRIBUTE=ON`。
+
+### 其他依赖
+
+使用 pip 安装其他依赖。
+
+```bash
+pip install -r requirements.txt
+```
+
+另外需要下载 nltk 的两个库，因为使用了 `punkt` 对文本进行 tokenization，并且使用了 `cmudict` 来将文本转为音位。
+
+```python
+import nltk
+nltk.download("punkt")
+nltk.download("cmudict")
+```
+
+## 模型结构
+
+![DeepVoice3 模型结构](./_images/model_architecture.png)
+
+模型包含 encoder, decoder, converter 几个部分，对于 multispeaker 数据集，还有一个 speaker embedding。其中 encoder 和 decoder 构成 seq2seq 部分，converter 构成 postnet 部分。
+
+## 项目结构
+
+```text
+├── audio.py                        # 用于处理处理音频的函数
+├── compute_timestamp_ratio.py      # 计算 position rate 的脚本
+├── conversion                      # 用于转换 pytorch 实现的参数
+├── requirements.txt                # 项目依赖
+├── hparams.py                      # DeepVoice3 运行超参数配置类的定义
+├── hparam_tf                       # 超参数相关
+├── ljspeech.py                     # ljspeech 数据集预处理
+├── preprocess.py                   # 通用预处理脚本
+├── presets                         # 预设超参数配置
+├── deepvoice3_paddle               # DeepVoice3 模型实现的主要文件
+├── eval_model.py                   # 模型测评相关函数
+├── synthesis.py                    # 用于语音合成的脚本
+├── train_model.py                  # 模型训练相关函数
+└── train.py                        # 用于模型训练的脚本
+```
+
+## 使用方法
+
+根据所使用的模型配置和数据集的不同，有不少超参数需要进行调节。我们提供已知结果较好的超参数设置，详见 `presets` 文件夹。目前我们只提供 LJSpeech 的预设配置 （`deepvoice3_ljspeech.json`）。后续将提供更多模型和数据集的预设配置。
+
+`preprocess.py`，`train.py`，`synthesis.py` 都接受 `--preset` 参数。为了保持一致性，最好在数据预处理，模型训练和语音合成时使用相同的预设配置。
+
+可以通过 `--hparams` 参数来覆盖预设的超参数配置，参数格式是逗号分隔的键值对 `${key}=${value}`，例如 `--hparams="batch_size=8, nepochs=500"`。
+
+部分参数只和训练有关，如 `batch_size`, `checkpoint_interval`, 用户在训练时可以使用不同的值。但部分参数和数据预处理相关，如 `num_mels` 和 `ref_level_db`, 这些参数在数据预处理和训练时候应该保持一致。
+
+关于超参数设置更多细节可以参考 `hparams.py` ，其中定义了 hparams。超参数的优先级序列是：通过命令行参数 `--hparams` 传入的参数优先级高于通过 `--preset` 参数传入的 json 配置文件，高于 `hparams.py` 中的定义。
+
+### 数据集
+
+下载并解压 [LJSpeech](https://keithito.com/LJ-Speech-Dataset/) 数据集。
+
+```bash
+wget https://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2
+tar xjvf LJSpeech-1.1.tar.bz2
+```
+
+使用 `preprocess.py`进行预处理。
+
+```bash
+python preprocess.py \
+    --preset=${preset_json_path} \
+    --hparams="hyper parameters you want to overwrite" \
+    ${name} ${in_dir} ${out_dir}
+```
+
+目前 `${name}$` 只支持 `ljspeech`。未来将会支持更多数据集。
+
+假设你使用 `presers/deepvoice3_ljspeech.json` 作为处理 LJSpeech 的预设配置文件，并且解压后的数据集位于  `./data/LJSpeech-1.1`, 那么使用如下的命令进行数据预处理。
+
+```bash
+python preprocess.py \
+    --preset=presets/deepvoice3_ljspeech.json \
+    ljspeech ./data/LJSpeech-1.1/ ./data/ljspeech
+```
+
+数据处理完成后，你会在 `./data/ljspeech` 看到提取的特征，包含如下文件。
+
+1. `train.txt`，包含文本和对应的音频特征的文件名。
+2. `ljspeech-mel-*.npy`，包含 mel 频谱。
+3. `ljspeech-spec-*.npy`，包含线性频谱。
+
+### 使用 GPU 单卡训练
+
+在单个 GPU 上训练整个模型的使用方法如下。
+
+```bash
+export CUDA_VISIBLE_DEVICES=0
+python train.py --data-root=${data-root} --use-gpu \
+    --preset=${preset_json_path} \
+    --hparams="parameters you may want to override"
+```
+
+用于可以通过 `python train.py --help` 查看 `train.py` 的详细使用方法。
+
+#### 加载保存的模型
+
+我们提供了使用默认的配置文件训练的模型 [dv3.single_frame](https://paddlespeech.bj.bcebos.com/Parakeet/dv3.single_frame.tar.gz) 供用户下载。使用 `tar xzvf dv3.single_frame.tar.gz` 解压下载的文件，会得到 `config.json`, `model.pdparams` and `README.md`。其中 `config.json` 是模型训练时使用的配置文件，`model.pdparams` 是参数文件，`README.md` 是模型的简要说明。
+
+用户可以通过 `--checkpoint` 参数加载保存的模型并恢复训练（注意：只需要传基础文件名，不需要扩展名，例如需要加载 `model.pdparams` 那么，只需要使用 `--checkpoint=model`）。如果同一个文件夹内有一个和参数文件基础文件名相同，而后缀为 `.pdopt` 的文件，（如 `model.pdopt`，即优化器文件），那么该文件也会被自动加载。如果你想要重置优化器的状态，在训练脚本加入 `--reset-optimizer` 参数。
+
+#### 训练模型的一部分
+
+用户可以通过 `--train-seq2seq-only` 或者 `--train-postnet-only` 来实现固定模型的其他部分，只训练需要训练的部分。但当只训练模型的一部分时，其他的部分需要从保存的模型中加载。
+
+当只训练模型的 `seq2seq` 部分或者 `postnet` 部分时，需要使用 `--checkpoint` 加载整个模型并保持相同的配置。注意，当只训练 `postnet` 的时候，需要保证配置中的`use_decoder_state_for_postnet_input=false`，因为在这种情况下，postnet 使用真实的 mel 频谱作为输入。注意，`use_decoder_state_for_postnet_input` 的默认值是 `True`。
+
+示例:
+
+```bash
+export CUDA_VISIBLE_DEVICES=0
+python train.py --data-root=${data-root} --use-gpu \
+    --preset=${preset_json_path} \
+    --hparams="parameters you may want to override" \
+    --train-seq2seq-only \
+    --output=${directory_to_save_results}
+```
+
+### 使用 GPU 多卡训练
+
+本模型支持使用多个 GPU 通过数据并行的方式训练。方法是使用 `paddle.distributed.launch` 模块来启动 `train.py`。
+
+```bash
+python -m paddle.distributed.launch \
+    --started_port ${port_of_the_first_worker} \
+    --selected_gpus ${logical_gpu_ids_to_choose} \
+    --log_dir ${path_to_write_log} \
+    training_script ...
+```
+
+paddle.distributed.launch 通过多进程的方式进行并行训练。`--selected_gpus` 指的是选择的 GPU 的逻辑序号，`started_port` 指的是 0 号显卡的使用的端口号，`--log_dir` 是日志保存的目录，每个进程的输出会在这个文件夹中保存为单独的文件。再在后面接上需要启动的脚本文件及其参数即可。这部分和单卡训练的脚本一致，但是需要传入 `--use-data-paralle` 以使用数据并行训练。示例命令如下。
+
+```bash
+export CUDA_VISIBLE_DEVICES=2,3,4,5    # The IDs of visible physical devices
+python -m paddle.distributed.launch \
+    --selected_gpus=0,1,2,3 --log_dir ${multi_gpu_log_dir} \
+    train.py --data-root=${data-root} \
+    --use-gpu --use-data-parallel \
+    --preset=${preset_json_path} \
+    --hparams="parameters you may want to override" \
+    --output=${directory_to_save_results}
+```
+
+上述的示例中，设置了 `2, 3, 4, 5` 号显卡为可见的 GPU。然后 `--selected_gpus=0,1,2,3` 选择的是 GPU 的逻辑序号，分别对应于  `2, 3, 4, 5` 号卡。
+
+模型 (模型参数保存为`*.pdparams` 文件，优化器被保存为 `*.pdopt` 文件)保存在 `${directory_to_save_results}/checkpoints` 文件夹中。多层平均的注意力机制对齐结果被保存为 `.png` 图片，默认保存在 `${directory_to_save_results}/checkpoints/alignment_ave` 中。每一层的注意力机制对齐结果默认被保存在 `${directory_to_save_results}/checkpoints/alignment_layer{attention_layer_num}`文件夹中。默认每 10000 步保存一次用于查看。
+
+对 6 个给定的句子的语音合成结果保存在 `${directory_to_save_results}/checkpoints/eval` 中，包含多层平均平均的注意力机制对齐结果，这被保存为名为  `step{step_num}_text{text_id}_single_alignment.png` 的图片；以及合成的音频文件，保存为名为 `step{step_num}_text{text_id}_single_predicted.wav` 的音频。
+
+
+### 使用 Tensorboard 查看训练
+
+Tensorboard 训练日志被保存在 `${directory_to_save_results}/log/` 文件夹，可以通过 tensorboard 查看。使用方法如下。
+
+```bash
+tensorboard --logdir=${log_dir} --host=$HOSTNAME --port=8888
+```
+
+### 从保存的模型合成语音
+
+给定一组文本，使用  `synthesis.py` 从一个训练好的模型来合成语音，使用方法如下。
+
+```bash
+python synthesis.py --use-gpu --preset=${preset_json_path} \
+    --hparams="parameters you may want to override" \
+      ${checkpoint} ${text_list_file} ${dst_dir}
+```
+
+示例文本文件如下:
+
+```text
+Generative adversarial network or variational auto-encoder.
+Once upon a time there was a dear little girl who was loved by every one who looked at her, but most of all by her grandmother, and there was nothing that she would not have given to the child.
+A text-to-speech synthesis system typically consists of multiple stages, such as a text analysis frontend, an acoustic model and an audio synthesis module.
+```
+
+合成的结果包含注意力机制对齐结果和音频文件，保存于 `${dst_dir}`。
+
+### 计算 position rate
+
+根据 [Deep Voice 3: Scaling Text-to-Speech with Convolutional Sequence Learning](https://arxiv.org/abs/1710.07654), 对于不同的数据集，会有不同的 position rate. 有两个不同的 position rate，一个用于 query 一个用于 key， 这在论文中称为 $\omega_1$ 和 $\omega_2$ ，在预设配置文件中的名字分别为 `query_position_rate` 和 `key_position_rate`。
+
+比如 LJSpeech 数据集的 `query_position_rate` 和 `key_position_rate` 分别为 `1.0` 和 `1.385`。固定 `query_position_rate` 为 1.0，`key_position_rate` 可以使用 `compute_timestamp_ratio.py` 计算，命令如下，其中 `${data_root}` 是预处理后的数据集路径。
+
+```bash
+python compute_timestamp_ratio.py --preset=${preset_json_path} \
+    --hparams="parameters you may want to override" ${data_root}
+```
+
+可以得到如下的结果。
+
+```text
+100%|██████████████████████████████████████████████████████████| 13047/13047 [00:12<00:00, 1058.19it/s]
+1345587 1863884.0 1.3851828235558161
+```
+
+然后在预设配置文件中设置 `key_position_rate=1.385` 以及 `query_position_rate=1.0`。
+
+## 致谢
+
+本实现包含及改写了 r9y9's 的 [deepvoice3_pytorch](https://github.com/r9y9/deepvoice3_pytorch) 中的部分文件，在此表示感谢。

deepvoice3/_ce.py

+# this file is only used for continuous evaluation test!
+
+import os
+import sys
+sys.path.append(os.environ['ceroot'])
+from kpi import CostKpi
+from kpi import DurationKpi
+from kpi import AccKpi
+
+each_epoch_duration_frame1_card1 = DurationKpi("each_epoch_duration_frame1_card1", 0.02, actived=True)
+train_cost_frame1_card1 = CostKpi("train_cost_frame1_card1", 0.02, actived=True)
+each_epoch_duration_frame4_card1 = DurationKpi("each_epoch_duration_frame4_card1", 0.05, actived=True)
+train_cost_frame4_card1 = CostKpi("train_cost_frame4_card1", 0.02, actived=True)
+
+
+tracking_kpis = [
+    each_epoch_duration_frame1_card1,
+    train_cost_frame1_card1,
+    each_epoch_duration_frame4_card1,
+    train_cost_frame4_card1,
+]
+
+
+def parse_log(log):
+    '''
+    This method should be implemented by model developers.
+
+    The suggestion:
+
+    each line in the log should be key, value, for example:
+
+    "
+    train_cost\t1.0
+    test_cost\t1.0
+    train_cost\t1.0
+    train_cost\t1.0
+    train_acc\t1.2
+    "
+    '''
+    for line in log.split('\n'):
+        fs = line.strip().split('\t')
+        print(fs)
+        if len(fs) == 3 and fs[0] == 'kpis':
+            kpi_name = fs[1]
+            kpi_value = float(fs[2])
+            yield kpi_name, kpi_value
+
+
+def log_to_ce(log):
+    kpi_tracker = {}
+    for kpi in tracking_kpis:
+        kpi_tracker[kpi.name] = kpi
+
+    for (kpi_name, kpi_value) in parse_log(log):
+        print(kpi_name, kpi_value)
+        kpi_tracker[kpi_name].add_record(kpi_value)
+        kpi_tracker[kpi_name].persist()
+
+
+if __name__ == '__main__':
+    log = sys.stdin.read()
+    log_to_ce(log)

deepvoice3/audio.py

+# This file was copied from https://github.com/r9y9/deepvoice3_pytorch/tree/master/audio.py
+# Copyright (c) 2017: Ryuichi Yamamoto.
+
+import librosa
+import librosa.filters
+import math
+import numpy as np
+from scipy import signal
+from hparams import hparams
+from scipy.io import wavfile
+
+import lws
+
+
+def load_wav(path):
+    return librosa.core.load(path, sr=hparams.sample_rate)[0]
+
+
+def save_wav(wav, path):
+    wav = wav * 32767 / max(0.01, np.max(np.abs(wav)))
+    wavfile.write(path, hparams.sample_rate, wav.astype(np.int16))
+
+
+def preemphasis(x):
+    from nnmnkwii.preprocessing import preemphasis
+    return preemphasis(x, hparams.preemphasis)
+
+
+def inv_preemphasis(x):
+    from nnmnkwii.preprocessing import inv_preemphasis
+    return inv_preemphasis(x, hparams.preemphasis)
+
+
+def spectrogram(y):
+    D = _lws_processor().stft(preemphasis(y)).T
+    S = _amp_to_db(np.abs(D)) - hparams.ref_level_db
+    return _normalize(S)
+
+
+def inv_spectrogram(spectrogram):
+    '''Converts spectrogram to waveform using librosa'''
+    S = _db_to_amp(_denormalize(spectrogram) +
+                   hparams.ref_level_db)  # Convert back to linear
+    processor = _lws_processor()
+    D = processor.run_lws(S.astype(np.float64).T**hparams.power)
+    y = processor.istft(D).astype(np.float32)
+    return inv_preemphasis(y)
+
+
+def melspectrogram(y):
+    D = _lws_processor().stft(preemphasis(y)).T
+    S = _amp_to_db(_linear_to_mel(np.abs(D))) - hparams.ref_level_db
+    if not hparams.allow_clipping_in_normalization:
+        assert S.max() <= 0 and S.min() - hparams.min_level_db >= 0
+    return _normalize(S)
+
+
+def _lws_processor():
+    return lws.lws(hparams.fft_size, hparams.hop_size, mode="speech")
+
+
+# Conversions:
+
+_mel_basis = None
+
+
+def _linear_to_mel(spectrogram):
+    global _mel_basis
+    if _mel_basis is None:
+        _mel_basis = _build_mel_basis()
+    return np.dot(_mel_basis, spectrogram)
+
+
+def _build_mel_basis():
+    if hparams.fmax is not None:
+        assert hparams.fmax <= hparams.sample_rate // 2
+    return librosa.filters.mel(hparams.sample_rate,
+                               hparams.fft_size,
+                               fmin=hparams.fmin,
+                               fmax=hparams.fmax,
+                               n_mels=hparams.num_mels)
+
+
+def _amp_to_db(x):
+    min_level = np.exp(hparams.min_level_db / 20 * np.log(10))
+    return 20 * np.log10(np.maximum(min_level, x))
+
+
+def _db_to_amp(x):
+    return np.power(10.0, x * 0.05)
+
+
+def _normalize(S):
+    return np.clip((S - hparams.min_level_db) / -hparams.min_level_db, 0, 1)
+
+
+def _denormalize(S):
+    return (np.clip(S, 0, 1) * -hparams.min_level_db) + hparams.min_level_db

deepvoice3/builder.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from deepvoice3 import DeepVoiceTTS, ConvSpec, WindowRange
+
+
+def deepvoice3(n_vocab,
+               embed_dim=256,
+               mel_dim=80,
+               linear_dim=513,
+               r=4,
+               downsample_step=1,
+               n_speakers=1,
+               speaker_dim=16,
+               padding_idx=0,
+               dropout=(1 - 0.96),
+               filter_size=5,
+               encoder_channels=128,
+               decoder_channels=256,
+               converter_channels=256,
+               query_position_rate=1.0,
+               key_position_rate=1.29,
+               use_memory_mask=False,
+               trainable_positional_encodings=False,
+               force_monotonic_attention=True,
+               use_decoder_state_for_postnet_input=True,
+               max_positions=512,
+               embedding_weight_std=0.1,
+               speaker_embedding_weight_std=0.01,
+               freeze_embedding=False,
+               window_range=WindowRange(-1, 3),
+               key_projection=False,
+               value_projection=False):
+    time_upsampling = max(downsample_step, 1)
+
+    h = encoder_channels
+    k = filter_size
+    encoder_convolutions = (ConvSpec(h, k, 1), ConvSpec(h, k, 3),
+                            ConvSpec(h, k, 9), ConvSpec(h, k, 27),
+                            ConvSpec(h, k, 1), ConvSpec(h, k, 3),
+                            ConvSpec(h, k, 9), ConvSpec(h, k, 27),
+                            ConvSpec(h, k, 1), ConvSpec(h, k, 3))
+
+    h = decoder_channels
+    prenet_convolutions = (ConvSpec(h, k, 1), ConvSpec(h, k, 3))
+    attentive_convolutions = (ConvSpec(h, k, 1), ConvSpec(h, k, 3),
+                              ConvSpec(h, k, 9), ConvSpec(h, k, 27),
+                              ConvSpec(h, k, 1))
+    attention = [True, False, False, False, True]
+
+    h = converter_channels
+    postnet_convolutions = (ConvSpec(h, k, 1), ConvSpec(h, k, 3),
+                            ConvSpec(2 * h, k, 1), ConvSpec(2 * h, k, 3))
+
+    model = DeepVoiceTTS(
+        "dv3", n_speakers, speaker_dim, speaker_embedding_weight_std, n_vocab,
+        embed_dim, padding_idx, embedding_weight_std, freeze_embedding,
+        encoder_convolutions, max_positions, padding_idx,
+        trainable_positional_encodings, mel_dim, r, prenet_convolutions,
+        attentive_convolutions, attention, use_memory_mask,
+        force_monotonic_attention, query_position_rate, key_position_rate,
+        window_range, key_projection, value_projection, linear_dim,
+        postnet_convolutions, time_upsampling, dropout,
+        use_decoder_state_for_postnet_input, "float32")
+    return model
+
+
+def deepvoice3_multispeaker(n_vocab,
+                            embed_dim=256,
+                            mel_dim=80,
+                            linear_dim=513,
+                            r=4,
+                            downsample_step=1,
+                            n_speakers=1,
+                            speaker_dim=16,
+                            padding_idx=0,
+                            dropout=(1 - 0.96),
+                            filter_size=5,
+                            encoder_channels=128,
+                            decoder_channels=256,
+                            converter_channels=256,
+                            query_position_rate=1.0,
+                            key_position_rate=1.29,
+                            use_memory_mask=False,
+                            trainable_positional_encodings=False,
+                            force_monotonic_attention=True,
+                            use_decoder_state_for_postnet_input=True,
+                            max_positions=512,
+                            embedding_weight_std=0.1,
+                            speaker_embedding_weight_std=0.01,
+                            freeze_embedding=False,
+                            window_range=WindowRange(-1, 3),
+                            key_projection=False,
+                            value_projection=False):
+    time_upsampling = max(downsample_step, 1)
+
+    h = encoder_channels
+    k = filter_size
+    encoder_convolutions = (ConvSpec(h, k, 1), ConvSpec(h, k, 3),
+                            ConvSpec(h, k, 9), ConvSpec(h, k, 27),
+                            ConvSpec(h, k, 1), ConvSpec(h, k, 3),
+                            ConvSpec(h, k, 9), ConvSpec(h, k, 27),
+                            ConvSpec(h, k, 1), ConvSpec(h, k, 3))
+
+    h = decoder_channels
+    prenet_convolutions = (ConvSpec(h, k, 1))
+    attentive_convolutions = (ConvSpec(h, k, 1), ConvSpec(h, k, 3),
+                              ConvSpec(h, k, 9), ConvSpec(h, k, 27),
+                              ConvSpec(h, k, 1))
+    attention = [True, False, False, False, False]
+
+    h = converter_channels
+    postnet_convolutions = (ConvSpec(h, k, 1), ConvSpec(h, k, 3),
+                            ConvSpec(2 * h, k, 1), ConvSpec(2 * h, k, 3))
+
+    model = DeepVoiceTTS(
+        "dv3", n_speakers, speaker_dim, speaker_embedding_weight_std, n_vocab,
+        embed_dim, padding_idx, embedding_weight_std, freeze_embedding,
+        encoder_convolutions, max_positions, padding_idx,
+        trainable_positional_encodings, mel_dim, r, prenet_convolutions,
+        attentive_convolutions, attention, use_memory_mask,
+        force_monotonic_attention, query_position_rate, key_position_rate,
+        window_range, key_projection, value_projection, linear_dim,
+        postnet_convolutions, time_upsampling, dropout,
+        use_decoder_state_for_postnet_input, "float32")
+    return model

deepvoice3/compute_timestamp_ratio.py

+# Part of code was adpated from https://github.com/r9y9/deepvoice3_pytorch/tree/master/compute_timestamp_ratio.py
+# Copyright (c) 2017: Ryuichi Yamamoto.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import sys
+import io
+import numpy as np
+sys.path.append("../")
+from hparams import hparams, hparams_debug_string
+from data.data import TextDataSource, MelSpecDataSource
+from nnmnkwii.datasets import FileSourceDataset
+from tqdm import trange
+from modules import frontend
+
+
+def build_parser():
+    parser = argparse.ArgumentParser(
+        description="Compute output/input timestamp ratio.")
+    parser.add_argument(
+        "--hparams", type=str, default="", help="Hyper parameters.")
+    parser.add_argument(
+        "--preset",
+        type=str,
+        required=True,
+        help="Path of preset parameters (json).")
+    parser.add_argument("data_root", type=str, help="path of the dataset.")
+    return parser
+
+
+if __name__ == "__main__":
+    parser = build_parser()
+    args, _ = parser.parse_known_args()
+
+    data_root = args.data_root
+    preset = args.preset
+
+    # Load preset if specified
+    if preset is not None:
+        with io.open(preset) as f:
+            hparams.parse_json(f.read())
+    # Override hyper parameters
+    hparams.parse(args.hparams)
+    assert hparams.name == "deepvoice3"
+
+    # Code below
+    X = FileSourceDataset(TextDataSource(data_root))
+    Mel = FileSourceDataset(MelSpecDataSource(data_root))
+
+    in_sizes = []
+    out_sizes = []
+    for i in trange(len(X)):
+        x, m = X[i], Mel[i]
+        if X.file_data_source.multi_speaker:
+            x = x[0]
+        in_sizes.append(x.shape[0])
+        out_sizes.append(m.shape[0])
+
+    in_sizes = np.array(in_sizes)
+    out_sizes = np.array(out_sizes)
+
+    input_timestamps = np.sum(in_sizes)
+    output_timestamps = np.sum(
+        out_sizes) / hparams.outputs_per_step / hparams.downsample_step
+
+    print(input_timestamps, output_timestamps,
+          output_timestamps / input_timestamps)
+    sys.exit(0)

deepvoice3/dry_run.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+from paddle import fluid
+import paddle.fluid.dygraph as dg
+
+from hparams import hparams, hparams_debug_string
+from modules import frontend
+from deepvoice3 import DeepVoiceTTS
+
+
+def dry_run(model):
+    """
+    Run the model once, just to get it initialized.
+    """
+    model.train()
+    _frontend = getattr(frontend, hparams.frontend)
+    batch_size = 4
+    enc_length = 157
+    snd_sample_length = 500
+
+    r = hparams.outputs_per_step
+    downsample_step = hparams.downsample_step
+    n_speakers = hparams.n_speakers
+
+    # make sure snd_sample_length can be divided by r * downsample_step
+    linear_shift = r * downsample_step
+    snd_sample_length += linear_shift - snd_sample_length % linear_shift
+    decoder_length = snd_sample_length // downsample_step // r
+    mel_length = snd_sample_length // downsample_step
+
+    n_vocab = _frontend.n_vocab
+    max_pos = hparams.max_positions
+    spker_embed = hparams.speaker_embed_dim
+    linear_dim = model.linear_dim
+    mel_dim = hparams.num_mels
+
+    x = np.random.randint(
+        low=0, high=n_vocab, size=(batch_size, enc_length, 1), dtype="int64")
+    input_lengths = np.arange(
+        enc_length - batch_size + 1, enc_length + 1, dtype="int64")
+    mel = np.random.randn(batch_size, mel_dim, 1, mel_length).astype("float32")
+    y = np.random.randn(batch_size, linear_dim, 1,
+                        snd_sample_length).astype("float32")
+
+    text_positions = np.tile(
+        np.arange(
+            0, enc_length, dtype="int64"), (batch_size, 1))
+    text_mask = text_positions > np.expand_dims(input_lengths, 1)
+    text_positions[text_mask] = 0
+    text_positions = np.expand_dims(text_positions, axis=-1)
+
+    frame_positions = np.tile(
+        np.arange(
+            1, decoder_length + 1, dtype="int64"), (batch_size, 1))
+    frame_positions = np.expand_dims(frame_positions, axis=-1)
+
+    done = np.zeros(shape=(batch_size, 1, 1, decoder_length), dtype="float32")
+    target_lengths = np.array([snd_sample_length] * batch_size).astype("int64")
+
+    speaker_ids = np.random.randint(
+        low=0, high=n_speakers, size=(batch_size, 1),
+        dtype="int64") if n_speakers > 1 else None
+
+    ismultispeaker = speaker_ids is not None
+
+    x = dg.to_variable(x)
+    input_lengths = dg.to_variable(input_lengths)
+    mel = dg.to_variable(mel)
+    y = dg.to_variable(y)
+    text_positions = dg.to_variable(text_positions)
+    frame_positions = dg.to_variable(frame_positions)
+    done = dg.to_variable(done)
+    target_lengths = dg.to_variable(target_lengths)
+    speaker_ids = dg.to_variable(
+        speaker_ids) if speaker_ids is not None else None
+
+    # these two fields are used as numpy ndarray
+    text_lengths = input_lengths.numpy()
+    decoder_lengths = target_lengths.numpy() // r // downsample_step
+
+    max_seq_len = max(text_lengths.max(), decoder_lengths.max())
+    if max_seq_len >= hparams.max_positions:
+        raise RuntimeError(
+            "max_seq_len ({}) >= max_posision ({})\n"
+            "Input text or decoder targget length exceeded the maximum length.\n"
+            "Please set a larger value for ``max_position`` in hyper parameters."
+            .format(max_seq_len, hparams.max_positions))
+
+    # cause paddle's embedding layer expect shape[-1] == 1
+
+    # first dry run runs the whole model
+    mel_outputs, linear_outputs, attn, done_hat = model(
+        x, input_lengths, mel, speaker_ids, text_positions, frame_positions)
+
+    num_parameters = 0
+    for k, v in model.state_dict().items():
+        print("{}|{}|{}".format(k, v.shape, np.prod(v.shape)))
+        num_parameters += np.prod(v.shape)
+    print("now model has {} parameters".format(len(model.state_dict())))
+    print("now model has {} elements".format(num_parameters))

deepvoice3/eval_model.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import sys
+import os
+from os.path import join, expanduser
+from warnings import warn
+from datetime import datetime
+
+import matplotlib
+# Force matplotlib not to use any Xwindows backend.
+matplotlib.use("Agg")
+from matplotlib import pyplot as plt
+from matplotlib import cm
+
+import audio
+import numpy as np
+from paddle import fluid
+import paddle.fluid.dygraph as dg
+import librosa.display
+from tensorboardX import SummaryWriter
+
+# import global hyper parameters
+from hparams import hparams
+from modules import frontend
+
+_frontend = getattr(frontend, hparams.frontend)
+
+
+def tts(model, text, p=0., speaker_id=None):
+    """
+    Convert text to speech waveform given a deepvoice3 model.
+
+    Args:
+        model (DeepVoiceTTS): Model used to synthesize waveform.
+        text (str) : Input text to be synthesized
+        p (float) : Replace word to pronounciation if p > 0. Default is 0.
+    
+    Returns:
+        waveform (numpy.ndarray): Shape(T_wav, ), predicted wave form, where
+            T_wav means the length of the synthesized wave form.
+        alignment (numpy.ndarray): Shape(T_dec, T_enc), predicted alignment
+            matrix, where T_dec means the time steps of decoder outputs, T_enc
+            means the time steps of encoder outoputs.
+        spectrogram (numpy.ndarray): Shape(T_lin, C_lin), predicted linear
+            spectrogram, where T__lin means the time steps of linear
+            spectrogram and C_lin mean sthe channels of linear spectrogram.
+        mel (numpy.ndarray): Shape(T_mel, C_mel), predicted mel spectrogram,
+            where T_mel means the time steps of mel spectrogram and C_mel means
+            the channels of mel spectrogram.
+    """
+    model.eval()
+
+    sequence = np.array(_frontend.text_to_sequence(text, p=p)).astype("int64")
+    sequence = np.reshape(sequence, (1, -1, 1))
+    text_positions = np.arange(1, sequence.shape[1] + 1, dtype="int64")
+    text_positions = np.reshape(text_positions, (1, -1, 1))
+
+    sequence = dg.to_variable(sequence)
+    text_positions = dg.to_variable(text_positions)
+    speaker_ids = None if speaker_id is None else fluid.layers.fill_constant(
+        shape=[1, 1], value=speaker_id)
+
+    # sequence: shape(1, input_length, 1)
+    # text_positions: shape(1, input_length, 1)
+    # Greedy decoding
+    mel_outputs, linear_outputs, alignments, done = model.transduce(
+        sequence, text_positions, speaker_ids)
+
+    # reshape to the desired shape
+    linear_output = linear_outputs.numpy().squeeze().T
+    spectrogram = audio._denormalize(linear_output)
+    alignment = alignments.numpy()[0]
+    mel = mel_outputs.numpy().squeeze().T
+    mel = audio._denormalize(mel)
+
+    # Predicted audio signal
+    waveform = audio.inv_spectrogram(linear_output.T)
+
+    return waveform, alignment, spectrogram, mel
+
+
+def prepare_spec_image(spectrogram):
+    """
+    Prepare an image from spectrogram to be written to tensorboardX 
+    summary writer.
+    
+    Args:
+        spectrogram (numpy.ndarray): Shape(T, C), spectrogram to be
+            visualized, where T means the time steps of the spectrogram,
+            and C means the channels of the spectrogram.
+    Return:
+        np.ndarray: Shape(C, T, 4), the generated image of the spectrogram,
+            where T means the time steps of the spectrogram. It is treated 
+            as the width of the image. And C means the channels of the
+            spectrogram, which is treated as the height of the image. And 4
+            means it is a 'ARGB' format.
+    """
+
+    # [0, 1]
+    spectrogram = (spectrogram - np.min(spectrogram)) / (
+        np.max(spectrogram) - np.min(spectrogram))
+    spectrogram = np.flip(spectrogram, axis=1)  # flip against freq axis
+    return np.uint8(cm.magma(spectrogram.T) * 255)
+
+
+def plot_alignment(alignment, path, info=None):
+    fig, ax = plt.subplots()
+    im = ax.imshow(
+        alignment, aspect="auto", origin="lower", interpolation="none")
+    fig.colorbar(im, ax=ax)
+    xlabel = "Decoder timestep"
+    if info is not None:
+        xlabel += "\n\n" + info
+    plt.xlabel(xlabel)
+    plt.ylabel("Encoder timestep")
+    plt.tight_layout()
+    plt.savefig(path, format="png")
+    plt.close()
+
+
+def time_string():
+    return datetime.now().strftime("%Y-%m-%d %H:%M")
+
+
+def save_alignment(global_step, path, attn):
+    plot_alignment(
+        attn.T,
+        path,
+        info="{}, {}, step={}".format(hparams.builder,
+                                      time_string(), global_step))
+
+
+def eval_model(global_step, writer, model, checkpoint_dir, ismultispeaker):
+    # hard coded text sequences
+    texts = [
+        "Scientists at the CERN laboratory say they have discovered a new particle.",
+        "There's a way to measure the acute emotional intelligence that has never gone out of style.",
+        "President Trump met with other leaders at the Group of 20 conference.",
+        "Generative adversarial network or variational auto-encoder.",
+        "Please call Stella.",
+        "Some have accepted this as a miracle without any physical explanation.",
+    ]
+
+    eval_output_dir = join(checkpoint_dir, "eval")
+    if not os.path.exists(eval_output_dir):
+        os.makedirs(eval_output_dir)
+    print("[eval] Evaluating the model, results are saved in {}".format(
+        eval_output_dir))
+
+    model.eval()
+    # hard coded
+    speaker_ids = [0, 1, 10] if ismultispeaker else [None]
+    for speaker_id in speaker_ids:
+        speaker_str = ("multispeaker{}".format(speaker_id)
+                       if speaker_id is not None else "single")
+
+        for idx, text in enumerate(texts):
+            signal, alignment, _, mel = tts(model,
+                                            text,
+                                            p=0,
+                                            speaker_id=speaker_id)
+            signal /= np.max(np.abs(signal))
+
+            # Alignment
+            path = join(eval_output_dir,
+                        "step{:09d}_text{}_{}_alignment.png".format(
+                            global_step, idx, speaker_str))
+            save_alignment(global_step, path, alignment)
+            tag = "eval_averaged_alignment_{}_{}".format(idx, speaker_str)
+            writer.add_image(
+                tag,
+                np.uint8(cm.viridis(np.flip(alignment, 1).T) * 255),
+                global_step,
+                dataformats='HWC')
+
+            # Mel
+            writer.add_image(
+                "(Eval) Predicted mel spectrogram text{}_{}".format(
+                    idx, speaker_str),
+                prepare_spec_image(mel),
+                global_step,
+                dataformats='HWC')
+
+            # Audio
+            path = join(eval_output_dir,
+                        "step{:09d}_text{}_{}_predicted.wav".format(
+                            global_step, idx, speaker_str))
+            audio.save_wav(signal, path)
+
+            try:
+                writer.add_audio(
+                    "(Eval) Predicted audio signal {}_{}".format(idx,
+                                                                 speaker_str),
+                    signal,
+                    global_step,
+                    sample_rate=hparams.sample_rate)
+            except Exception as e:
+                warn(str(e))
+                pass
+
+
+def save_states(global_step,
+                writer,
+                mel_outputs,
+                linear_outputs,
+                attn,
+                mel,
+                y,
+                input_lengths,
+                checkpoint_dir=None):
+    """
+    Save states for the trainning process.
+    """
+    print("[train] Saving intermediate states at step {}".format(global_step))
+
+    idx = min(1, len(input_lengths) - 1)
+    input_length = input_lengths[idx]
+
+    # Alignment, Multi-hop attention
+    if attn is not None and len(attn.shape) == 4:
+        attn = attn.numpy()
+        for i in range(attn.shape[0]):
+            alignment = attn[i]
+            alignment = alignment[idx]
+            tag = "alignment_layer{}".format(i + 1)
+            writer.add_image(
+                tag,
+                np.uint8(cm.viridis(np.flip(alignment, 1).T) * 255),
+                global_step,
+                dataformats='HWC')
+
+            alignment_dir = join(checkpoint_dir,
+                                 "alignment_layer{}".format(i + 1))
+            if not os.path.exists(alignment_dir):
+                os.makedirs(alignment_dir)
+            path = join(
+                alignment_dir,
+                "step{:09d}_layer_{}_alignment.png".format(global_step, i + 1))
+            save_alignment(global_step, path, alignment)
+
+        alignment_dir = join(checkpoint_dir, "alignment_ave")
+        if not os.path.exists(alignment_dir):
+            os.makedirs(alignment_dir)
+        path = join(alignment_dir,
+                    "step{:09d}_alignment.png".format(global_step))
+        alignment = np.mean(attn, axis=0)[idx]
+        save_alignment(global_step, path, alignment)
+
+        tag = "averaged_alignment"
+        writer.add_image(
+            tag,
+            np.uint8(cm.viridis(np.flip(alignment, 1).T) * 255),
+            global_step,
+            dataformats="HWC")
+
+    if mel_outputs is not None:
+        mel_output = mel_outputs[idx].numpy().squeeze().T
+        mel_output = prepare_spec_image(audio._denormalize(mel_output))
+        writer.add_image(
+            "Predicted mel spectrogram",
+            mel_output,
+            global_step,
+            dataformats="HWC")
+
+    if linear_outputs is not None:
+        linear_output = linear_outputs[idx].numpy().squeeze().T
+        spectrogram = prepare_spec_image(audio._denormalize(linear_output))
+        writer.add_image(
+            "Predicted linear spectrogram",
+            spectrogram,
+            global_step,
+            dataformats="HWC")
+
+        signal = audio.inv_spectrogram(linear_output.T)
+        signal /= np.max(np.abs(signal))
+        path = join(checkpoint_dir,
+                    "step{:09d}_predicted.wav".format(global_step))
+        try:
+            writer.add_audio(
+                "Predicted audio signal",
+                signal,
+                global_step,
+                sample_rate=hparams.sample_rate)
+        except Exception as e:
+            warn(str(e))
+            pass
+        audio.save_wav(signal, path)
+
+    if mel_outputs is not None:
+        mel_output = mel[idx].numpy().squeeze().T
+        mel_output = prepare_spec_image(audio._denormalize(mel_output))
+        writer.add_image(
+            "Target mel spectrogram",
+            mel_output,
+            global_step,
+            dataformats="HWC")
+
+    if linear_outputs is not None:
+        linear_output = y[idx].numpy().squeeze().T
+        spectrogram = prepare_spec_image(audio._denormalize(linear_output))
+        writer.add_image(
+            "Target linear spectrogram",
+            spectrogram,
+            global_step,
+            dataformats="HWC")

deepvoice3/hparams.py

+# Part of code was adpated from https://github.com/r9y9/deepvoice3_pytorch/tree/master/hparams.py
+# Copyright (c) 2017: Ryuichi Yamamoto.
+
+import hparam_tf.hparam
+# NOTE: If you want full control for model architecture. please take a look
+# at the code and change whatever you want. Some hyper parameters are hardcoded.
+
+# Default hyperparameters:
+hparams = hparam_tf.hparam.HParams(
+    name="deepvoice3",
+
+    # Text:
+    # [en, jp]
+    frontend='en',
+
+    # Replace words to its pronunciation with fixed probability.
+    # e.g., 'hello' to 'HH AH0 L OW1'
+    # [en, jp]
+    # en: Word -> pronunciation using CMUDict
+    # jp: Word -> pronounciation usnig MeCab
+    # [0 ~ 1.0]: 0 means no replacement happens.
+    replace_pronunciation_prob=0.5,
+
+    # Convenient model builder
+    # [deepvoice3, deepvoice3_multispeaker, nyanko]
+    # Definitions can be found at deepvoice3_pytorch/builder.py
+    # deepvoice3: DeepVoice3 https://arxiv.org/abs/1710.07654
+    # deepvoice3_multispeaker: Multi-speaker version of DeepVoice3
+    # nyanko: https://arxiv.org/abs/1710.08969
+    builder="deepvoice3",
+
+    # Must be configured depends on the dataset and model you use
+    n_speakers=1,
+    speaker_embed_dim=16,
+
+    # Audio:
+    num_mels=80,
+    fmin=125,
+    fmax=7600,
+    fft_size=1024,
+    hop_size=256,
+    sample_rate=22050,
+    preemphasis=0.97,
+    min_level_db=-100,
+    ref_level_db=20,
+    # whether to rescale waveform or not.
+    # Let x is an input waveform, rescaled waveform y is given by:
+    # y = x / np.abs(x).max() * rescaling_max
+    rescaling=False,
+    rescaling_max=0.999,
+    # mel-spectrogram is normalized to [0, 1] for each utterance and clipping may
+    # happen depends on min_level_db and ref_level_db, causing clipping noise.
+    # If False, assertion is added to ensure no clipping happens.
+    allow_clipping_in_normalization=True,
+
+    # Model:
+    downsample_step=4,  # must be 4 when builder="nyanko"
+    outputs_per_step=1,  # must be 1 when builder="nyanko"
+    embedding_weight_std=0.1,
+    speaker_embedding_weight_std=0.01,
+    padding_idx=0,
+    # Maximum number of input text length
+    # try setting larger value if you want to give very long text input
+    max_positions=512,
+    dropout=1 - 0.95,
+    kernel_size=3,
+    text_embed_dim=128,
+    encoder_channels=256,
+    decoder_channels=256,
+    # Note: large converter channels requires significant computational cost
+    converter_channels=256,
+    query_position_rate=1.0,
+    # can be computed by `compute_timestamp_ratio.py`.
+    key_position_rate=1.385,  # 2.37 for jsut
+    key_projection=False,
+    value_projection=False,
+    use_memory_mask=True,
+    trainable_positional_encodings=False,
+    freeze_embedding=False,
+    # If True, use decoder's internal representation for postnet inputs,
+    # otherwise use mel-spectrogram.
+    use_decoder_state_for_postnet_input=True,
+
+    # Data loader
+    random_seed=1234,
+    pin_memory=True,
+    # Set it to 1 when in Windows (MemoryError, THAllocator.c 0x5)
+    num_workers=2,
+
+    # Loss
+    masked_loss_weight=0.5,  # (1-w)*loss + w * masked_loss
+    # heuristic: priotrize [0 ~ priotiry_freq] for linear loss
+    priority_freq=3000,
+    priority_freq_weight=0.0,  # (1-w)*linear_loss + w*priority_linear_loss
+    # https://arxiv.org/pdf/1710.08969.pdf
+    # Adding the divergence to the loss stabilizes training, expecially for
+    # very deep (> 10 layers) networks.
+    # Binary div loss seems has approx 10x scale compared to L1 loss, so I choose 0.1.
+    binary_divergence_weight=0.1,  # set 0 to disable
+    use_guided_attention=True,
+    guided_attention_sigma=0.2,
+
+    # Training:
+    batch_size=16,
+    adam_beta1=0.5,
+    adam_beta2=0.9,
+    adam_eps=1e-6,
+    amsgrad=False,
+    initial_learning_rate=5e-4,  # 0.001,
+    lr_schedule="noam_learning_rate_decay",
+    lr_schedule_kwargs={},
+    nepochs=2000,
+    weight_decay=0.0,
+    clip_thresh=0.1,
+
+    # Save
+    checkpoint_interval=10000,
+    eval_interval=10000,
+    save_optimizer_state=True,
+
+    # Eval:
+    # this can be list for multple layers of attention
+    # e.g., [True, False, False, False, True]
+    force_monotonic_attention=True,
+    # Attention constraint for incremental decoding
+    window_ahead=3,
+    # 0 tends to prevent word repretetion, but sometime causes skip words
+    window_backward=1,
+    power=1.4,  # Power to raise magnitudes to prior to phase retrieval
+
+    # GC:
+    # Forced garbage collection probability
+    # Use only when MemoryError continues in Windows (Disabled by default)
+    #gc_probability = 0.001,
+
+    # json_meta mode only
+    # 0: "use all",
+    # 1: "ignore only unmatched_alignment",
+    # 2: "fully ignore recognition",
+    ignore_recognition_level=2,
+    # when dealing with non-dedicated speech dataset(e.g. movie excerpts), setting min_text above 15 is desirable. Can be adjusted by dataset.
+    min_text=20,
+    # if true, data without phoneme alignment file(.lab) will be ignored
+    process_only_htk_aligned=False)
+
+
+def hparams_debug_string():
+    values = hparams.values()
+    hp = ['  %s: %s' % (name, values[name]) for name in sorted(values)]
+    return 'Hyperparameters:\n' + '\n'.join(hp)

deepvoice3/ljspeech.py

+# This file is copied from https://github.com/r9y9/deepvoice3_pytorch/tree/master/ljspeech.py
+# Copyright (c) 2017: Ryuichi Yamamoto.
+
+from concurrent.futures import ProcessPoolExecutor
+from functools import partial
+import numpy as np
+import io
+import os
+import audio
+from hparams import hparams
+
+
+def build_from_path(in_dir, out_dir, num_workers=1, tqdm=lambda x: x):
+    '''Preprocesses the LJ Speech dataset from a given input path into a given output directory.
+
+      Args:
+        in_dir: The directory where you have downloaded the LJ Speech dataset
+        out_dir: The directory to write the output into
+        num_workers: Optional number of worker processes to parallelize across
+        tqdm: You can optionally pass tqdm to get a nice progress bar
+
+      Returns:
+        A list of tuples describing the training examples. This should be written to train.txt
+    '''
+
+    # We use ProcessPoolExecutor to parallize across processes. This is just an optimization and you
+    # can omit it and just call _process_utterance on each input if you want.
+    executor = ProcessPoolExecutor(max_workers=num_workers)
+    futures = []
+    index = 1
+    with io.open(
+            os.path.join(in_dir, 'metadata.csv'), "rt", encoding='utf-8') as f:
+        for line in f:
+            parts = line.strip().split('|')
+            wav_path = os.path.join(in_dir, 'wavs', '%s.wav' % parts[0])
+            text = parts[2]
+            if len(text) < hparams.min_text:
+                continue
+            futures.append(
+                executor.submit(
+                    partial(_process_utterance, out_dir, index, wav_path,
+                            text)))
+            index += 1
+    return [future.result() for future in tqdm(futures)]
+
+
+def _process_utterance(out_dir, index, wav_path, text):
+    '''Preprocesses a single utterance audio/text pair.
+
+    This writes the mel and linear scale spectrograms to disk and returns a tuple to write
+    to the train.txt file.
+
+    Args:
+      out_dir: The directory to write the spectrograms into
+      index: The numeric index to use in the spectrogram filenames.
+      wav_path: Path to the audio file containing the speech input
+      text: The text spoken in the input audio file
+
+    Returns:
+      A (spectrogram_filename, mel_filename, n_frames, text) tuple to write to train.txt
+    '''
+
+    # Load the audio to a numpy array:
+    wav = audio.load_wav(wav_path)
+
+    if hparams.rescaling:
+        wav = wav / np.abs(wav).max() * hparams.rescaling_max
+
+    # Compute the linear-scale spectrogram from the wav:
+    spectrogram = audio.spectrogram(wav).astype(np.float32)
+    n_frames = spectrogram.shape[1]
+
+    # Compute a mel-scale spectrogram from the wav:
+    mel_spectrogram = audio.melspectrogram(wav).astype(np.float32)
+
+    # Write the spectrograms to disk:
+    spectrogram_filename = 'ljspeech-spec-%05d.npy' % index
+    mel_filename = 'ljspeech-mel-%05d.npy' % index
+    np.save(
+        os.path.join(out_dir, spectrogram_filename),
+        spectrogram.T,
+        allow_pickle=False)
+    np.save(
+        os.path.join(out_dir, mel_filename),
+        mel_spectrogram.T,
+        allow_pickle=False)
+
+    # Return a tuple describing this training example:
+    return (spectrogram_filename, mel_filename, n_frames, text)

deepvoice3/preprocess.py

+# Part of code was adpated from https://github.com/r9y9/deepvoice3_pytorch/tree/master/preprocess.py
+# Copyright (c) 2017: Ryuichi Yamamoto.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import io
+import six
+import os
+from multiprocessing import cpu_count
+from tqdm import tqdm
+import importlib
+from hparams import hparams, hparams_debug_string
+
+
+def build_parser():
+    parser = argparse.ArgumentParser(description="Data Preprocessing")
+    parser.add_argument("--num-workers", type=int, help="Num workers.")
+    parser.add_argument(
+        "--hparams",
+        type=str,
+        default="",
+        help="Hyper parameters to overwrite.")
+    parser.add_argument(
+        "--preset",
+        type=str,
+        required=True,
+        help="Path of preset parameters (json)")
+    parser.add_argument("name", type=str, help="Dataset name")
+    parser.add_argument("in_dir", type=str, help="Dataset path.")
+    parser.add_argument(
+        "out_dir", type=str, help="Path of preprocessed dataset.")
+    return parser
+
+
+def preprocess(mod, in_dir, out_root, num_workers):
+    if not os.path.exists(out_dir):
+        os.makedirs(out_dir)
+    metadata = mod.build_from_path(in_dir, out_dir, num_workers, tqdm=tqdm)
+    write_metadata(metadata, out_dir)
+
+
+def write_metadata(metadata, out_dir):
+    if six.PY3:
+        string_type = str
+    elif six.PY2:
+        string_type = unicode
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")
+    with io.open(
+            os.path.join(out_dir, 'train.txt'), 'wt', encoding='utf-8') as f:
+        for m in metadata:
+            f.write(u'|'.join([string_type(x) for x in m]) + '\n')
+
+    frames = sum([m[2] for m in metadata])
+    frame_shift_ms = hparams.hop_size / hparams.sample_rate * 1000
+    hours = frames * frame_shift_ms / (3600 * 1000)
+    print('Wrote %d utterances, %d frames (%.2f hours)' %
+          (len(metadata), frames, hours))
+    print('Max input length:  %d' % max(len(m[3]) for m in metadata))
+    print('Max output length: %d' % max(m[2] for m in metadata))
+
+
+if __name__ == "__main__":
+    parser = build_parser()
+    args, _ = parser.parse_known_args()
+
+    name = args.name
+    in_dir = args.in_dir
+    out_dir = args.out_dir
+    num_workers = args.num_workers
+    if num_workers is None:
+        num_workers = cpu_count()
+    preset = args.preset
+
+    # Load preset if specified
+    if preset is not None:
+        with io.open(preset) as f:
+            hparams.parse_json(f.read())
+    # Override hyper parameters
+    hparams.parse(args.hparams)
+    assert hparams.name == "deepvoice3"
+    print(hparams_debug_string())
+
+    assert name in ["ljspeech"], "now we only supports ljspeech"
+    mod = importlib.import_module(name)
+    preprocess(mod, in_dir, out_dir, num_workers)

deepvoice3/presets/deepvoice3_ljspeech.json

+{
+  "name": "deepvoice3",
+  "frontend": "en",
+  "replace_pronunciation_prob": 0.5,
+  "builder": "deepvoice3",
+  "n_speakers": 1,
+  "speaker_embed_dim": 16,
+  "num_mels": 80,
+  "fmin": 125,
+  "fmax": 7600,
+  "fft_size": 1024,
+  "hop_size": 256,
+  "sample_rate": 22050,
+  "preemphasis": 0.97,
+  "min_level_db": -100,
+  "ref_level_db": 20,
+  "rescaling": false,
+  "rescaling_max": 0.999,
+  "allow_clipping_in_normalization": true,
+  "downsample_step": 4,
+  "outputs_per_step": 1,
+  "embedding_weight_std": 0.1,
+  "speaker_embedding_weight_std": 0.01,
+  "padding_idx": 0,
+  "max_positions": 512,
+  "dropout": 0.050000000000000044,
+  "kernel_size": 3,
+  "text_embed_dim": 256,
+  "encoder_channels": 512,
+  "decoder_channels": 256,
+  "converter_channels": 256,
+  "query_position_rate": 1.0,
+  "key_position_rate": 1.385,
+  "key_projection": true,
+  "value_projection": true,
+  "use_memory_mask": true,
+  "trainable_positional_encodings": false,
+  "freeze_embedding": false,
+  "use_decoder_state_for_postnet_input": true,
+  "pin_memory": true,
+  "num_workers": 2,
+  "masked_loss_weight": 0.5,
+  "priority_freq": 3000,
+  "priority_freq_weight": 0.0,
+  "binary_divergence_weight": 0.1,
+  "use_guided_attention": true,
+  "guided_attention_sigma": 0.2,
+  "batch_size": 16,
+  "adam_beta1": 0.5,
+  "adam_beta2": 0.9,
+  "adam_eps": 1e-06,
+  "initial_learning_rate": 0.0005,
+  "lr_schedule": "noam_learning_rate_decay",
+  "lr_schedule_kwargs": {},
+  "nepochs": 2000,
+  "weight_decay": 0.0,
+  "clip_thresh": 0.1,
+  "checkpoint_interval": 10000,
+  "eval_interval": 10000,
+  "save_optimizer_state": true,
+  "force_monotonic_attention": true,
+  "window_ahead": 3,
+  "window_backward": 1,
+  "power": 1.4
+}

deepvoice3/synthesis.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import sys
+import os
+import io
+from os.path import dirname, join, basename, splitext, exists
+from tqdm import tqdm
+import numpy as np
+import nltk
+
+from paddle import fluid
+import paddle.fluid.dygraph as dg
+
+sys.path.append("../")
+import audio
+from modules import frontend
+import dry_run
+
+from hparams import hparams
+from train import make_deepvoice3_from_hparams
+from eval_model import tts, plot_alignment
+
+
+def build_parser():
+    parser = argparse.ArgumentParser(
+        description="Synthesis waveform from trained model.")
+    parser.add_argument(
+        "--hparams", type=str, default="", help="Hyper parameters.")
+    parser.add_argument(
+        "--preset",
+        type=str,
+        required=True,
+        help="Path of preset parameters (json).")
+    parser.add_argument(
+        "--use-gpu",
+        action="store_true",
+        help="Whether to use gpu for generation.")
+    parser.add_argument(
+        "--file-name-suffix", type=str, default="", help="File name suffix.")
+    parser.add_argument(
+        "--max-decoder-steps", type=int, default=500, help="Max decoder steps.")
+    parser.add_argument(
+        "--replace_pronunciation_prob",
+        type=float,
+        default=0.,
+        help="Probility to replace text with pronunciation.")
+    parser.add_argument(
+        "--speaker-id", type=int, help="Speaker ID (for multi-speaker model).")
+    parser.add_argument(
+        "--output-html", action="store_true", help="Output html for blog post.")
+    parser.add_argument(
+        "checkpoint", type=str, help="The checkpoint used for synthesis")
+    parser.add_argument(
+        "text_list_file",
+        type=str,
+        help="Text file to synthesis, a sentence per line.")
+    parser.add_argument(
+        "dst_dir", type=str, help="Directory to save synthesis results.")
+    return parser
+
+
+if __name__ == "__main__":
+    parser = build_parser()
+    args, _ = parser.parse_known_args()
+
+    checkpoint_path = args.checkpoint
+    text_list_file_path = args.text_list_file
+    dst_dir = args.dst_dir
+    use_gpu = args.use_gpu
+
+    max_decoder_steps = args.max_decoder_steps
+    file_name_suffix = args.file_name_suffix
+    replace_pronunciation_prob = args.replace_pronunciation_prob
+    output_html = args.output_html
+    speaker_id = args.speaker_id
+    preset = args.preset
+
+    print("Command Line Args:")
+    for k, v in vars(args).items():
+        print("  {}: {}".format(k, v))
+
+    # Load preset if specified
+    if preset is not None:
+        with io.open(preset) as f:
+            hparams.parse_json(f.read())
+    # Override hyper parameters
+    hparams.parse(args.hparams)
+    assert hparams.name == "deepvoice3"
+
+    place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
+    with dg.guard(place):
+        # Model
+        model = make_deepvoice3_from_hparams(hparams)
+        dry_run(model)
+        model_dict, _ = dg.load_dygraph(args.checkpoint)
+        model.set_dict(model_dict)
+
+        checkpoint_name = splitext(basename(checkpoint_path))[0]
+
+        model.seq2seq.decoder.max_decoder_steps = max_decoder_steps
+
+        if not os.path.exists(dst_dir):
+            os.makedirs(dst_dir)
+        with io.open(text_list_file_path, "rt", encoding="utf-8") as f:
+            lines = f.readlines()
+            for idx, line in enumerate(lines):
+                text = line[:-1]
+                words = nltk.word_tokenize(text)
+                waveform, alignment, _, _ = tts(model,
+                                                text,
+                                                p=replace_pronunciation_prob,
+                                                speaker_id=speaker_id)
+
+                dst_wav_path = join(dst_dir, "{}_{}{}.wav".format(
+                    idx, checkpoint_name, file_name_suffix))
+                dst_alignment_path = join(
+                    dst_dir, "{}_{}{}_alignment.png".format(
+                        idx, checkpoint_name, file_name_suffix))
+                plot_alignment(
+                    alignment.T,
+                    dst_alignment_path,
+                    info="{}, {}".format(hparams.builder,
+                                         basename(checkpoint_path)))
+                audio.save_wav(waveform, dst_wav_path)
+                name = splitext(basename(text_list_file_path))[0]
+                if output_html:
+                    print("""
+                    {}
+                    
+                    ({} chars, {} words)
+                    
+                    <audio controls="controls" >
+                    <source src="/audio/{}/{}/{}" autoplay/>
+                    Your browser does not support the audio element.
+                    </audio>
+                    
+                    <div align="center"><img src="/audio/{}/{}/{}" /></div>
+                      """.format(text,
+                                 len(text),
+                                 len(words), hparams.builder, name,
+                                 basename(dst_wav_path), hparams.builder, name,
+                                 basename(dst_alignment_path)))
+                else:
+                    print(idx, ": {}\n ({} chars, {} words)".format(text,
+                                                                    len(text),
+                                                                    len(words)))
+
+        print("Finished! Check out {} for generated audio samples.".format(
+            dst_dir))
+        sys.exit(0)

deepvoice3/train.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import io
+
+from paddle import fluid
+import paddle.fluid.dygraph as dg
+
+import sys
+sys.path.append("../")
+from argparse import ArgumentParser
+from hparams import hparams, hparams_debug_string
+
+from nnmnkwii.datasets import FileSourceDataset
+from data.data import (TextDataSource, MelSpecDataSource,
+                                    LinearSpecDataSource,
+                                    PartialyRandomizedSimilarTimeLengthSampler,
+                                    Dataset, make_loader, create_batch)
+from modules import frontend
+from builder import deepvoice3, WindowRange
+from dry_run import dry_run
+from train_model import train_model
+from modules.loss import TTSLoss
+from tensorboardX import SummaryWriter
+
+
+def build_arg_parser():
+    parser = ArgumentParser(description="Train deepvoice 3 model.")
+
+    parser.add_argument(
+        "--data-root",
+        type=str,
+        required=True,
+        help="Directory contains preprocessed features.")
+    parser.add_argument(
+        "--use-data-parallel",
+        action="store_true",
+        help="Whether to use data parallel training.")
+    parser.add_argument(
+        "--use-gpu", action="store_true", help="Whether to use gpu training.")
+    parser.add_argument(
+        "--output",
+        type=str,
+        default="result",
+        help="Directory to save results")
+    parser.add_argument(
+        "--preset",
+        type=str,
+        required=True,
+        help="Path of preset parameters in json format.")
+    parser.add_argument(
+        "--hparams",
+        type=str,
+        default="",
+        help="Hyper parameters to override preset.")
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        help="Restore model from checkpoint path if given.")
+    parser.add_argument(
+        "--reset-optimizer", action="store_true", help="Reset optimizer.")
+    # mutually exclusive option
+    train_opt = parser.add_mutually_exclusive_group()
+    train_opt.add_argument(
+        "--train-seq2seq-only",
+        action="store_true",
+        help="Train only seq2seq model")
+    train_opt.add_argument(
+        "--train-postnet-only",
+        action="store_true",
+        help="Train only postnet model.")
+    parser.add_argument(
+        "--speaker-id",
+        type=int,
+        help="Use specific speaker of data in case for multi-speaker datasets.",
+    )
+    return parser
+
+
+def make_deepvoice3_from_hparams(hparams):
+    n_vocab = getattr(frontend, hparams.frontend).n_vocab
+    model = deepvoice3(
+        n_vocab, hparams.text_embed_dim, hparams.num_mels,
+        hparams.fft_size // 2 + 1, hparams.outputs_per_step,
+        hparams.downsample_step, hparams.n_speakers, hparams.speaker_embed_dim,
+        hparams.padding_idx, hparams.dropout, hparams.kernel_size,
+        hparams.encoder_channels, hparams.decoder_channels,
+        hparams.converter_channels, hparams.query_position_rate,
+        hparams.key_position_rate, hparams.use_memory_mask,
+        hparams.trainable_positional_encodings,
+        hparams.force_monotonic_attention,
+        hparams.use_decoder_state_for_postnet_input, hparams.max_positions,
+        hparams.embedding_weight_std, hparams.speaker_embedding_weight_std,
+        hparams.freeze_embedding,
+        WindowRange(-hparams.window_backward, hparams.window_ahead),
+        hparams.key_projection, hparams.value_projection)
+    return model
+
+
+def noam_learning_rate_decay(init_lr, warmup_steps=4000):
+    # Noam scheme from tensor2tensor:
+    warmup_steps = float(warmup_steps)
+    return dg.NoamDecay(1 / (warmup_steps * (init_lr**2)), warmup_steps)
+
+
+def make_optimizer_from_hparams(hparams):
+    if hparams.lr_schedule is not None:
+        learning_rate = noam_learning_rate_decay(hparams.initial_learning_rate,
+                                                 **hparams.lr_schedule_kwargs)
+    else:
+        learning_rate = hparams.initial_learning_rate
+
+    if hparams.weight_decay > 0.0:
+        regularization = fluid.regularizer.L2DecayRegularizer(
+            hparams.weight_decay)
+    else:
+        regularization = None
+
+    optim = fluid.optimizer.Adam(
+        learning_rate=learning_rate,
+        beta1=hparams.adam_beta1,
+        beta2=hparams.adam_beta2,
+        regularization=regularization)
+
+    if hparams.clip_thresh > 0.0:
+        clipper = fluid.dygraph_grad_clip.GradClipByGlobalNorm(
+            hparams.clip_thresh)
+    else:
+        clipper = None
+
+    return optim, clipper
+
+
+def make_loss_from_hparams(hparams):
+    criterion = TTSLoss(
+        hparams.masked_loss_weight, hparams.priority_freq_weight,
+        hparams.binary_divergence_weight, hparams.guided_attention_sigma)
+    return criterion
+
+
+class MyDataParallel(dg.parallel.DataParallel):
+    """
+    A data parallel proxy for model.
+    """
+
+    def __init__(self, layers, strategy):
+        super(MyDataParallel, self).__init__(layers, strategy)
+
+    def __getattr__(self, key):
+        if key in self.__dict__:
+            return object.__getattribute__(self, key)
+        elif key is "_layers":
+            return object.__getattribute__(self, "_sub_layers")["_layers"]
+        else:
+            return getattr(
+                object.__getattribute__(self, "_sub_layers")["_layers"], key)
+
+
+if __name__ == "__main__":
+    parser = build_arg_parser()
+    args, _ = parser.parse_known_args()
+
+    print("Command Line Args:")
+    for k, v in vars(args).items():
+        print("  {}: {}".format(k, v))
+
+    # Load preset if specified
+    if args.preset is not None:
+        with io.open(args.preset) as f:
+            hparams.parse_json(f.read())
+    # Override hyper parameters
+    hparams.parse(args.hparams)
+    print(hparams_debug_string())
+
+    checkpoint_dir = os.path.join(args.output, "checkpoints")
+    tensorboard_dir = os.path.join(args.output, "log")
+    if not os.path.exists(checkpoint_dir):
+        os.makedirs(checkpoint_dir)
+    if not os.path.exists(tensorboard_dir):
+        os.makedirs(tensorboard_dir)
+
+    data_root = args.data_root
+    speaker_id = args.speaker_id
+    X = FileSourceDataset(TextDataSource(data_root, speaker_id))
+    Mel = FileSourceDataset(MelSpecDataSource(data_root, speaker_id))
+    Y = FileSourceDataset(LinearSpecDataSource(data_root, speaker_id))
+
+    frame_lengths = Mel.file_data_source.frame_lengths
+    sampler = PartialyRandomizedSimilarTimeLengthSampler(
+        frame_lengths, batch_size=hparams.batch_size)
+
+    dataset = Dataset(X, Mel, Y)
+    n_trainers = dg.parallel.Env().nranks
+    local_rank = dg.parallel.Env().local_rank
+    data_loader = make_loader(
+        dataset,
+        batch_size=hparams.batch_size,
+        shuffle=False,
+        sampler=sampler,
+        create_batch_fn=create_batch,
+        trainer_count=n_trainers,
+        local_rank=local_rank)
+
+    place = (fluid.CUDAPlace(dg.parallel.Env().dev_id)
+             if args.use_data_parallel else fluid.CUDAPlace(0)
+             if args.use_gpu else fluid.CPUPlace())
+    with dg.guard(place) as g:
+        pyreader = fluid.io.PyReader(capacity=10, return_list=True)
+        pyreader.decorate_batch_generator(data_loader, place)
+
+        model = make_deepvoice3_from_hparams(hparams)
+        optimizer, clipper = make_optimizer_from_hparams(hparams)
+        print("Log event path: {}".format(tensorboard_dir))
+        writer = SummaryWriter(tensorboard_dir) if local_rank == 0 else None
+        criterion = make_loss_from_hparams(hparams)
+
+        # loading saved model
+        if args.train_postnet_only or args.train_seq2seq_only:
+            assert args.checkpoint is not None, \
+                "you must train part of the model from a trained whole model"
+        if args.train_postnet_only:
+            assert hparams.use_decoder_state_for_postnet_input is False, \
+                "when training only the postnet, there is no decoder states"
+
+        if args.checkpoint is not None:
+            model_dict, optimizer_dict = dg.load_dygraph(args.checkpoint)
+
+        if args.use_data_parallel:
+            strategy = dg.parallel.prepare_context()
+            model = MyDataParallel(model, strategy)
+
+        train_model(model, pyreader, criterion, optimizer, clipper, writer,
+                    args, hparams)
+    print("Done!")

deepvoice3/train.sh

+export LD_LIBRARY_PATH=/fluid13_workspace/cuda-9.0/lib64/:/fluid13_workspace/cudnnv7.5_cuda9.0/lib64/:$LD_LIBRARY_PATH
+#export PYTHONPATH=/dv3_workspace/paddle_for_dv3/build/python/
+
+export PYTHONPATH=/fluid13_workspace/paddle_cherry_pick/build/python/:../
+
+export CUDA_VISIBLE_DEVICES=7
+GLOG_v=0 python -u train.py \
+    --use-gpu \
+    --reset-optimizer \
+    --preset=presets/deepvoice3_ljspeech.json \
+    --checkpoint-dir=checkpoint_single_1014 \
+    --data-root="/fluid13_workspace/dv3_workspace/deepvoice3_pytorch/data/ljspeech/" \
+    --hparams="batch_size=16"

deepvoice3/train_model.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import time
+from itertools import chain
+
+from paddle import fluid
+import paddle.fluid.dygraph as dg
+
+from tqdm import tqdm
+from eval_model import eval_model, save_states
+
+
+def train_model(model, loader, criterion, optimizer, clipper, writer, args,
+                hparams):
+    assert fluid.framework.in_dygraph_mode(
+    ), "this function must be run within dygraph guard"
+
+    n_trainers = dg.parallel.Env().nranks
+    local_rank = dg.parallel.Env().local_rank
+
+    # amount of shifting when compute losses
+    linear_shift = hparams.outputs_per_step
+    mel_shift = hparams.outputs_per_step
+
+    global_step = 0
+    global_epoch = 0
+    ismultispeaker = model.n_speakers > 1
+    checkpoint_dir = os.path.join(args.output, "checkpoints")
+    tensorboard_dir = os.path.join(args.output, "log")
+
+    ce_loss = 0
+    start_time = time.time()
+
+    for epoch in range(hparams.nepochs):
+        epoch_loss = 0.
+        for step, inputs in tqdm(enumerate(loader())):
+
+            if len(inputs) == 9:
+                (text, input_lengths, mel, linear, text_positions,
+                 frame_positions, done, target_lengths, speaker_ids) = inputs
+            else:
+                (text, input_lengths, mel, linear, text_positions,
+                 frame_positions, done, target_lengths) = inputs
+                speaker_ids = None
+
+            model.train()
+            if not (args.train_seq2seq_only or args.train_postnet_only):
+                results = model(text, input_lengths, mel, speaker_ids,
+                                text_positions, frame_positions)
+                mel_outputs, linear_outputs, alignments, done_hat = results
+            elif args.train_seq2seq_only:
+
+                if speaker_ids is not None:
+                    speaker_embed = model.speaker_embedding(speaker_ids)
+                else:
+                    speaker_embed = None
+                results = model.seq2seq(text, input_lengths, mel, speaker_embed,
+                                        text_positions, frame_positions)
+                mel_outputs, alignments, done_hat, decoder_states = results
+                if model.r > 1:
+                    mel_outputs = fluid.layers.transpose(mel_outputs,
+                                                         [0, 3, 2, 1])
+                    mel_outputs = fluid.layers.reshape(
+                        mel_outputs,
+                        [mel_outputs.shape[0], -1, 1, model.mel_dim])
+                    mel_outputs = fluid.layers.transpose(mel_outputs,
+                                                         [0, 3, 2, 1])
+
+                linear_outputs = None
+            else:
+                assert (
+                    model.use_decoder_state_for_postnet_input is False
+                ), "when train only the converter, you have no decoder states"
+
+                if speaker_ids is not None:
+                    speaker_embed = model.speaker_embedding(speaker_ids)
+                else:
+                    speaker_embed = None
+                linear_outputs = model.converter(mel, speaker_embed)
+                alignments = None
+                mel_outputs = None
+                done_hat = None
+
+            if not args.train_seq2seq_only:
+                n_priority_freq = int(hparams.priority_freq /
+                                      (hparams.sample_rate * 0.5) *
+                                      model.linear_dim)
+                linear_mask = fluid.layers.sequence_mask(
+                    target_lengths, maxlen=linear.shape[-1], dtype="float32")
+                linear_mask = linear_mask[:, linear_shift:]
+                linear_predicted = linear_outputs[:, :, :, :-linear_shift]
+                linear_target = linear[:, :, :, linear_shift:]
+                lin_l1_loss = criterion.l1_loss(
+                    linear_predicted,
+                    linear_target,
+                    linear_mask,
+                    priority_bin=n_priority_freq)
+                lin_div = criterion.binary_divergence(
+                    linear_predicted, linear_target, linear_mask)
+                lin_loss = criterion.binary_divergence_weight * lin_div \
+                    + (1 - criterion.binary_divergence_weight) * lin_l1_loss
+                if writer is not None and local_rank == 0:
+                    writer.add_scalar("linear_loss",
+                                      float(lin_loss.numpy()), global_step)
+                    writer.add_scalar("linear_l1_loss",
+                                      float(lin_l1_loss.numpy()), global_step)
+                    writer.add_scalar("linear_binary_div_loss",
+                                      float(lin_div.numpy()), global_step)
+
+            if not args.train_postnet_only:
+                mel_lengths = target_lengths // hparams.downsample_step
+                mel_mask = fluid.layers.sequence_mask(
+                    mel_lengths, maxlen=mel.shape[-1], dtype="float32")
+                mel_mask = mel_mask[:, mel_shift:]
+                mel_predicted = mel_outputs[:, :, :, :-mel_shift]
+                mel_target = mel[:, :, :, mel_shift:]
+                mel_l1_loss = criterion.l1_loss(mel_predicted, mel_target,
+                                                mel_mask)
+                mel_div = criterion.binary_divergence(mel_predicted, mel_target,
+                                                      mel_mask)
+                mel_loss = criterion.binary_divergence_weight * mel_div \
+                    + (1 - criterion.binary_divergence_weight) * mel_l1_loss
+                if writer is not None and local_rank == 0:
+                    writer.add_scalar("mel_loss",
+                                      float(mel_loss.numpy()), global_step)
+                    writer.add_scalar("mel_l1_loss",
+                                      float(mel_l1_loss.numpy()), global_step)
+                    writer.add_scalar("mel_binary_div_loss",
+                                      float(mel_div.numpy()), global_step)
+
+                done_loss = criterion.done_loss(done_hat, done)
+                if writer is not None and local_rank == 0:
+                    writer.add_scalar("done_loss",
+                                      float(done_loss.numpy()), global_step)
+
+                if hparams.use_guided_attention:
+                    decoder_length = target_lengths.numpy() / (
+                        hparams.outputs_per_step * hparams.downsample_step)
+                    attn_loss = criterion.attention_loss(alignments,
+                                                         input_lengths.numpy(),
+                                                         decoder_length)
+                    if writer is not None and local_rank == 0:
+                        writer.add_scalar("attention_loss",
+                                          float(attn_loss.numpy()), global_step)
+
+            if not (args.train_seq2seq_only or args.train_postnet_only):
+                if hparams.use_guided_attention:
+                    loss = lin_loss + mel_loss + done_loss + attn_loss
+                else:
+                    loss = lin_loss + mel_loss + done_loss
+            elif args.train_seq2seq_only:
+                if hparams.use_guided_attention:
+                    loss = mel_loss + done_loss + attn_loss
+                else:
+                    loss = mel_loss + done_loss
+            else:
+                loss = lin_loss
+
+            if writer is not None and local_rank == 0:
+                writer.add_scalar("loss", float(loss.numpy()), global_step)
+
+            if isinstance(optimizer._learning_rate,
+                          fluid.optimizer.LearningRateDecay):
+                current_lr = optimizer._learning_rate.step().numpy()
+            else:
+                current_lr = optimizer._learning_rate
+            if writer is not None and local_rank == 0:
+                writer.add_scalar("learning_rate", current_lr, global_step)
+
+            epoch_loss += loss.numpy()[0]
+
+            if (local_rank == 0 and global_step > 0 and
+                    global_step % hparams.checkpoint_interval == 0):
+
+                save_states(global_step, writer, mel_outputs, linear_outputs,
+                            alignments, mel, linear,
+                            input_lengths.numpy(), checkpoint_dir)
+                step_path = os.path.join(
+                    checkpoint_dir, "checkpoint_{:09d}".format(global_step))
+                dg.save_dygraph(model.state_dict(), step_path)
+                dg.save_dygraph(optimizer.state_dict(), step_path)
+
+            if (local_rank == 0 and global_step > 0 and
+                    global_step % hparams.eval_interval == 0):
+                eval_model(global_step, writer, model, checkpoint_dir,
+                           ismultispeaker)
+
+            if args.use_data_parallel:
+                loss = model.scale_loss(loss)
+                loss.backward()
+                model.apply_collective_grads()
+            else:
+                loss.backward()
+
+            if not (args.train_seq2seq_only or args.train_postnet_only):
+                param_list = model.parameters()
+            elif args.train_seq2seq_only:
+                if ismultispeaker:
+                    param_list = chain(model.speaker_embedding.parameters(),
+                                       model.seq2seq.parameters())
+                else:
+                    param_list = model.seq2seq.parameters()
+            else:
+                if ismultispeaker:
+                    param_list = chain(model.speaker_embedding.parameters(),
+                                       model.seq2seq.parameters())
+                else:
+                    param_list = model.converter.parameters()
+
+            optimizer.minimize(
+                loss, grad_clip=clipper, parameter_list=param_list)
+
+            if not (args.train_seq2seq_only or args.train_postnet_only):
+                model.clear_gradients()
+            elif args.train_seq2seq_only:
+                if ismultispeaker:
+                    model.speaker_embedding.clear_gradients()
+                model.seq2seq.clear_gradients()
+            else:
+                if ismultispeaker:
+                    model.speaker_embedding.clear_gradients()
+                model.converter.clear_gradients()
+
+            global_step += 1
+
+        average_loss_in_epoch = epoch_loss / (step + 1)
+        print("Epoch loss: {}".format(average_loss_in_epoch))
+        if writer is not None and local_rank == 0:
+            writer.add_scalar("average_loss_in_epoch", average_loss_in_epoch,
+                              global_epoch)
+        ce_loss = average_loss_in_epoch
+        global_epoch += 1
+    
+    end_time = time.time()
+    epoch_time = (end_time - start_time) / global_epoch 
+    print("kpis\teach_epoch_duration_frame%s_card%s\t%s" %
+            (hparams.outputs_per_step, n_trainers, epoch_time))
+    print("kpis\ttrain_cost_frame%s_card%s\t%f" %
+            (hparams.outputs_per_step, n_trainers, ce_loss))
+
+

hparam_tf/readme.md

+Source: hparam.py copied from tensorflow v1.12.0.
+
+https://github.com/tensorflow/tensorflow/blob/v1.12.0/tensorflow/contrib/training/python/training/hparam.py
+
+with the following:
+wget https://github.com/tensorflow/tensorflow/raw/v1.12.0/tensorflow/contrib/training/python/training/hparam.py
+
+Once all other tensorflow dependencies of these file are removed, the class keeps its goal. Functions not available due to this process are not used in this project.

modules/frontend/README.md

+This package is adapted from https://github.com/r9y9/deepvoice3_pytorch/tree/master/deepvoice3_pytorch/frontend, Copyright (c) 2017: Ryuichi Yamamoto, whose license applies.

modules/frontend/__init__.py

+# coding: utf-8
+"""Text processing frontend
+
+All frontend module should have the following functions:
+
+- text_to_sequence(text, p)
+- sequence_to_text(sequence)
+
+and the property:
+
+- n_vocab
+
+"""
+from . import en
+
+# optinoal Japanese frontend
+try:
+    from . import jp
+except ImportError:
+    jp = None
+
+try:
+    from . import ko
+except ImportError:
+    ko = None
+
+# if you are going to use the frontend, you need to modify _characters in
+# symbol.py:
+# _characters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!\'(),-.:;? ' + '¡¿ñáéíóúÁÉÍÓÚÑ'
+try:
+    from . import es
+except ImportError:
+    es = None

modules/frontend/en/__init__.py

+# coding: utf-8
+from modules.frontend.text.symbols import symbols
+
+import nltk
+from random import random
+
+n_vocab = len(symbols)
+
+_arpabet = nltk.corpus.cmudict.dict()
+
+
+def _maybe_get_arpabet(word, p):
+    try:
+        phonemes = _arpabet[word][0]
+        phonemes = " ".join(phonemes)
+    except KeyError:
+        return word
+
+    return '{%s}' % phonemes if random() < p else word
+
+
+def mix_pronunciation(text, p):
+    text = ' '.join(_maybe_get_arpabet(word, p) for word in text.split(' '))
+    return text
+
+
+def text_to_sequence(text, p=0.0):
+    if p >= 0:
+        text = mix_pronunciation(text, p)
+    from modules.frontend.text import text_to_sequence
+    text = text_to_sequence(text, ["english_cleaners"])
+    return text
+
+
+from modules.frontend.text import sequence_to_text