transform parse to argparse

examples/FastSpeech/config/fastspeech.yaml

@@ -1,46 +0,0 @@
-audio:
-  num_mels: 80  #the number of mel bands when calculating mel spectrograms.
-  n_fft: 2048 #the number of fft components.
-  sr: 22050 #the sampling rate of audio data file.
-  preemphasis: 0.97 #the preemphasis coefficient.
-  hop_length: 256 #the number of samples to advance between frames.
-  win_length: 1024 #the length (width) of the window function.
-  power: 1.2 #the power to raise before griffin-lim.
-  min_level_db: -100 #the minimum level db.
-  ref_level_db: 20 #the reference level db.
-  outputs_per_step: 1 #the outputs per step.
-
-encoder_n_layer: 6
-encoder_head: 2
-encoder_conv1d_filter_size: 1536
-max_sep_len: 2048
-decoder_n_layer: 6
-decoder_head: 2
-decoder_conv1d_filter_size: 1536
-fs_hidden_size: 384
-duration_predictor_output_size: 256
-duration_predictor_filter_size: 3
-fft_conv1d_filter: 3
-fft_conv1d_padding: 1
-dropout: 0.1
-transformer_head: 4
-
-embedding_size: 512
-hidden_size: 256
-
-warm_up_step: 4000
-grad_clip_thresh: 0.1
-batch_size: 32
-epochs: 10000
-lr: 0.001
-save_step: 500
-use_gpu: True
-use_data_parallel: True
-
-data_path: ../../dataset/LJSpeech-1.1
-transtts_path: ../TransformerTTS/checkpoint/
-transformer_step: 160000
-save_path: ./checkpoint
-log_dir: ./log
-#checkpoint_path: ./checkpoint
-#transformer_step: 97000

examples/FastSpeech/parse.py

@@ -1,97 +0,0 @@
-import jsonargparse
-
-def add_config_options_to_parser(parser):
-    parser.add_argument('--audio.num_mels', type=int, default=80,
-        help="the number of mel bands when calculating mel spectrograms.")
-    parser.add_argument('--audio.n_fft', type=int, default=2048,
-        help="the number of fft components.")
-    parser.add_argument('--audio.sr', type=int, default=22050,
-        help="the sampling rate of audio data file.")
-    parser.add_argument('--audio.preemphasis', type=float, default=0.97,
-        help="the preemphasis coefficient.")
-    parser.add_argument('--audio.hop_length', type=int, default=128,
-        help="the number of samples to advance between frames.")
-    parser.add_argument('--audio.win_length', type=int, default=1024,
-        help="the length (width) of the window function.")
-    parser.add_argument('--audio.power', type=float, default=1.4,
-        help="the power to raise before griffin-lim.")
-    parser.add_argument('--audio.min_level_db', type=int, default=-100,
-        help="the minimum level db.")
-    parser.add_argument('--audio.ref_level_db', type=int, default=20,
-        help="the reference level db.")
-    parser.add_argument('--audio.outputs_per_step', type=int, default=1,
-        help="the outputs per step.")
-    
-    parser.add_argument('--encoder_n_layer', type=int, default=6,
-        help="the number of FFT Block in encoder.")
-    parser.add_argument('--encoder_head', type=int, default=2,
-        help="the attention head number in encoder.")
-    parser.add_argument('--encoder_conv1d_filter_size', type=int, default=1024,
-        help="the filter size of conv1d in encoder.")
-    parser.add_argument('--max_sep_len', type=int, default=2048,
-        help="the max length of sequence.")
-    parser.add_argument('--decoder_n_layer', type=int, default=6,
-        help="the number of FFT Block in decoder.")
-    parser.add_argument('--decoder_head', type=int, default=2,
-        help="the attention head number in decoder.")
-    parser.add_argument('--decoder_conv1d_filter_size', type=int, default=1024,
-        help="the filter size of conv1d in decoder.")
-    parser.add_argument('--fs_hidden_size', type=int, default=256,
-        help="the hidden size in model of fastspeech.")
-    parser.add_argument('--duration_predictor_output_size', type=int, default=256,
-        help="the output size of duration predictior.")
-    parser.add_argument('--duration_predictor_filter_size', type=int, default=3,
-        help="the filter size of conv1d in duration prediction.")
-    parser.add_argument('--fft_conv1d_filter', type=int, default=3,
-        help="the filter size of conv1d in fft.")
-    parser.add_argument('--fft_conv1d_padding', type=int, default=1,
-        help="the padding size of conv1d in fft.")
-    parser.add_argument('--dropout', type=float, default=0.1,
-        help="the dropout in network.")
-    parser.add_argument('--transformer_head', type=int, default=4,
-        help="the attention head num of transformerTTS.")
-    parser.add_argument('--alpha', type=float, default=1.0,
-        help="the hyperparameter to determine the length of the expanded sequence\
-                mel, thereby controlling the voice speed.")
-
-    parser.add_argument('--hidden_size', type=int, default=256,
-        help="the hidden size in model of transformerTTS.")
-    parser.add_argument('--embedding_size', type=int, default=256,
-        help="the dim size of embedding of transformerTTS.")
-
-    parser.add_argument('--warm_up_step', type=int, default=4000,
-        help="the warm up step of learning rate.")
-    parser.add_argument('--grad_clip_thresh', type=float, default=1.0,
-        help="the threshold of grad clip.")
-    parser.add_argument('--batch_size', type=int, default=32,
-        help="batch size for training.")
-    parser.add_argument('--epochs', type=int, default=10000,
-        help="the number of epoch for training.")
-    parser.add_argument('--lr', type=float, default=0.001,
-        help="the learning rate for training.")
-    parser.add_argument('--save_step', type=int, default=500,
-        help="checkpointing interval during training.")
-    parser.add_argument('--fastspeech_step', type=int, default=160000,
-        help="Global step to restore checkpoint of fastspeech.")
-    parser.add_argument('--use_gpu', type=bool, default=True,
-        help="use gpu or not during training.")
-    parser.add_argument('--use_data_parallel', type=bool, default=False,
-        help="use data parallel or not during training.")
-
-    parser.add_argument('--data_path', type=str, default='./dataset/LJSpeech-1.1',
-        help="the path of dataset.")
-    parser.add_argument('--checkpoint_path', type=str, default=None,
-        help="the path to load checkpoint or pretrain model.")
-    parser.add_argument('--save_path', type=str, default='./checkpoint',
-        help="the path to save checkpoint.")
-    parser.add_argument('--log_dir', type=str, default='./log',
-        help="the directory to save tensorboard log.")
-    parser.add_argument('--sample_path', type=str, default='./sample',
-        help="the directory to save audio sample in synthesis.")
-    parser.add_argument('--transtts_path', type=str, default='./log',
-        help="the directory to load pretrain transformerTTS model.")
-    parser.add_argument('--transformer_step', type=int, default=70000,
-        help="the step to load transformerTTS model.")
-    
-
-    parser.add_argument('-c', '--config', action=jsonargparse.ActionConfigFile)

examples/TransformerTTS/config/train_transformer.yaml

@@ -1,35 +0,0 @@
-audio:
-  num_mels: 80
-  n_fft: 2048
-  sr: 22050
-  preemphasis: 0.97
-  hop_length: 275
-  win_length: 1102
-  power: 1.2
-  min_level_db: -100
-  ref_level_db: 20
-  outputs_per_step: 1
-
-
-hidden_size: 256
-embedding_size: 512
-
-
-warm_up_step: 4000
-grad_clip_thresh: 1.0
-batch_size: 32
-epochs: 10000
-lr: 0.001
-save_step: 1000
-image_step: 2000
-use_gpu: True
-use_data_parallel: False
-stop_token: False
-
-data_path: ../../dataset/LJSpeech-1.1
-save_path: ./checkpoint
-log_dir: ./log
-#checkpoint_path: ./checkpoint
-#ransformer_step: 97000
-
-   

examples/TransformerTTS/config/train_vocoder.yaml

@@ -1,29 +0,0 @@
-audio:
-  num_mels: 80
-  n_fft: 2048
-  sr: 22050
-  preemphasis: 0.97
-  hop_length: 275
-  win_length: 1102
-  power: 1.2
-  min_level_db: -100
-  ref_level_db: 20
-  outputs_per_step: 1
-
-hidden_size: 256
-embedding_size: 512
-
-warm_up_step: 4000
-grad_clip_thresh: 1.0
-batch_size: 32
-epochs: 10000
-lr: 0.001
-save_step: 10
-use_gpu: True
-use_data_parallel: True
-
-data_path: ../../dataset/LJSpeech-1.1
-save_path: ./checkpoint
-log_dir: ./log
-#checkpoint_path: ./checkpoint
-#transformer_step: 27000

examples/TransformerTTS/parse.py

@@ -1,69 +0,0 @@
-import jsonargparse
-
-def add_config_options_to_parser(parser):
-    parser.add_argument('--audio.num_mels', type=int, default=80,
-        help="the number of mel bands when calculating mel spectrograms.")
-    parser.add_argument('--audio.n_fft', type=int, default=2048,
-        help="the number of fft components.")
-    parser.add_argument('--audio.sr', type=int, default=22050,
-        help="the sampling rate of audio data file.")
-    parser.add_argument('--audio.preemphasis', type=float, default=0.97,
-        help="the preemphasis coefficient.")
-    parser.add_argument('--audio.hop_length', type=int, default=128,
-        help="the number of samples to advance between frames.")
-    parser.add_argument('--audio.win_length', type=int, default=1024,
-        help="the length (width) of the window function.")
-    parser.add_argument('--audio.power', type=float, default=1.4,
-        help="the power to raise before griffin-lim.")
-    parser.add_argument('--audio.min_level_db', type=int, default=-100,
-        help="the minimum level db.")
-    parser.add_argument('--audio.ref_level_db', type=int, default=20,
-        help="the reference level db.")
-    parser.add_argument('--audio.outputs_per_step', type=int, default=1,
-        help="the outputs per step.")
-    
-    parser.add_argument('--hidden_size', type=int, default=256,
-        help="the hidden size in network.")
-    parser.add_argument('--embedding_size', type=int, default=512,
-        help="the embedding vector size.")
-
-    parser.add_argument('--warm_up_step', type=int, default=4000,
-        help="the warm up step of learning rate.")
-    parser.add_argument('--grad_clip_thresh', type=float, default=1.0,
-        help="the threshold of grad clip.")
-    parser.add_argument('--batch_size', type=int, default=32,
-        help="batch size for training.")
-    parser.add_argument('--epochs', type=int, default=10000,
-        help="the number of epoch for training.")
-    parser.add_argument('--lr', type=float, default=0.001,
-        help="the learning rate for training.")
-    parser.add_argument('--save_step', type=int, default=500,
-        help="checkpointing interval during training.")
-    parser.add_argument('--image_step', type=int, default=2000,
-        help="attention image interval during training.")
-    parser.add_argument('--max_len', type=int, default=400,
-        help="The max length of audio when synthsis.")
-    parser.add_argument('--transformer_step', type=int, default=160000,
-        help="Global step to restore checkpoint of transformer.")
-    parser.add_argument('--postnet_step', type=int, default=90000,
-        help="Global step to restore checkpoint of postnet.")
-    parser.add_argument('--use_gpu', type=bool, default=True,
-        help="use gpu or not during training.")
-    parser.add_argument('--use_data_parallel', type=bool, default=False,
-        help="use data parallel or not during training.")
-    parser.add_argument('--stop_token', type=bool, default=False,
-        help="use stop token loss in network or not.")
-
-    parser.add_argument('--data_path', type=str, default='./dataset/LJSpeech-1.1',
-        help="the path of dataset.")
-    parser.add_argument('--checkpoint_path', type=str, default=None,
-        help="the path to load checkpoint or pretrain model.")
-    parser.add_argument('--save_path', type=str, default='./checkpoint',
-        help="the path to save checkpoint.")
-    parser.add_argument('--log_dir', type=str, default='./log',
-        help="the directory to save tensorboard log.")
-    parser.add_argument('--sample_path', type=str, default='./log',
-        help="the directory to save audio sample in synthesis.")
-    
-
-    parser.add_argument('-c', '--config', action=jsonargparse.ActionConfigFile)

examples/fastspeech/config/fastspeech.yaml

@@ -0,0 +1,32 @@
+audio:
+  num_mels: 80  #the number of mel bands when calculating mel spectrograms.
+  n_fft: 2048 #the number of fft components.
+  sr: 22050 #the sampling rate of audio data file.
+  preemphasis: 0.97 #the preemphasis coefficient.
+  hop_length: 256 #the number of samples to advance between frames.
+  win_length: 1024 #the length (width) of the window function.
+  power: 1.2 #the power to raise before griffin-lim.
+  min_level_db: -100 #the minimum level db.
+  ref_level_db: 20 #the reference level db.
+  outputs_per_step: 1 #the outputs per step.
+
+encoder_n_layer: 6 #the number of FFT Block in encoder.
+encoder_head: 2 #the attention head number in encoder.
+encoder_conv1d_filter_size: 1536 #the filter size of conv1d in encoder.
+max_seq_len: 2048 #the max length of sequence.
+decoder_n_layer: 6 #the number of FFT Block in decoder.
+decoder_head: 2 #the attention head number in decoder.
+decoder_conv1d_filter_size: 1536 #the filter size of conv1d in decoder.
+fs_hidden_size: 384 #the hidden size in model of fastspeech.
+duration_predictor_output_size: 256 #the output size of duration predictior.
+duration_predictor_filter_size: 3 #the filter size of conv1d in duration prediction.
+fft_conv1d_filter: 3 #the filter size of conv1d in fft.
+fft_conv1d_padding: 1 #the padding size of conv1d in fft.
+dropout: 0.1 #the dropout in network.
+transformer_head: 4 #the attention head num of transformerTTS.
+
+embedding_size: 512 #the dim size of embedding of transformerTTS.
+hidden_size: 256 #the hidden size in model of transformerTTS.
+warm_up_step: 4000 #the warm up step of learning rate.
+grad_clip_thresh: 0.1 #the threshold of grad clip.
+

examples/fastspeech/config/synthesis.yaml

@@ -13,7 +13,7 @@ audio:
 encoder_n_layer: 6
 encoder_head: 2
 encoder_conv1d_filter_size: 1536
-max_sep_len: 2048
+max_seq_len: 2048
 decoder_n_layer: 6
 decoder_head: 2
 decoder_conv1d_filter_size: 1536
@@ -23,11 +23,4 @@ duration_predictor_filter_size: 3
 fft_conv1d_filter: 3
 fft_conv1d_padding: 1
 dropout: 0.1
-transformer_head: 4
+transformer_head: 4
-
-use_gpu: True
-alpha: 1.0
-
-checkpoint_path: checkpoint/
-fastspeech_step: 71000
-log_dir: ./log

examples/fastspeech/parse.py

@@ -0,0 +1,36 @@
+import argparse
+
+def add_config_options_to_parser(parser):
+    parser.add_argument('--config_path', type=str, default='config/fastspeech.yaml',
+        help="the yaml config file path.")
+    parser.add_argument('--batch_size', type=int, default=32,
+        help="batch size for training.")
+    parser.add_argument('--epochs', type=int, default=10000,
+        help="the number of epoch for training.")
+    parser.add_argument('--lr', type=float, default=0.001,
+        help="the learning rate for training.")
+    parser.add_argument('--save_step', type=int, default=500,
+        help="checkpointing interval during training.")
+    parser.add_argument('--fastspeech_step', type=int, default=70000,
+        help="Global step to restore checkpoint of fastspeech.")
+    parser.add_argument('--use_gpu', type=int, default=1,
+        help="use gpu or not during training.")
+    parser.add_argument('--use_data_parallel', type=int, default=0,
+        help="use data parallel or not during training.")
+
+    parser.add_argument('--data_path', type=str, default='./dataset/LJSpeech-1.1',
+        help="the path of dataset.")
+    parser.add_argument('--checkpoint_path', type=str, default=None,
+        help="the path to load checkpoint or pretrain model.")
+    parser.add_argument('--save_path', type=str, default='./checkpoint',
+        help="the path to save checkpoint.")
+    parser.add_argument('--log_dir', type=str, default='./log',
+        help="the directory to save tensorboard log.")
+    parser.add_argument('--sample_path', type=str, default='./sample',
+        help="the directory to save audio sample in synthesis.")
+    parser.add_argument('--transtts_path', type=str, default='./log',
+        help="the directory to load pretrain transformerTTS model.")
+    parser.add_argument('--transformer_step', type=int, default=160000,
+        help="the step to load transformerTTS model.")
+    
+

examples/fastspeech/synthesis.py

@@ -1,15 +1,16 @@
 import os
 from tensorboardX import SummaryWriter
 from collections import OrderedDict
-import jsonargparse
+import argparse
 from parse import add_config_options_to_parser
 from pprint import pprint
+from ruamel import yaml
 import numpy as np
 import paddle.fluid as fluid
 import paddle.fluid.dygraph as dg
 from parakeet.g2p.en import text_to_sequence
 from parakeet import audio
-from network import FastSpeech
+from parakeet.models.fastspeech.fastspeech import FastSpeech
 
 def load_checkpoint(step, model_path):
     model_dict, _ = fluid.dygraph.load_dygraph(os.path.join(model_path, step))
@@ -21,19 +22,22 @@ def load_checkpoint(step, model_path):
             new_state_dict[param] = model_dict[param]
     return new_state_dict
 
-def synthesis(text_input, cfg):
+def synthesis(text_input, args):
-    place = (fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace())
+    place = (fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace())
 
     # tensorboard
-    if not os.path.exists(cfg.log_dir):
+    if not os.path.exists(args.log_dir):
-            os.mkdir(cfg.log_dir)
+            os.mkdir(args.log_dir)
-    path = os.path.join(cfg.log_dir,'synthesis')
+    path = os.path.join(args.log_dir,'synthesis')
+
+    with open(args.config_path) as f:
+        cfg = yaml.load(f, Loader=yaml.Loader)
 
     writer = SummaryWriter(path)
 
     with dg.guard(place):
         model = FastSpeech(cfg)
-        model.set_dict(load_checkpoint(str(cfg.fastspeech_step), os.path.join(cfg.checkpoint_path, "fastspeech")))
+        model.set_dict(load_checkpoint(str(args.fastspeech_step), os.path.join(args.checkpoint_path, "fastspeech")))
         model.eval()
 
         text = np.asarray(text_to_sequence(text_input))
@@ -41,18 +45,18 @@ def synthesis(text_input, cfg):
         pos_text = np.arange(1, text.shape[1]+1)
         pos_text = fluid.layers.unsqueeze(dg.to_variable(pos_text),[0])
 
-        mel_output, mel_output_postnet = model(text, pos_text, alpha=cfg.alpha)
+        mel_output, mel_output_postnet = model(text, pos_text, alpha=args.alpha)
 
         _ljspeech_processor = audio.AudioProcessor(
-            sample_rate=cfg.audio.sr, 
+            sample_rate=cfg['audio']['sr'], 
-            num_mels=cfg.audio.num_mels, 
+            num_mels=cfg['audio']['num_mels'], 
-            min_level_db=cfg.audio.min_level_db, 
+            min_level_db=cfg['audio']['min_level_db'], 
-            ref_level_db=cfg.audio.ref_level_db, 
+            ref_level_db=cfg['audio']['ref_level_db'], 
-            n_fft=cfg.audio.n_fft, 
+            n_fft=cfg['audio']['n_fft'], 
-            win_length= cfg.audio.win_length, 
+            win_length= cfg['audio']['win_length'], 
-            hop_length= cfg.audio.hop_length,
+            hop_length= cfg['audio']['hop_length'],
-            power=cfg.audio.power,
+            power=cfg['audio']['power'],
-            preemphasis=cfg.audio.preemphasis,
+            preemphasis=cfg['audio']['preemphasis'],
             signal_norm=True,
             symmetric_norm=False,
             max_norm=1.,
@@ -65,12 +69,12 @@ def synthesis(text_input, cfg):
 
         mel_output_postnet = fluid.layers.transpose(fluid.layers.squeeze(mel_output_postnet,[0]), [1,0])
         wav = _ljspeech_processor.inv_melspectrogram(mel_output_postnet.numpy())
-        writer.add_audio(text_input, wav, 0, cfg.audio.sr)
+        writer.add_audio(text_input, wav, 0, cfg['audio']['sr'])
         print("Synthesis completed !!!")
     writer.close()
 
 if __name__ == '__main__':
-    parser = jsonargparse.ArgumentParser(description="Synthesis model", formatter_class='default_argparse')
+    parser = argparse.ArgumentParser(description="Train Fastspeech model")
     add_config_options_to_parser(parser)
-    cfg = parser.parse_args('-c ./config/synthesis.yaml'.split())
+    args = parser.parse_args()
-    synthesis("Transformer model is so fast!", cfg)
+    synthesis("Transformer model is so fast!", args)

examples/fastspeech/train.py

@@ -3,10 +3,10 @@ import argparse
 import os
 import time
 import math
-import jsonargparse
 from pathlib import Path
 from parse import add_config_options_to_parser
 from pprint import pprint
+from ruamel import yaml
 from tqdm import tqdm
 from collections import OrderedDict
 from tensorboardX import SummaryWriter
@@ -14,7 +14,7 @@ import paddle.fluid.dygraph as dg
 import paddle.fluid.layers as layers
 import paddle.fluid as fluid
 from parakeet.models.dataloader.ljspeech import LJSpeechLoader
-from parakeet.models.transformerTTS.transformerTTS import TransformerTTS
+from parakeet.models.transformer_tts.transformerTTS import TransformerTTS
 from parakeet.models.fastspeech.fastspeech import FastSpeech
 from parakeet.models.fastspeech.utils import get_alignment
 
@@ -28,50 +28,49 @@ def load_checkpoint(step, model_path):
             new_state_dict[param] = model_dict[param]
     return new_state_dict, opti_dict
 
-def main(cfg):
+def main(args):
-    local_rank = dg.parallel.Env().local_rank if cfg.use_data_parallel else 0
+    local_rank = dg.parallel.Env().local_rank if args.use_data_parallel else 0
-    nranks = dg.parallel.Env().nranks if cfg.use_data_parallel else 1
+    nranks = dg.parallel.Env().nranks if args.use_data_parallel else 1
 
-    if local_rank == 0:
+    with open(args.config_path) as f:
-        # Print the whole config setting.
+        cfg = yaml.load(f, Loader=yaml.Loader)
-        pprint(jsonargparse.namespace_to_dict(cfg))
 
     global_step = 0
     place = (fluid.CUDAPlace(dg.parallel.Env().dev_id)
-             if cfg.use_data_parallel else fluid.CUDAPlace(0)
+             if args.use_data_parallel else fluid.CUDAPlace(0)
-             if cfg.use_gpu else fluid.CPUPlace())
+             if args.use_gpu else fluid.CPUPlace())
 
-    if not os.path.exists(cfg.log_dir):
+    if not os.path.exists(args.log_dir):
-            os.mkdir(cfg.log_dir)
+            os.mkdir(args.log_dir)
-    path = os.path.join(cfg.log_dir,'fastspeech')
+    path = os.path.join(args.log_dir,'fastspeech')
 
     writer = SummaryWriter(path) if local_rank == 0 else None
 
     with dg.guard(place):
         with fluid.unique_name.guard():
             transformerTTS = TransformerTTS(cfg)
-            model_dict, _ = load_checkpoint(str(cfg.transformer_step), os.path.join(cfg.transtts_path, "transformer"))
+            model_dict, _ = load_checkpoint(str(args.transformer_step), os.path.join(args.transtts_path, "transformer"))
             transformerTTS.set_dict(model_dict)
             transformerTTS.eval()
 
         model = FastSpeech(cfg)
         model.train()
-        optimizer = fluid.optimizer.AdamOptimizer(learning_rate=dg.NoamDecay(1/(cfg.warm_up_step *( cfg.lr ** 2)), cfg.warm_up_step),
+        optimizer = fluid.optimizer.AdamOptimizer(learning_rate=dg.NoamDecay(1/(cfg['warm_up_step'] *( args.lr ** 2)), cfg['warm_up_step']),
                                                   parameter_list=model.parameters())
-        reader = LJSpeechLoader(cfg, nranks, local_rank, shuffle=True).reader()
+        reader = LJSpeechLoader(cfg, args, nranks, local_rank, shuffle=True).reader()
         
-        if cfg.checkpoint_path is not None:
+        if args.checkpoint_path is not None:
-            model_dict, opti_dict = load_checkpoint(str(cfg.fastspeech_step), os.path.join(cfg.checkpoint_path, "fastspeech"))
+            model_dict, opti_dict = load_checkpoint(str(args.fastspeech_step), os.path.join(args.checkpoint_path, "fastspeech"))
             model.set_dict(model_dict)
             optimizer.set_dict(opti_dict)
-            global_step = cfg.fastspeech_step
+            global_step = args.fastspeech_step
             print("load checkpoint!!!")
 
-        if cfg.use_data_parallel:
+        if args.use_data_parallel:
             strategy = dg.parallel.prepare_context()
             model = fluid.dygraph.parallel.DataParallel(model, strategy)
 
-        for epoch in range(cfg.epochs):
+        for epoch in range(args.epochs):
             pbar = tqdm(reader)
 
             for i, data in enumerate(pbar):
@@ -79,7 +78,7 @@ def main(cfg):
                 character, mel, mel_input, pos_text, pos_mel, text_length, mel_lens = data
 
                 _, _, attn_probs, _, _, _ = transformerTTS(character, mel_input, pos_text, pos_mel)
-                alignment = dg.to_variable(get_alignment(attn_probs, mel_lens, cfg.transformer_head)).astype(np.float32)
+                alignment = dg.to_variable(get_alignment(attn_probs, mel_lens, cfg['transformer_head'])).astype(np.float32)
 
                 global_step += 1
                     
@@ -101,20 +100,20 @@ def main(cfg):
                     writer.add_scalar('learning_rate', optimizer._learning_rate.step().numpy(), global_step)
 
 
-                if cfg.use_data_parallel:
+                if args.use_data_parallel:
                     total_loss = model.scale_loss(total_loss)
                     total_loss.backward()
                     model.apply_collective_grads()
                 else:
                     total_loss.backward()
-                optimizer.minimize(total_loss, grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(cfg.grad_clip_thresh))
+                optimizer.minimize(total_loss, grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(cfg['grad_clip_thresh']))
                 model.clear_gradients()
 
                  # save checkpoint
-                if local_rank==0 and global_step % cfg.save_step == 0:
+                if local_rank==0 and global_step % args.save_step == 0:
-                    if not os.path.exists(cfg.save_path):
+                    if not os.path.exists(args.save_path):
-                        os.mkdir(cfg.save_path)
+                        os.mkdir(args.save_path)
-                    save_path = os.path.join(cfg.save_path,'fastspeech/%d' % global_step)
+                    save_path = os.path.join(args.save_path,'fastspeech/%d' % global_step)
                     dg.save_dygraph(model.state_dict(), save_path)
                     dg.save_dygraph(optimizer.state_dict(), save_path)
         if local_rank==0:
@@ -122,7 +121,9 @@ def main(cfg):
 
 
 if __name__ =='__main__':
-    parser = jsonargparse.ArgumentParser(description="Train Fastspeech model", formatter_class='default_argparse')
+    parser = argparse.ArgumentParser(description="Train Fastspeech model")
     add_config_options_to_parser(parser)
-    cfg = parser.parse_args('-c config/fastspeech.yaml'.split())
+    args = parser.parse_args()
-    main(cfg)
+    # Print the whole config setting.
+    pprint(args)
+    main(args)

examples/transformer_tts/config/synthesis.yaml

@@ -0,0 +1,11 @@
+audio:
+  num_mels: 80
+  n_fft: 2048
+  sr: 22050
+  preemphasis: 0.97
+  hop_length: 275
+  win_length: 1102
+  power: 1.2
+  min_level_db: -100
+  ref_level_db: 20
+  outputs_per_step: 1

examples/transformer_tts/config/train_transformer.yaml

@@ -10,11 +10,11 @@ audio:
   ref_level_db: 20
   outputs_per_step: 1
 
-max_len: 50
-transformer_step: 10
-postnet_step: 10
-use_gpu: True
 
-checkpoint_path: ./checkpoint
+hidden_size: 256
-log_dir: ./log
+embedding_size: 512
-sample_path: ./sample
+warm_up_step: 4000
+grad_clip_thresh: 1.0
+
+
+   

examples/transformer_tts/config/train_vocoder.yaml

@@ -0,0 +1,16 @@
+audio:
+  num_mels: 80
+  n_fft: 2048
+  sr: 22050
+  preemphasis: 0.97
+  hop_length: 275
+  win_length: 1102
+  power: 1.2
+  min_level_db: -100
+  ref_level_db: 20
+  outputs_per_step: 1
+
+hidden_size: 256
+embedding_size: 512
+warm_up_step: 4000
+grad_clip_thresh: 1.0

examples/transformer_tts/parse.py

@@ -0,0 +1,38 @@
+import argparse
+
+def add_config_options_to_parser(parser):
+    parser.add_argument('--config_path', type=str, default='config/train_transformer.yaml',
+        help="the yaml config file path.")
+    parser.add_argument('--batch_size', type=int, default=32,
+        help="batch size for training.")
+    parser.add_argument('--epochs', type=int, default=10000,
+        help="the number of epoch for training.")
+    parser.add_argument('--lr', type=float, default=0.001,
+        help="the learning rate for training.")
+    parser.add_argument('--save_step', type=int, default=500,
+        help="checkpointing interval during training.")
+    parser.add_argument('--image_step', type=int, default=2000,
+        help="attention image interval during training.")
+    parser.add_argument('--max_len', type=int, default=400,
+        help="The max length of audio when synthsis.")
+    parser.add_argument('--transformer_step', type=int, default=160000,
+        help="Global step to restore checkpoint of transformer.")
+    parser.add_argument('--vocoder_step', type=int, default=90000,
+        help="Global step to restore checkpoint of postnet.")
+    parser.add_argument('--use_gpu', type=int, default=1,
+        help="use gpu or not during training.")
+    parser.add_argument('--use_data_parallel', type=int, default=0,
+        help="use data parallel or not during training.")
+    parser.add_argument('--stop_token', type=int, default=0,
+        help="use stop token loss in network or not.")
+
+    parser.add_argument('--data_path', type=str, default='./dataset/LJSpeech-1.1',
+        help="the path of dataset.")
+    parser.add_argument('--checkpoint_path', type=str, default=None,
+        help="the path to load checkpoint or pretrain model.")
+    parser.add_argument('--save_path', type=str, default='./checkpoint',
+        help="the path to save checkpoint.")
+    parser.add_argument('--log_dir', type=str, default='./log',
+        help="the directory to save tensorboard log.")
+    parser.add_argument('--sample_path', type=str, default='./sample',
+        help="the directory to save audio sample in synthesis.")

examples/transformer_tts/synthesis.py

@@ -2,17 +2,19 @@ import os
 from scipy.io.wavfile import write
 from parakeet.g2p.en import text_to_sequence
 import numpy as np
-from network import TransformerTTS, ModelPostNet
 from tqdm import tqdm
 from tensorboardX import SummaryWriter
+from ruamel import yaml
 import paddle.fluid as fluid
 import paddle.fluid.dygraph as dg
 from pathlib import Path
-import jsonargparse
+import argparse
 from parse import add_config_options_to_parser
 from pprint import pprint
 from collections import OrderedDict
 from parakeet import audio
+from parakeet.models.transformer_tts.vocoder import Vocoder
+from parakeet.models.transformer_tts.transformerTTS import TransformerTTS
 
 def load_checkpoint(step, model_path):
     model_dict, _ = fluid.dygraph.load_dygraph(os.path.join(model_path, step))
@@ -24,25 +26,28 @@ def load_checkpoint(step, model_path):
             new_state_dict[param] = model_dict[param]
     return new_state_dict
 
-def synthesis(text_input, cfg):
+def synthesis(text_input, args):
-    place = (fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace())
+    place = (fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace())
+
+    with open(args.config_path) as f:
+        cfg = yaml.load(f, Loader=yaml.Loader)
 
     # tensorboard
-    if not os.path.exists(cfg.log_dir):
+    if not os.path.exists(args.log_dir):
-            os.mkdir(cfg.log_dir)
+            os.mkdir(args.log_dir)
-    path = os.path.join(cfg.log_dir,'synthesis')
+    path = os.path.join(args.log_dir,'synthesis')
 
     writer = SummaryWriter(path)
 
     with dg.guard(place):
         with fluid.unique_name.guard():
             model = TransformerTTS(cfg)
-            model.set_dict(load_checkpoint(str(cfg.transformer_step), os.path.join(cfg.checkpoint_path, "nostop_token/transformer")))
+            model.set_dict(load_checkpoint(str(args.transformer_step), os.path.join(args.checkpoint_path, "nostop_token/transformer")))
             model.eval()
         
         with fluid.unique_name.guard():
-            model_postnet = ModelPostNet(cfg)
+            model_postnet = Vocoder(cfg, args.batch_size)
-            model_postnet.set_dict(load_checkpoint(str(cfg.postnet_step), os.path.join(cfg.checkpoint_path, "postnet")))
+            model_postnet.set_dict(load_checkpoint(str(args.postnet_step), os.path.join(args.checkpoint_path, "postnet")))
             model_postnet.eval()
         # init input
         text = np.asarray(text_to_sequence(text_input))
@@ -52,7 +57,7 @@ def synthesis(text_input, cfg):
         pos_text = fluid.layers.unsqueeze(dg.to_variable(pos_text),[0])
         
 
-        pbar = tqdm(range(cfg.max_len))
+        pbar = tqdm(range(args.max_len))
 
         for i in pbar:
             pos_mel = np.arange(1, mel_input.shape[1]+1)
@@ -62,15 +67,15 @@ def synthesis(text_input, cfg):
         mag_pred = model_postnet(postnet_pred)
 
         _ljspeech_processor = audio.AudioProcessor(
-            sample_rate=cfg.audio.sr, 
+            sample_rate=cfg['audio']['sr'], 
-            num_mels=cfg.audio.num_mels, 
+            num_mels=cfg['audio']['num_mels'], 
-            min_level_db=cfg.audio.min_level_db, 
+            min_level_db=cfg['audio']['min_level_db'], 
-            ref_level_db=cfg.audio.ref_level_db, 
+            ref_level_db=cfg['audio']['ref_level_db'], 
-            n_fft=cfg.audio.n_fft, 
+            n_fft=cfg['audio']['n_fft'], 
-            win_length= cfg.audio.win_length, 
+            win_length= cfg['audio']['win_length'], 
-            hop_length= cfg.audio.hop_length,
+            hop_length= cfg['audio']['hop_length'],
-            power=cfg.audio.power,
+            power=cfg['audio']['power'],
-            preemphasis=cfg.audio.preemphasis,
+            preemphasis=cfg['audio']['preemphasis'],
             signal_norm=True,
             symmetric_norm=False,
             max_norm=1.,
@@ -82,14 +87,14 @@ def synthesis(text_input, cfg):
             sound_norm=False)
 
         wav = _ljspeech_processor.inv_spectrogram(fluid.layers.transpose(fluid.layers.squeeze(mag_pred,[0]), [1,0]).numpy())
-        writer.add_audio(text_input, wav, 0, cfg.audio.sr)
+        writer.add_audio(text_input, wav, 0, cfg['audio']['sr'])
-        if not os.path.exists(cfg.sample_path):
+        if not os.path.exists(args.sample_path):
-            os.mkdir(cfg.sample_path)
+            os.mkdir(args.sample_path)
-        write(os.path.join(cfg.sample_path,'test.wav'), cfg.audio.sr, wav)
+        write(os.path.join(args.sample_path,'test.wav'), cfg['audio']['sr'], wav)
     writer.close()
 
 if __name__ == '__main__':
-    parser = jsonargparse.ArgumentParser(description="Synthesis model", formatter_class='default_argparse')
+    parser = argparse.ArgumentParser(description="Synthesis model")
     add_config_options_to_parser(parser)
-    cfg = parser.parse_args('-c ./config/synthesis.yaml'.split())
+    args = parser.parse_args()
-    synthesis("Transformer model is so fast!", cfg)
+    synthesis("Transformer model is so fast!", args)

examples/transformer_tts/train_transformer.py

@@ -3,9 +3,10 @@ from tqdm import tqdm
 from tensorboardX import SummaryWriter
 from pathlib import Path
 from collections import OrderedDict
-import jsonargparse
+import argparse
 from parse import add_config_options_to_parser
 from pprint import pprint
+from ruamel import yaml
 from matplotlib import cm
 import numpy as np
 import paddle.fluid as fluid
@@ -13,7 +14,7 @@ import paddle.fluid.dygraph as dg
 import paddle.fluid.layers as layers
 from parakeet.modules.utils import cross_entropy
 from parakeet.models.dataloader.ljspeech import LJSpeechLoader
-from parakeet.models.transformerTTS.transformerTTS import TransformerTTS
+from parakeet.models.transformer_tts.transformerTTS import TransformerTTS
 
 def load_checkpoint(step, model_path):
     model_dict, opti_dict = fluid.dygraph.load_dygraph(os.path.join(model_path, step))
@@ -26,22 +27,21 @@ def load_checkpoint(step, model_path):
     return new_state_dict, opti_dict
 
 
-def main(cfg):
+def main(args):
-    local_rank = dg.parallel.Env().local_rank if cfg.use_data_parallel else 0
+    local_rank = dg.parallel.Env().local_rank if args.use_data_parallel else 0
-    nranks = dg.parallel.Env().nranks if cfg.use_data_parallel else 1
+    nranks = dg.parallel.Env().nranks if args.use_data_parallel else 1
 
-    if local_rank == 0:
+    with open(args.config_path) as f:
-        # Print the whole config setting.
+        cfg = yaml.load(f, Loader=yaml.Loader)
-        pprint(jsonargparse.namespace_to_dict(cfg))
 
     global_step = 0
     place = (fluid.CUDAPlace(dg.parallel.Env().dev_id)
-             if cfg.use_data_parallel else fluid.CUDAPlace(0)
+             if args.use_data_parallel else fluid.CUDAPlace(0)
-             if cfg.use_gpu else fluid.CPUPlace())
+             if args.use_gpu else fluid.CPUPlace())
 
-    if not os.path.exists(cfg.log_dir):
+    if not os.path.exists(args.log_dir):
-            os.mkdir(cfg.log_dir)
+            os.mkdir(args.log_dir)
-    path = os.path.join(cfg.log_dir,'transformer')
+    path = os.path.join(args.log_dir,'transformer')
 
     writer = SummaryWriter(path) if local_rank == 0 else None
     
@@ -49,23 +49,23 @@ def main(cfg):
         model = TransformerTTS(cfg)
 
         model.train()
-        optimizer = fluid.optimizer.AdamOptimizer(learning_rate=dg.NoamDecay(1/(cfg.warm_up_step *( cfg.lr ** 2)), cfg.warm_up_step), 
+        optimizer = fluid.optimizer.AdamOptimizer(learning_rate=dg.NoamDecay(1/(cfg['warm_up_step'] *( args.lr ** 2)), cfg['warm_up_step']), 
                                                   parameter_list=model.parameters())
         
-        reader = LJSpeechLoader(cfg, nranks, local_rank, shuffle=True).reader()
+        reader = LJSpeechLoader(cfg, args, nranks, local_rank, shuffle=True).reader()
-        
+
-        if cfg.checkpoint_path is not None:
+        if args.checkpoint_path is not None:
-            model_dict, opti_dict = load_checkpoint(str(cfg.transformer_step), os.path.join(cfg.checkpoint_path, "transformer"))
+            model_dict, opti_dict = load_checkpoint(str(args.transformer_step), os.path.join(args.checkpoint_path, "transformer"))
             model.set_dict(model_dict)
             optimizer.set_dict(opti_dict)
-            global_step = cfg.transformer_step
+            global_step = args.transformer_step
             print("load checkpoint!!!")
 
-        if cfg.use_data_parallel:
+        if args.use_data_parallel:
             strategy = dg.parallel.prepare_context()
             model = fluid.dygraph.parallel.DataParallel(model, strategy)
         
-        for epoch in range(cfg.epochs):
+        for epoch in range(args.epochs):
             pbar = tqdm(reader)
             for i, data in enumerate(pbar):
                 pbar.set_description('Processing at epoch %d'%epoch)
@@ -81,7 +81,7 @@ def main(cfg):
                 post_mel_loss = layers.mean(layers.abs(layers.elementwise_sub(postnet_pred, mel)))
                 loss = mel_loss + post_mel_loss
                 # Note: When used stop token loss the learning did not work.
-                if cfg.stop_token:
+                if args.stop_token:
                     stop_loss = cross_entropy(stop_preds, label)
                     loss = loss + stop_loss
 
@@ -91,7 +91,7 @@ def main(cfg):
                         'post_mel_loss':post_mel_loss.numpy()
                     }, global_step)
 
-                    if cfg.stop_token:
+                    if args.stop_token:
                         writer.add_scalar('stop_loss', stop_loss.numpy(), global_step)
 
                     writer.add_scalars('alphas', {
@@ -101,7 +101,7 @@ def main(cfg):
 
                     writer.add_scalar('learning_rate', optimizer._learning_rate.step().numpy(), global_step)
 
-                    if global_step % cfg.image_step == 1:
+                    if global_step % args.image_step == 1:
                         for i, prob in enumerate(attn_probs):
                             for j in range(4):
                                     x = np.uint8(cm.viridis(prob.numpy()[j*16]) * 255)
@@ -117,20 +117,20 @@ def main(cfg):
                                 x = np.uint8(cm.viridis(prob.numpy()[j*16]) * 255)
                                 writer.add_image('Attention_dec_%d_0'%global_step, x, i*4+j, dataformats="HWC")
                                 
-                if cfg.use_data_parallel:
+                if args.use_data_parallel:
                     loss = model.scale_loss(loss)
                     loss.backward()
                     model.apply_collective_grads()
                 else:
                     loss.backward()
-                optimizer.minimize(loss, grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(cfg.grad_clip_thresh))
+                optimizer.minimize(loss, grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(cfg['grad_clip_thresh']))
                 model.clear_gradients()
                 
                 # save checkpoint
-                if local_rank==0 and global_step % cfg.save_step == 0:
+                if local_rank==0 and global_step % args.save_step == 0:
-                    if not os.path.exists(cfg.save_path):
+                    if not os.path.exists(args.save_path):
-                        os.mkdir(cfg.save_path)
+                        os.mkdir(args.save_path)
-                    save_path = os.path.join(cfg.save_path,'transformer/%d' % global_step)
+                    save_path = os.path.join(args.save_path,'transformer/%d' % global_step)
                     dg.save_dygraph(model.state_dict(), save_path)
                     dg.save_dygraph(optimizer.state_dict(), save_path)
         if local_rank==0:
@@ -138,7 +138,10 @@ def main(cfg):
                     
 
 if __name__ =='__main__':
-    parser = jsonargparse.ArgumentParser(description="Train TransformerTTS model", formatter_class='default_argparse')
+    parser = argparse.ArgumentParser(description="Train TransformerTTS model")
     add_config_options_to_parser(parser)
-    cfg = parser.parse_args('-c ./config/train_transformer.yaml'.split())
+
-    main(cfg)
+    args = parser.parse_args()
+    # Print the whole config setting.
+    pprint(args)
+    main(args)

examples/transformer_tts/train_vocoder.py

@@ -3,14 +3,15 @@ import os
 from tqdm import tqdm
 from pathlib import Path
 from collections import OrderedDict
-import jsonargparse
+import argparse
+from ruamel import yaml
 from parse import add_config_options_to_parser
 from pprint import pprint
 import paddle.fluid as fluid
 import paddle.fluid.dygraph as dg
 import paddle.fluid.layers as layers
 from parakeet.models.dataloader.ljspeech import LJSpeechLoader
-from parakeet.models.transformerTTS.vocoder import Vocoder
+from parakeet.models.transformer_tts.vocoder import Vocoder
 
 def load_checkpoint(step, model_path):
     model_dict, opti_dict = dg.load_dygraph(os.path.join(model_path, step))
@@ -22,48 +23,47 @@ def load_checkpoint(step, model_path):
             new_state_dict[param] = model_dict[param]
     return new_state_dict, opti_dict
 
-def main(cfg):
+def main(args):
     
-    local_rank = dg.parallel.Env().local_rank if cfg.use_data_parallel else 0
+    local_rank = dg.parallel.Env().local_rank if args.use_data_parallel else 0
-    nranks = dg.parallel.Env().nranks if cfg.use_data_parallel else 1
+    nranks = dg.parallel.Env().nranks if args.use_data_parallel else 1
-    
+
-    if local_rank == 0:
+    with open(args.config_path) as f:
-        # Print the whole config setting.
+        cfg = yaml.load(f, Loader=yaml.Loader)
-        pprint(jsonargparse.namespace_to_dict(cfg))
 
     global_step = 0
     place = (fluid.CUDAPlace(dg.parallel.Env().dev_id)
-             if cfg.use_data_parallel else fluid.CUDAPlace(0)
+             if args.use_data_parallel else fluid.CUDAPlace(0)
-             if cfg.use_gpu else fluid.CPUPlace())
+             if args.use_gpu else fluid.CPUPlace())
     
-    if not os.path.exists(cfg.log_dir):
+    if not os.path.exists(args.log_dir):
-            os.mkdir(cfg.log_dir)
+            os.mkdir(args.log_dir)
-    path = os.path.join(cfg.log_dir,'postnet')
+    path = os.path.join(args.log_dir,'postnet')
 
     writer = SummaryWriter(path) if local_rank == 0 else None
 
     with dg.guard(place):   
-        model = Vocoder(cfg)
+        model = Vocoder(cfg, args.batch_size)
 
         model.train()
-        optimizer = fluid.optimizer.AdamOptimizer(learning_rate=dg.NoamDecay(1/(cfg.warm_up_step *( cfg.lr ** 2)), cfg.warm_up_step),
+        optimizer = fluid.optimizer.AdamOptimizer(learning_rate=dg.NoamDecay(1/(cfg['warm_up_step'] *( args.lr ** 2)), cfg['warm_up_step']),
                                                   parameter_list=model.parameters())
 
 
-        if cfg.checkpoint_path is not None:
+        if args.checkpoint_path is not None:
-            model_dict, opti_dict = load_checkpoint(str(cfg.postnet_step), os.path.join(cfg.checkpoint_path, "postnet"))
+            model_dict, opti_dict = load_checkpoint(str(args.vocoder_step), os.path.join(args.checkpoint_path, "postnet"))
             model.set_dict(model_dict)
             optimizer.set_dict(opti_dict)
-            global_step = cfg.postnet_step
+            global_step = args.vocoder_step
             print("load checkpoint!!!")
 
-        if cfg.use_data_parallel:
+        if args.use_data_parallel:
             strategy = dg.parallel.prepare_context()
             model = fluid.dygraph.parallel.DataParallel(model, strategy)
 
-        reader = LJSpeechLoader(cfg, nranks, local_rank, is_vocoder=True).reader()
+        reader = LJSpeechLoader(cfg, args, nranks, local_rank, is_vocoder=True).reader()
 
-        for epoch in range(cfg.epochs):
+        for epoch in range(args.epochs):
             pbar = tqdm(reader)
             for i, data in enumerate(pbar):
                 pbar.set_description('Processing at epoch %d'%epoch)
@@ -75,13 +75,13 @@ def main(cfg):
                 mag_pred = model(mel)
                 loss = layers.mean(layers.abs(layers.elementwise_sub(mag_pred, mag)))
                 
-                if cfg.use_data_parallel:
+                if args.use_data_parallel:
                     loss = model.scale_loss(loss)
                     loss.backward()
                     model.apply_collective_grads()
                 else:
                     loss.backward()
-                optimizer.minimize(loss, grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(cfg.grad_clip_thresh))
+                optimizer.minimize(loss, grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(cfg['grad_clip_thresh']))
                 model.clear_gradients()
                 
                 if local_rank==0:
@@ -89,10 +89,10 @@ def main(cfg):
                         'loss':loss.numpy(),
                     }, global_step)
 
-                    if global_step % cfg.save_step == 0:
+                    if global_step % args.save_step == 0:
-                        if not os.path.exists(cfg.save_path):
+                        if not os.path.exists(args.save_path):
-                            os.mkdir(cfg.save_path)
+                            os.mkdir(args.save_path)
-                        save_path = os.path.join(cfg.save_path,'postnet/%d' % global_step)
+                        save_path = os.path.join(args.save_path,'postnet/%d' % global_step)
                         dg.save_dygraph(model.state_dict(), save_path)
                         dg.save_dygraph(optimizer.state_dict(), save_path)
 
@@ -100,7 +100,9 @@ def main(cfg):
             writer.close()
 
 if __name__ == '__main__':
-    parser = jsonargparse.ArgumentParser(description="Train postnet model", formatter_class='default_argparse')
+    parser = argparse.ArgumentParser(description="Train postnet model")
     add_config_options_to_parser(parser)
-    cfg = parser.parse_args('-c ./config/train_vocoder.yaml'.split())
+    args = parser.parse_args()
-    main(cfg)
+    # Print the whole config setting.
+    pprint(args)
+    main(args)

parakeet/models/dataloader/ljspeech.py

@@ -13,17 +13,17 @@ from parakeet.data.batch import TextIDBatcher, SpecBatcher
 from parakeet.data.dataset import DatasetMixin, TransformDataset
 
 class LJSpeechLoader:
-    def __init__(self, config, nranks, rank, is_vocoder=False, shuffle=True):
+    def __init__(self, config, args, nranks, rank, is_vocoder=False, shuffle=True):
-        place = fluid.CUDAPlace(rank) if config.use_gpu else fluid.CPUPlace()
+        place = fluid.CUDAPlace(rank) if args.use_gpu else fluid.CPUPlace()
 
-        LJSPEECH_ROOT = Path(config.data_path)
+        LJSPEECH_ROOT = Path(args.data_path)
         metadata = LJSpeechMetaData(LJSPEECH_ROOT)
         transformer = LJSpeech(config)
         dataset = TransformDataset(metadata, transformer)
         sampler = DistributedSampler(len(metadata), nranks, rank, shuffle=shuffle)
 
-        assert config.batch_size % nranks == 0
+        assert args.batch_size % nranks == 0
-        each_bs = config.batch_size // nranks
+        each_bs = args.batch_size // nranks
         if is_vocoder:
             dataloader = DataCargo(dataset, sampler=sampler, batch_size=each_bs, shuffle=shuffle, batch_fn=batch_examples_vocoder, drop_last=True)
         else:
@@ -63,15 +63,15 @@ class LJSpeech(object):
         super(LJSpeech, self).__init__()
         self.config = config
         self._ljspeech_processor = audio.AudioProcessor(
-            sample_rate=config.audio.sr, 
+            sample_rate=config['audio']['sr'], 
-            num_mels=config.audio.num_mels, 
+            num_mels=config['audio']['num_mels'], 
-            min_level_db=config.audio.min_level_db, 
+            min_level_db=config['audio']['min_level_db'], 
-            ref_level_db=config.audio.ref_level_db, 
+            ref_level_db=config['audio']['ref_level_db'], 
-            n_fft=config.audio.n_fft, 
+            n_fft=config['audio']['n_fft'], 
-            win_length= config.audio.win_length, 
+            win_length= config['audio']['win_length'], 
-            hop_length= config.audio.hop_length,
+            hop_length= config['audio']['hop_length'],
-            power=config.audio.power,
+            power=config['audio']['power'],
-            preemphasis=config.audio.preemphasis,
+            preemphasis=config['audio']['preemphasis'],
             signal_norm=True,
             symmetric_norm=False,
             max_norm=1.,

parakeet/models/fastspeech/fastspeech.py

@@ -2,7 +2,7 @@ import math
 import paddle.fluid.dygraph as dg
 import paddle.fluid as fluid
 from parakeet.g2p.text.symbols import symbols
-from parakeet.models.transformerTTS.post_convnet import PostConvNet
+from parakeet.models.transformer_tts.post_convnet import PostConvNet
 from parakeet.models.fastspeech.LengthRegulator import LengthRegulator
 from parakeet.models.fastspeech.encoder import Encoder
 from parakeet.models.fastspeech.decoder import Decoder
@@ -13,43 +13,43 @@ class FastSpeech(dg.Layer):
         super(FastSpeech, self).__init__()
 
         self.encoder = Encoder(n_src_vocab=len(symbols)+1,
-                               len_max_seq=cfg.max_sep_len,
+                               len_max_seq=cfg['max_seq_len'],
-                               n_layers=cfg.encoder_n_layer,
+                               n_layers=cfg['encoder_n_layer'],
-                               n_head=cfg.encoder_head,
+                               n_head=cfg['encoder_head'],
-                               d_k=cfg.fs_hidden_size // cfg.encoder_head,
+                               d_k=cfg['fs_hidden_size'] // cfg['encoder_head'],
-                               d_v=cfg.fs_hidden_size // cfg.encoder_head,
+                               d_v=cfg['fs_hidden_size'] // cfg['encoder_head'],
-                               d_model=cfg.fs_hidden_size,
+                               d_model=cfg['fs_hidden_size'],
-                               d_inner=cfg.encoder_conv1d_filter_size,
+                               d_inner=cfg['encoder_conv1d_filter_size'],
-                               fft_conv1d_kernel=cfg.fft_conv1d_filter, 
+                               fft_conv1d_kernel=cfg['fft_conv1d_filter'], 
-                               fft_conv1d_padding=cfg.fft_conv1d_padding,
+                               fft_conv1d_padding=cfg['fft_conv1d_padding'],
                                dropout=0.1)
-        self.length_regulator = LengthRegulator(input_size=cfg.fs_hidden_size, 
+        self.length_regulator = LengthRegulator(input_size=cfg['fs_hidden_size'], 
-                                                out_channels=cfg.duration_predictor_output_size, 
+                                                out_channels=cfg['duration_predictor_output_size'], 
-                                                filter_size=cfg.duration_predictor_filter_size, 
+                                                filter_size=cfg['duration_predictor_filter_size'], 
-                                                dropout=cfg.dropout)
+                                                dropout=cfg['dropout'])
-        self.decoder = Decoder(len_max_seq=cfg.max_sep_len,
+        self.decoder = Decoder(len_max_seq=cfg['max_seq_len'],
-                                n_layers=cfg.decoder_n_layer,
+                                n_layers=cfg['decoder_n_layer'],
-                                n_head=cfg.decoder_head,
+                                n_head=cfg['decoder_head'],
-                                d_k=cfg.fs_hidden_size // cfg.decoder_head,
+                                d_k=cfg['fs_hidden_size'] // cfg['decoder_head'],
-                                d_v=cfg.fs_hidden_size // cfg.decoder_head,
+                                d_v=cfg['fs_hidden_size'] // cfg['decoder_head'],
-                                d_model=cfg.fs_hidden_size,
+                                d_model=cfg['fs_hidden_size'],
-                                d_inner=cfg.decoder_conv1d_filter_size,
+                                d_inner=cfg['decoder_conv1d_filter_size'],
-                                fft_conv1d_kernel=cfg.fft_conv1d_filter, 
+                                fft_conv1d_kernel=cfg['fft_conv1d_filter'], 
-                                fft_conv1d_padding=cfg.fft_conv1d_padding,
+                                fft_conv1d_padding=cfg['fft_conv1d_padding'],
                                 dropout=0.1)
         self.weight = fluid.ParamAttr(initializer = fluid.initializer.XavierInitializer())
-        k = math.sqrt(1 / cfg.fs_hidden_size)
+        k = math.sqrt(1 / cfg['fs_hidden_size'])
         self.bias = fluid.ParamAttr(initializer = fluid.initializer.Uniform(low=-k, high=k))
-        self.mel_linear = dg.Linear(cfg.fs_hidden_size, 
+        self.mel_linear = dg.Linear(cfg['fs_hidden_size'], 
-                                    cfg.audio.num_mels * cfg.audio.outputs_per_step,
+                                    cfg['audio']['num_mels']* cfg['audio']['outputs_per_step'],
                                     param_attr = self.weight,
                                     bias_attr = self.bias,)
-        self.postnet = PostConvNet(n_mels=cfg.audio.num_mels,
+        self.postnet = PostConvNet(n_mels=cfg['audio']['num_mels'],
                  num_hidden=512,
                  filter_size=5,
                  padding=int(5 / 2),
                  num_conv=5,
-                 outputs_per_step=cfg.audio.outputs_per_step,
+                 outputs_per_step=cfg['audio']['outputs_per_step'],
                  use_cudnn=True,
                  dropout=0.1,
                  batchnorm_last=True)

parakeet/models/transformer_tts/decoder.py

@@ -4,8 +4,8 @@ import paddle.fluid as fluid
 from parakeet.modules.utils import *
 from parakeet.modules.multihead_attention import MultiheadAttention
 from parakeet.modules.ffn import PositionwiseFeedForward
-from parakeet.models.transformerTTS.prenet import PreNet
+from parakeet.models.transformer_tts.prenet import PreNet
-from parakeet.models.transformerTTS.post_convnet import PostConvNet
+from parakeet.models.transformer_tts.post_convnet import PostConvNet
 
 class Decoder(dg.Layer):
     def __init__(self, num_hidden, config, num_head=4):
@@ -20,7 +20,7 @@ class Decoder(dg.Layer):
                                  param_attr=fluid.ParamAttr(
                                      initializer=fluid.initializer.NumpyArrayInitializer(self.pos_inp),
                                      trainable=False))
-        self.decoder_prenet = PreNet(input_size = config.audio.num_mels, 
+        self.decoder_prenet = PreNet(input_size = config['audio']['num_mels'], 
                                             hidden_size = num_hidden * 2, 
                                             output_size = num_hidden, 
                                             dropout_rate=0.2)
@@ -38,17 +38,17 @@ class Decoder(dg.Layer):
         self.ffns = [PositionwiseFeedForward(num_hidden, num_hidden*num_head, filter_size=1) for _ in range(3)]
         for i, layer in enumerate(self.ffns):
             self.add_sublayer("ffns_{}".format(i), layer)
-        self.mel_linear = dg.Linear(num_hidden, config.audio.num_mels * config.audio.outputs_per_step,
+        self.mel_linear = dg.Linear(num_hidden, config['audio']['num_mels'] * config['audio']['outputs_per_step'],
                                 param_attr=fluid.ParamAttr(initializer = fluid.initializer.XavierInitializer()),
                                 bias_attr=fluid.ParamAttr(initializer = fluid.initializer.Uniform(low=-k, high=k)))
         self.stop_linear = dg.Linear(num_hidden, 1,
                                   param_attr=fluid.ParamAttr(initializer = fluid.initializer.XavierInitializer()),
                                   bias_attr=fluid.ParamAttr(initializer = fluid.initializer.Uniform(low=-k, high=k)))
 
-        self.postconvnet = PostConvNet(config.audio.num_mels, config.hidden_size, 
+        self.postconvnet = PostConvNet(config['audio']['num_mels'], config['hidden_size'], 
                                        filter_size = 5, padding = 4, num_conv=5, 
-                                       outputs_per_step=config.audio.outputs_per_step, 
+                                       outputs_per_step=config['audio']['outputs_per_step'], 
-                                       use_cudnn = config.use_gpu)
+                                       use_cudnn = True)
 
     def forward(self, key, value, query, c_mask, positional):
 

parakeet/models/transformer_tts/encoder.py

@@ -3,10 +3,10 @@ import paddle.fluid as fluid
 from parakeet.modules.utils import *
 from parakeet.modules.multihead_attention import MultiheadAttention
 from parakeet.modules.ffn import PositionwiseFeedForward
-from parakeet.models.transformerTTS.encoderprenet import EncoderPrenet
+from parakeet.models.transformer_tts.encoderprenet import EncoderPrenet
 
 class Encoder(dg.Layer):
-    def __init__(self, embedding_size, num_hidden, config, num_head=4):
+    def __init__(self, embedding_size, num_hidden, num_head=4):
         super(Encoder, self).__init__()
         self.num_hidden = num_hidden
         param = fluid.ParamAttr(initializer=fluid.initializer.Constant(value=1.0))
@@ -19,11 +19,11 @@ class Encoder(dg.Layer):
                                      trainable=False))
         self.encoder_prenet = EncoderPrenet(embedding_size = embedding_size, 
                                             num_hidden = num_hidden, 
-                                            use_cudnn=config.use_gpu)
+                                            use_cudnn=True)
         self.layers = [MultiheadAttention(num_hidden, num_hidden//num_head, num_hidden//num_head) for _ in range(3)]
         for i, layer in enumerate(self.layers):
             self.add_sublayer("self_attn_{}".format(i), layer)
-        self.ffns = [PositionwiseFeedForward(num_hidden, num_hidden*num_head, filter_size=1, use_cudnn = config.use_gpu) for _ in range(3)]
+        self.ffns = [PositionwiseFeedForward(num_hidden, num_hidden*num_head, filter_size=1, use_cudnn = True) for _ in range(3)]
         for i, layer in enumerate(self.ffns):
             self.add_sublayer("ffns_{}".format(i), layer)
 

parakeet/models/transformer_tts/transformerTTS.py

@@ -1,13 +1,13 @@
 import paddle.fluid.dygraph as dg
 import paddle.fluid as fluid
-from parakeet.models.transformerTTS.encoder import Encoder
+from parakeet.models.transformer_tts.encoder import Encoder
-from parakeet.models.transformerTTS.decoder import Decoder
+from parakeet.models.transformer_tts.decoder import Decoder
 
 class TransformerTTS(dg.Layer):
     def __init__(self, config):
         super(TransformerTTS, self).__init__()
-        self.encoder = Encoder(config.embedding_size, config.hidden_size, config)
+        self.encoder = Encoder(config['embedding_size'], config['hidden_size'])
-        self.decoder = Decoder(config.hidden_size, config)
+        self.decoder = Decoder(config['hidden_size'], config)
         self.config = config
 
     def forward(self, characters, mel_input, pos_text, pos_mel):

parakeet/models/transformer_tts/vocoder.py

@@ -2,20 +2,20 @@ import paddle.fluid.dygraph as dg
 import paddle.fluid as fluid
 from parakeet.modules.customized import Conv1D
 from parakeet.modules.utils import *
-from parakeet.models.transformerTTS.CBHG import CBHG
+from parakeet.models.transformer_tts.CBHG import CBHG
 
 class Vocoder(dg.Layer):
     """
     CBHG Network (mel -> linear)
     """
-    def __init__(self, config):
+    def __init__(self, config, batch_size):
         super(Vocoder, self).__init__()
-        self.pre_proj = Conv1D(num_channels = config.audio.num_mels, 
+        self.pre_proj = Conv1D(num_channels = config['audio']['num_mels'], 
-                             num_filters = config.hidden_size,
+                             num_filters = config['hidden_size'],
                              filter_size=1)
-        self.cbhg = CBHG(config.hidden_size, config.batch_size)
+        self.cbhg = CBHG(config['hidden_size'], batch_size)
-        self.post_proj = Conv1D(num_channels = config.hidden_size, 
+        self.post_proj = Conv1D(num_channels = config['hidden_size'], 
-                             num_filters = (config.audio.n_fft // 2) + 1,
+                             num_filters = (config['audio']['n_fft'] // 2) + 1,
                              filter_size=1)
 
     def forward(self, mel):