forked from lturing/tacotronv2_wavernn_chinese
-
Notifications
You must be signed in to change notification settings - Fork 0
/
tacotron_hparams.py
244 lines (209 loc) · 18.6 KB
/
tacotron_hparams.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
import numpy as np
import tensorflow as tf
# Default hyperparameters
hparams = tf.contrib.training.HParams(
# Comma-separated list of cleaners to run on text prior to training and eval. For non-English
# text, you may want to use "basic_cleaners" or "transliteration_cleaners".
dataset = '/home/spurs/tts/dataset/bznsyp',
base_dir = '/home/spurs/tts/dataset',
feat_out_dir = 'training_data_v1',
#tacotron_input = '/home/spurs/tts/dataset/bznsyp/training_data_v1/train.txt',
tacotron_input = './train.txt',
#If you only have 1 GPU or want to use only one GPU, please set num_gpus=0 and specify the GPU idx on run. example:
#expample 1 GPU of index 2 (train on "/gpu2" only): CUDA_VISIBLE_DEVICES=2 python train.py --model='Tacotron' --hparams='tacotron_gpu_start_idx=2'
#If you want to train on multiple GPUs, simply specify the number of GPUs available, and the idx of the first GPU to use. example:
#example 4 GPUs starting from index 0 (train on "/gpu0"->"/gpu3"): python train.py --model='Tacotron' --hparams='tacotron_num_gpus=4, tacotron_gpu_start_idx=0'
#The hparams arguments can be directly modified on this hparams.py file instead of being specified on run if preferred!
#If one wants to train both Tacotron and WaveNet in parallel (provided WaveNet will be trained on True mel spectrograms), one needs to specify different GPU idxes.
#example Tacotron+WaveNet on a machine with 4 or more GPUs. Two GPUs for each model:
# CUDA_VISIBLE_DEVICES=0,1 python train.py --model='Tacotron' --hparams='tacotron_num_gpus=2'
# Cuda_VISIBLE_DEVICES=2,3 python train.py --model='WaveNet' --hparams='wavenet_num_gpus=2'
#IMPORTANT NOTES: The Multi-GPU performance highly depends on your hardware and optimal parameters change between rigs. Default are optimized for servers.
#If using N GPUs, please multiply the tacotron_batch_size by N below in the hparams! (tacotron_batch_size = 32 * N)
#Never use lower batch size than 32 on a single GPU!
#Same applies for Wavenet: wavenet_batch_size = 8 * N (wavenet_batch_size can be smaller than 8 if GPU is having OOM, minimum 2)
#Please also apply the synthesis batch size modification likewise. (if N GPUs are used for synthesis, minimal batch size must be N, minimum of 1 sample per GPU)
#We did not add an automatic multi-GPU batch size computation to avoid confusion in the user's mind and to provide more control to the user for
#resources related decisions.
#Acknowledgement:
# Many thanks to @MlWoo for his awesome work on multi-GPU Tacotron which showed to work a little faster than the original
# pipeline for a single GPU as well. Great work!
#Hardware setup: Default supposes user has only one GPU: "/gpu:0" (Both Tacotron and WaveNet can be trained on multi-GPU: data parallelization)
#Synthesis also uses the following hardware parameters for multi-GPU parallel synthesis.
tacotron_num_gpus = 1, #Determines the number of gpus in use for Tacotron training.
wavenet_num_gpus = 1, #Determines the number of gpus in use for WaveNet training.
split_on_cpu = True, #Determines whether to split data on CPU or on first GPU. This is automatically True when more than 1 GPU is used.
#(Recommend: False on slow CPUs/Disks, True otherwise for small speed boost)
###########################################################################################################################################
#Audio
#Audio parameters are the most important parameters to tune when using this work on your personal data. Below are the beginner steps to adapt
#this work to your personal data:
# 1- Determine my data sample rate: First you need to determine your audio sample_rate (how many samples are in a second of audio). This can be done using sox: "sox --i <filename>"
# (For this small tuto, I will consider 24kHz (24000 Hz), and defaults are 22050Hz, so there are plenty of examples to refer to)
# 2- set sample_rate parameter to your data correct sample rate
# 3- Fix win_size and and hop_size accordingly: (Supposing you will follow our advice: 50ms window_size, and 12.5ms frame_shift(hop_size))
# a- win_size = 0.05 * sample_rate. In the tuto example, 0.05 * 24000 = 1200
# b- hop_size = 0.25 * win_size. Also equal to 0.0125 * sample_rate. In the tuto example, 0.25 * 1200 = 0.0125 * 24000 = 300 (Can set frame_shift_ms=12.5 instead)
# 4- Fix n_fft, num_freq and upsample_scales parameters accordingly.
# a- n_fft can be either equal to win_size or the first power of 2 that comes after win_size. I usually recommend using the latter
# to be more consistent with signal processing friends. No big difference to be seen however. For the tuto example: n_fft = 2048 = 2**11
# b- num_freq = (n_fft / 2) + 1. For the tuto example: num_freq = 2048 / 2 + 1 = 1024 + 1 = 1025.
# c- For WaveNet, upsample_scales products must be equal to hop_size. For the tuto example: upsample_scales=[15, 20] where 15 * 20 = 300
# it is also possible to use upsample_scales=[3, 4, 5, 5] instead. One must only keep in mind that upsample_kernel_size[0] = 2*upsample_scales[0]
# so the training segments should be long enough (2.8~3x upsample_scales[0] * hop_size or longer) so that the first kernel size can see the middle
# of the samples efficiently. The length of WaveNet training segments is under the parameter "max_time_steps".
# 5- Finally comes the silence trimming. This very much data dependent, so I suggest trying preprocessing (or part of it, ctrl-C to stop), then use the
# .ipynb provided in the repo to listen to some inverted mel/linear spectrograms. That will first give you some idea about your above parameters, and
# it will also give you an idea about trimming. If silences persist, try reducing trim_top_db slowly. If samples are trimmed mid words, try increasing it.
# 6- If audio quality is too metallic or fragmented (or if linear spectrogram plots are showing black silent regions on top), then restart from step 2.
num_mels = 80, #Number of mel-spectrogram channels and local conditioning dimensionality
num_freq = 1025, # (= n_fft / 2 + 1) only used when adding linear spectrograms post processing network
rescale = True, #Whether to rescale audio prior to preprocessing
rescaling_max = 0.999, #Rescaling value
#train samples of lengths between 3sec and 14sec are more than enough to make a model capable of generating consistent speech.
clip_mels_length = False, #For cases of OOM (Not really recommended, only use if facing unsolvable OOM errors, also consider clipping your samples to smaller chunks)
max_mel_frames = 900, #Only relevant when clip_mels_length = True, please only use after trying output_per_steps=3 and still getting OOM errors.
# Use LWS (https://github.com/Jonathan-LeRoux/lws) for STFT and phase reconstruction
# It's preferred to set True to use with https://github.com/r9y9/wavenet_vocoder
# Does not work if n_ffit is not multiple of hop_size!!
use_lws=False, #Only used to set as True if using WaveNet, no difference in performance is observed in either cases.
silence_threshold=2, #silence threshold used for sound trimming for wavenet preprocessing
#Mel spectrogram
n_fft = 2048, #Extra window size is filled with 0 paddings to match this parameter
hop_size = 275, #For 22050Hz, 275 ~= 12.5 ms (0.0125 * sample_rate)
win_size = 1100, #For 22050Hz, 1100 ~= 50 ms (If None, win_size = n_fft) (0.05 * sample_rate)
sample_rate = 22050, #22050 Hz (corresponding to ljspeech dataset) (sox --i <filename>)
frame_shift_ms = None, #Can replace hop_size parameter. (Recommended: 12.5)
magnitude_power = 2., #The power of the spectrogram magnitude (1. for energy, 2. for power)
#M-AILABS (and other datasets) trim params (there parameters are usually correct for any data, but definitely must be tuned for specific speakers)
trim_silence = True, #Whether to clip silence in Audio (at beginning and end of audio only, not the middle)
trim_fft_size = 2048, #Trimming window size
trim_hop_size = 512, #Trimmin hop length
trim_top_db = 25, #Trimming db difference from reference db (smaller==harder trim.)
#Mel and Linear spectrograms normalization/scaling and clipping
signal_normalization = True, #Whether to normalize mel spectrograms to some predefined range (following below parameters)
allow_clipping_in_normalization = True, #Only relevant if mel_normalization = True
symmetric_mels = True, #Whether to scale the data to be symmetric around 0. (Also multiplies the output range by 2, faster and cleaner convergence)
max_abs_value = 4., #max absolute value of data. If symmetric, data will be [-max, max] else [0, max] (Must not be too big to avoid gradient explosion,
#not too small for fast convergence)
normalize_for_wavenet = True, #whether to rescale to [0, 1] for wavenet. (better audio quality)
clip_for_wavenet = True, #whether to clip [-max, max] before training/synthesizing with wavenet (better audio quality)
wavenet_pad_sides = 1, #Can be 1 or 2. 1 for pad right only, 2 for both sides padding.
#Contribution by @begeekmyfriend
#Spectrogram Pre-Emphasis (Lfilter: Reduce spectrogram noise and helps model certitude levels. Also allows for better G&L phase reconstruction)
preemphasize = True, #whether to apply filter
preemphasis = 0.97, #filter coefficient.
#Limits
min_level_db = -100,
ref_level_db = 20,
fmin = 95, #Set this to 55 if your speaker is male! if female, 95 should help taking off noise. (To test depending on dataset. Pitch info: male~[65, 260], female~[100, 525])
fmax = 7600, #To be increased/reduced depending on data.
#Griffin Lim
power = 1.5, #Only used in G&L inversion, usually values between 1.2 and 1.5 are a good choice.
griffin_lim_iters = 60, #Number of G&L iterations, typically 30 is enough but we use 60 to ensure convergence.
GL_on_GPU = False, #Whether to use G&L GPU version as part of tensorflow graph. (Usually much faster than CPU but slightly worse quality too).
###########################################################################################################################################
#Tacotron
#Model general type
outputs_per_step = 1, #number of frames to generate at each decoding step (increase to speed up computation and allows for higher batch size, decreases G&L audio quality)
use_all_outputs = False,
stop_at_any = True, #Determines whether the decoder should stop when predicting <stop> to any frame or to all of them (True works pretty well)
batch_norm_position = 'after', #Can be in ('before', 'after'). Determines whether we use batch norm before or after the activation function (relu). Matter for debate.
clip_outputs = True, #Whether to clip spectrograms to T2_output_range (even in loss computation). ie: Don't penalize model for exceeding output range and bring back to borders.
lower_bound_decay = 0.1, #Small regularizer for noise synthesis by adding small range of penalty for silence regions. Set to 0 to clip in Tacotron range.
#Input parameters
embedding_dim = 128, #dimension of embedding space
#Encoder parameters
enc_conv_num_layers = 3, #number of encoder convolutional layers
enc_conv_kernel_size = (5, ), #size of encoder convolution filters for each layer
enc_conv_channels = 256, #number of encoder convolutions filters for each layer
encoder_lstm_units = 256, #number of lstm units for each direction (forward and backward)
#Attention mechanism
smoothing = False, #Whether to smooth the attention normalization function
attention_dim = 128, #dimension of attention space
attention_filters = 32, #number of attention convolution filters
attention_kernel = (31, ), #kernel size of attention convolution
cumulative_weights = True, #Whether to cumulate (sum) all previous attention weights or simply feed previous weights (Recommended: True)
#Attention synthesis constraints
#"Monotonic" constraint forces the model to only look at the forwards attention_win_size steps.
#"Window" allows the model to look at attention_win_size neighbors, both forward and backward steps.
synthesis_constraint = False, #Whether to use attention windows constraints in synthesis only (Useful for long utterances synthesis)
synthesis_constraint_type = 'window', #can be in ('window', 'monotonic').
attention_win_size = 2, #Side of the window. Current step does not count. If mode is window and attention_win_size is not pair, the 1 extra is provided to backward part of the window.
#Decoder
prenet_layers = [256, 256], #number of layers and number of units of prenet
decoder_layers = 2, #number of decoder lstm layers
decoder_lstm_units = 256, #number of decoder lstm units on each layer
max_iters = 2000, #Max decoder steps during inference (Just for safety from infinite loop cases)
#Residual postnet
postnet_num_layers = 5, #number of postnet convolutional layers
postnet_kernel_size = (5, ), #size of postnet convolution filters for each layer
postnet_channels = 256, #number of postnet convolution filters for each layer
#CBHG mel->linear postnet
cbhg_kernels = 8, #All kernel sizes from 1 to cbhg_kernels will be used in the convolution bank of CBHG to act as "K-grams"
cbhg_conv_channels = 128, #Channels of the convolution bank
cbhg_pool_size = 2, #pooling size of the CBHG
cbhg_projection = 256, #projection channels of the CBHG (1st projection, 2nd is automatically set to num_mels)
cbhg_projection_kernel_size = 3, #kernel_size of the CBHG projections
cbhg_highwaynet_layers = 4, #Number of HighwayNet layers
cbhg_highway_units = 128, #Number of units used in HighwayNet fully connected layers
cbhg_rnn_units = 128, #Number of GRU units used in bidirectional RNN of CBHG block. CBHG output is 2x rnn_units in shape
#Loss params
mask_encoder = True, #whether to mask encoder padding while computing attention. Set to True for better prosody but slower convergence.
mask_decoder = False, #Whether to use loss mask for padded sequences (if False, <stop_token> loss function will not be weighted, else recommended pos_weight = 20)
cross_entropy_pos_weight = 1, #Use class weights to reduce the stop token classes imbalance (by adding more penalty on False Negatives (FN)) (1 = disabled)
predict_linear = False, #Whether to add a post-processing network to the Tacotron to predict linear spectrograms (True mode Not tested!!)
###########################################################################################################################################
#Tacotron Training
#Reproduction seeds
tacotron_random_seed = 5339, #Determines initial graph and operations (i.e: model) random state for reproducibility
tacotron_data_random_state = 1234, #random state for train test split repeatability
#performance parameters
tacotron_swap_with_cpu = False, #Whether to use cpu as support to gpu for decoder computation (Not recommended: may cause major slowdowns! Only use when critical!)
#train/test split ratios, mini-batches sizes
tacotron_batch_size = 32, #number of training samples on each training steps
#Tacotron Batch synthesis supports ~16x the training batch size (no gradients during testing).
#Training Tacotron with unmasked paddings makes it aware of them, which makes synthesis times different from training. We thus recommend masking the encoder.
tacotron_synthesis_batch_size = 1, #DO NOT MAKE THIS BIGGER THAN 1 IF YOU DIDN'T TRAIN TACOTRON WITH "mask_encoder=True"!!
tacotron_test_size = 0.05, #% of data to keep as test data, if None, tacotron_test_batches must be not None. (5% is enough to have a good idea about overfit)
tacotron_test_batches = None, #number of test batches.
#Learning rate schedule
tacotron_decay_learning_rate = True, #boolean, determines if the learning rate will follow an exponential decay
tacotron_start_decay = 66000, #Step at which learning decay starts
tacotron_decay_steps = 20000, #Determines the learning rate decay slope (UNDER TEST)
tacotron_decay_rate = 0.5, #learning rate decay rate (UNDER TEST)
tacotron_initial_learning_rate = 1e-3, #starting learning rate
tacotron_final_learning_rate = 1e-5, #minimal learning rate
#Optimization parameters
tacotron_adam_beta1 = 0.9, #AdamOptimizer beta1 parameter
tacotron_adam_beta2 = 0.999, #AdamOptimizer beta2 parameter
tacotron_adam_epsilon = 1e-6, #AdamOptimizer Epsilon parameter
#Regularization parameters
tacotron_reg_weight = 1e-6, #regularization weight (for L2 regularization)
tacotron_scale_regularization = False, #Whether to rescale regularization weight to adapt for outputs range (used when reg_weight is high and biasing the model)
tacotron_zoneout_rate = 0.1, #zoneout rate for all LSTM cells in the network
tacotron_dropout_rate = 0.5, #dropout rate for all convolutional layers + prenet
tacotron_clip_gradients = True, #whether to clip gradients
#Evaluation parameters
tacotron_natural_eval = True, #Whether to use 100% natural eval (to evaluate Curriculum Learning performance) or with same teacher-forcing ratio as in training (just for overfit)
#Decoder RNN learning can take be done in one of two ways:
# Teacher Forcing: vanilla teacher forcing (usually with ratio = 1). mode='constant'
# Scheduled Sampling Scheme: From Teacher-Forcing to sampling from previous outputs is function of global step. (teacher forcing ratio decay) mode='scheduled'
#The second approach is inspired by:
#Bengio et al. 2015: Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks.
#Can be found under: https://arxiv.org/pdf/1506.03099.pdf
tacotron_teacher_forcing_mode = 'constant', #Can be ('constant' or 'scheduled'). 'scheduled' mode applies a cosine teacher forcing ratio decay. (Preference: scheduled)
tacotron_teacher_forcing_ratio = 1., #Value from [0., 1.], 0.=0%, 1.=100%, determines the % of times we force next decoder inputs, Only relevant if mode='constant'
tacotron_teacher_forcing_init_ratio = 1., #initial teacher forcing ratio. Relevant if mode='scheduled'
tacotron_teacher_forcing_final_ratio = 0.3, #final teacher forcing ratio. (Set None to use alpha instead) Relevant if mode='scheduled'
tacotron_teacher_forcing_start_decay = 70000, #starting point of teacher forcing ratio decay. Relevant if mode='scheduled'
tacotron_teacher_forcing_decay_steps = 150000, #Determines the teacher forcing ratio decay slope. Relevant if mode='scheduled'
tacotron_teacher_forcing_decay_alpha = None, #teacher forcing ratio decay rate. Defines the final tfr as a ratio of initial tfr. Relevant if mode='scheduled'
#Speaker adaptation parameters
tacotron_fine_tuning = False, #Set to True to freeze encoder and only keep training pretrained decoder. Used for speaker adaptation with small data.
###########################################################################################################################################
)
def hparams_debug_string():
values = hparams.values()
hp = [' %s: %s' % (name, values[name]) for name in sorted(values) if name != 'sentences']
return 'Hyperparameters:\n' + '\n'.join(hp)