dathudeptrai commited on
Commit
14292f0
1 Parent(s): d5efa8b

🍒 Upload config, checkpoint for mb_melgan Chinese.

Browse files
Files changed (2) hide show
  1. config.yml +106 -0
  2. model.h5 +3 -0
config.yml ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # This is the hyperparameter configuration file for Multi-Band MelGAN.
2
+ # Please make sure this is adjusted for the Baker dataset. If you want to
3
+ # apply to the other dataset, you might need to carefully change some parameters.
4
+ # This configuration performs 1000k iters.
5
+
6
+ ###########################################################
7
+ # FEATURE EXTRACTION SETTING #
8
+ ###########################################################
9
+ sampling_rate: 24000
10
+ hop_size: 300 # Hop size.
11
+ format: "npy"
12
+
13
+
14
+ ###########################################################
15
+ # GENERATOR NETWORK ARCHITECTURE SETTING #
16
+ ###########################################################
17
+ model_type: "multiband_melgan_generator"
18
+
19
+ multiband_melgan_generator_params:
20
+ out_channels: 4 # Number of output channels (number of subbands).
21
+ kernel_size: 7 # Kernel size of initial and final conv layers.
22
+ filters: 384 # Initial number of channels for conv layers.
23
+ upsample_scales: [3, 5, 5] # List of Upsampling scales.
24
+ stack_kernel_size: 3 # Kernel size of dilated conv layers in residual stack.
25
+ stacks: 4 # Number of stacks in a single residual stack module.
26
+ is_weight_norm: false # Use weight-norm or not.
27
+
28
+ ###########################################################
29
+ # DISCRIMINATOR NETWORK ARCHITECTURE SETTING #
30
+ ###########################################################
31
+ multiband_melgan_discriminator_params:
32
+ out_channels: 1 # Number of output channels.
33
+ scales: 3 # Number of multi-scales.
34
+ downsample_pooling: "AveragePooling1D" # Pooling type for the input downsampling.
35
+ downsample_pooling_params: # Parameters of the above pooling function.
36
+ pool_size: 4
37
+ strides: 2
38
+ kernel_sizes: [5, 3] # List of kernel size.
39
+ filters: 16 # Number of channels of the initial conv layer.
40
+ max_downsample_filters: 512 # Maximum number of channels of downsampling layers.
41
+ downsample_scales: [4, 4, 4] # List of downsampling scales.
42
+ nonlinear_activation: "LeakyReLU" # Nonlinear activation function.
43
+ nonlinear_activation_params: # Parameters of nonlinear activation function.
44
+ alpha: 0.2
45
+ is_weight_norm: false # Use weight-norm or not.
46
+
47
+ ###########################################################
48
+ # STFT LOSS SETTING #
49
+ ###########################################################
50
+ stft_loss_params:
51
+ fft_lengths: [1024, 2048, 512] # List of FFT size for STFT-based loss.
52
+ frame_steps: [120, 240, 50] # List of hop size for STFT-based loss
53
+ frame_lengths: [600, 1200, 240] # List of window length for STFT-based loss.
54
+
55
+ subband_stft_loss_params:
56
+ fft_lengths: [384, 683, 171] # List of FFT size for STFT-based loss.
57
+ frame_steps: [30, 60, 10] # List of hop size for STFT-based loss
58
+ frame_lengths: [150, 300, 60] # List of window length for STFT-based loss.
59
+
60
+ ###########################################################
61
+ # ADVERSARIAL LOSS SETTING #
62
+ ###########################################################
63
+ lambda_feat_match: 10.0 # Loss balancing coefficient for feature matching loss
64
+ lambda_adv: 2.5 # Loss balancing coefficient for adversarial loss.
65
+
66
+ ###########################################################
67
+ # DATA LOADER SETTING #
68
+ ###########################################################
69
+ batch_size: 64 # Batch size for each GPU with assuming that gradient_accumulation_steps == 1.
70
+ batch_max_steps: 9600 # Length of each audio in batch for training. Make sure dividable by hop_size.
71
+ batch_max_steps_valid: 48000 # Length of each audio for validation. Make sure dividable by hope_size.
72
+ remove_short_samples: true # Whether to remove samples the length of which are less than batch_max_steps.
73
+ allow_cache: true # Whether to allow cache in dataset. If true, it requires cpu memory.
74
+ is_shuffle: true # shuffle dataset after each epoch.
75
+
76
+ ###########################################################
77
+ # OPTIMIZER & SCHEDULER SETTING #
78
+ ###########################################################
79
+ generator_optimizer_params:
80
+ lr_fn: "PiecewiseConstantDecay"
81
+ lr_params:
82
+ boundaries: [100000, 200000, 300000, 400000, 500000, 600000, 700000]
83
+ values: [0.001, 0.0005, 0.00025, 0.000125, 0.0000625, 0.00003125, 0.000015625, 0.000001]
84
+ amsgrad: false
85
+
86
+ discriminator_optimizer_params:
87
+ lr_fn: "PiecewiseConstantDecay"
88
+ lr_params:
89
+ boundaries: [100000, 200000, 300000, 400000, 500000]
90
+ values: [0.00025, 0.000125, 0.0000625, 0.00003125, 0.000015625, 0.000001]
91
+ amsgrad: false
92
+
93
+ gradient_accumulation_steps: 1
94
+ ###########################################################
95
+ # INTERVAL SETTING #
96
+ ###########################################################
97
+ discriminator_train_start_steps: 200000 # steps begin training discriminator
98
+ train_max_steps: 4000000 # Number of training steps.
99
+ save_interval_steps: 20000 # Interval steps to save checkpoint.
100
+ eval_interval_steps: 5000 # Interval steps to evaluate the network.
101
+ log_interval_steps: 200 # Interval steps to record the training log.
102
+
103
+ ###########################################################
104
+ # OTHER SETTING #
105
+ ###########################################################
106
+ num_save_intermediate_results: 1 # Number of batch to be saved as intermediate results.
model.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:52785c0e4d278b5ce12baa3f74bc4b412c29a44512e4615a59f9fbad0cdd985b
3
+ size 7543376