Book a Demo!
CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutPoliciesSign UpSign In
TensorSpeech
GitHub Repository: TensorSpeech/TensorFlowTTS
 Path: blob/master/examples/fastspeech/conf/fastspeech.v1.yaml
1559 views
1
# This is the hyperparameter configuration file for FastSpeech v1.

2
# Please make sure this is adjusted for the LJSpeech dataset. If you want to

3
# apply to the other dataset, you might need to carefully change some parameters.

4
# This configuration performs 200k iters but a best checkpoint is around 150k iters.

5


6
###########################################################

7
#                FEATURE EXTRACTION SETTING               #

8
###########################################################

9
hop_size: 256            # Hop size.

10
format: "npy"

11


12


13
###########################################################

14
#              NETWORK ARCHITECTURE SETTING               #

15
###########################################################

16
model_type: "fastspeech"

17


18
fastspeech_params:

19
    n_speakers: 1

20
    encoder_hidden_size: 384

21
    encoder_num_hidden_layers: 4

22
    encoder_num_attention_heads: 2

23
    encoder_attention_head_size: 192  # hidden_size // num_attention_heads

24
    encoder_intermediate_size: 1024

25
    encoder_intermediate_kernel_size: 3

26
    encoder_hidden_act: "mish"

27
    decoder_hidden_size: 384

28
    decoder_num_hidden_layers: 4

29
    decoder_num_attention_heads: 2

30
    decoder_attention_head_size: 192  # hidden_size // num_attention_heads

31
    decoder_intermediate_size: 1024

32
    decoder_intermediate_kernel_size: 3

33
    decoder_hidden_act: "mish"

34
    num_duration_conv_layers: 2

35
    duration_predictor_filters: 256

36
    duration_predictor_kernel_sizes: 3

37
    num_mels: 80

38
    hidden_dropout_prob: 0.1

39
    attention_probs_dropout_prob: 0.1

40
    duration_predictor_dropout_probs: 0.1

41
    max_position_embeddings: 2048

42
    initializer_range: 0.02

43
    output_attentions: False

44
    output_hidden_states: False

45


46
###########################################################

47
#                  DATA LOADER SETTING                    #

48
###########################################################

49
batch_size: 16              # Batch size for each GPU with asuming that gradient_accumulation_steps is 1

50
remove_short_samples: true  # Whether to remove samples the length of which are less than batch_max_steps.

51
allow_cache: true           # Whether to allow cache in dataset. If true, it requires cpu memory.

52
mel_length_threshold: 32    # remove all targets has mel_length <= 32 

53
is_shuffle: true            # shuffle dataset after each epoch.

54
###########################################################

55
#             OPTIMIZER & SCHEDULER SETTING               #

56
###########################################################

57
optimizer_params:

58
    initial_learning_rate: 0.001

59
    end_learning_rate: 0.00005

60
    decay_steps: 150000          # < train_max_steps is recommend.

61
    warmup_proportion: 0.02

62
    weight_decay: 0.001

63


64
gradient_accumulation_steps: 1

65
var_train_expr: null  # trainable variable expr (eg. 'embeddings|encoder|decoder' )

66
                      # must separate by |. if var_train_expr is null then we 

67
                      # training all variable 

68
###########################################################

69
#                    INTERVAL SETTING                     #

70
###########################################################

71
train_max_steps: 200000               # Number of training steps.

72
save_interval_steps: 5000             # Interval steps to save checkpoint.

73
eval_interval_steps: 500              # Interval steps to evaluate the network.

74
log_interval_steps: 200               # Interval steps to record the training log.

75


76
###########################################################

77
#                     OTHER SETTING                       #

78
###########################################################

79
num_save_intermediate_results: 1  # Number of batch to be saved as intermediate results.

80


81