Upload hyperparams.yaml
Browse files- hyperparams.yaml +157 -0
hyperparams.yaml
ADDED
@@ -0,0 +1,157 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# ############################################################################
|
2 |
+
# Model: E2E ST JA->EN with Conformer
|
3 |
+
# Encoder: Conformer Encoder
|
4 |
+
# Decoder: Conformer Decoder + (CTC/ATT joint)
|
5 |
+
# Tokens: BPE
|
6 |
+
# losses: CTC
|
7 |
+
# Training: Custom JA->EN youtube scrape, ~600h
|
8 |
+
# Authors: Eric Engelhart, 2022
|
9 |
+
# ############################################################################
|
10 |
+
|
11 |
+
# Tokenier initialization
|
12 |
+
tokenizer: !new:sentencepiece.SentencePieceProcessor
|
13 |
+
|
14 |
+
# Features
|
15 |
+
sample_rate: 16000
|
16 |
+
n_fft: 400
|
17 |
+
n_mels: 80
|
18 |
+
|
19 |
+
# normalization
|
20 |
+
normalizer: !new:speechbrain.processing.features.InputNormalization
|
21 |
+
norm_type: global
|
22 |
+
|
23 |
+
compute_features: !new:speechbrain.lobes.features.Fbank
|
24 |
+
sample_rate: !ref <sample_rate>
|
25 |
+
n_fft: !ref <n_fft>
|
26 |
+
n_mels: !ref <n_mels>
|
27 |
+
|
28 |
+
####################### Model parameters ###########################
|
29 |
+
# Transformer
|
30 |
+
d_model: 384
|
31 |
+
nhead: 6
|
32 |
+
num_encoder_layers: 12
|
33 |
+
num_decoder_layers: 6
|
34 |
+
d_ffn: 1536
|
35 |
+
transformer_dropout: 0.1
|
36 |
+
activation: !name:torch.nn.GELU
|
37 |
+
output_neurons: 5000
|
38 |
+
vocab_size: 5000
|
39 |
+
attention_type: "regularMHA" # "RelPosMHAXL" or "regularMHA"
|
40 |
+
kernel_size: 15
|
41 |
+
encoder_module: conformer
|
42 |
+
|
43 |
+
|
44 |
+
# Outputs
|
45 |
+
blank_index: 0
|
46 |
+
label_smoothing: 0.1
|
47 |
+
pad_index: 0
|
48 |
+
bos_index: 1
|
49 |
+
eos_index: 2
|
50 |
+
unk_index: 0
|
51 |
+
|
52 |
+
# Decoding parameters
|
53 |
+
min_decode_ratio: 0.0
|
54 |
+
max_decode_ratio: 1.0
|
55 |
+
valid_search_interval: 2
|
56 |
+
valid_beam_size: 1
|
57 |
+
test_beam_size: 25
|
58 |
+
|
59 |
+
############################## models ################################
|
60 |
+
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
|
61 |
+
input_shape: (8, 10, 80)
|
62 |
+
num_blocks: 2
|
63 |
+
num_layers_per_block: 1
|
64 |
+
out_channels: (256, 256)
|
65 |
+
kernel_sizes: (3, 3)
|
66 |
+
strides: (2, 2)
|
67 |
+
residuals: (False, False)
|
68 |
+
|
69 |
+
Transformer: !new:speechbrain.lobes.models.transformer.TransformerST.TransformerST # yamllint disable-line rule:line-length
|
70 |
+
input_size: 5120
|
71 |
+
tgt_vocab: !ref <output_neurons>
|
72 |
+
d_model: !ref <d_model>
|
73 |
+
nhead: !ref <nhead>
|
74 |
+
num_encoder_layers: !ref <num_encoder_layers>
|
75 |
+
num_decoder_layers: !ref <num_decoder_layers>
|
76 |
+
d_ffn: !ref <d_ffn>
|
77 |
+
dropout: !ref <transformer_dropout>
|
78 |
+
activation: !ref <activation>
|
79 |
+
ctc_weight: 0
|
80 |
+
asr_weight: 0
|
81 |
+
mt_weight: 0
|
82 |
+
asr_tgt_vocab: !ref <output_neurons>
|
83 |
+
mt_src_vocab: !ref <output_neurons>
|
84 |
+
attention_type: !ref <attention_type>
|
85 |
+
kernel_size: !ref <kernel_size>
|
86 |
+
encoder_module: !ref <encoder_module>
|
87 |
+
normalize_before: True
|
88 |
+
causal: False
|
89 |
+
max_length: 5000
|
90 |
+
|
91 |
+
# only when multi-task setting is used
|
92 |
+
ctc_lin: !new:speechbrain.nnet.linear.Linear
|
93 |
+
input_size: !ref <d_model>
|
94 |
+
n_neurons: !ref <output_neurons>
|
95 |
+
|
96 |
+
seq_lin: !new:speechbrain.nnet.linear.Linear
|
97 |
+
input_size: !ref <d_model>
|
98 |
+
n_neurons: !ref <output_neurons>
|
99 |
+
|
100 |
+
# when asr-weight > 0 and ctc-weight < 1
|
101 |
+
asr_seq_lin: !new:speechbrain.nnet.linear.Linear
|
102 |
+
input_size: !ref <d_model>
|
103 |
+
n_neurons: !ref <vocab_size>
|
104 |
+
|
105 |
+
|
106 |
+
st_model: !new:torch.nn.ModuleList
|
107 |
+
- [!ref <CNN>, !ref <Transformer>, !ref <seq_lin>]
|
108 |
+
|
109 |
+
|
110 |
+
Tencoder: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
|
111 |
+
transformer: !ref <Transformer>
|
112 |
+
|
113 |
+
encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
|
114 |
+
input_shape: [null, null, !ref <n_mels>]
|
115 |
+
compute_features: !ref <compute_features>
|
116 |
+
normalize: !ref <normalizer>
|
117 |
+
cnn: !ref <CNN>
|
118 |
+
transformer_encoder: !ref <Tencoder>
|
119 |
+
|
120 |
+
|
121 |
+
|
122 |
+
decoder: !new:speechbrain.decoders.S2STransformerBeamSearch
|
123 |
+
modules: [!ref <Transformer>, !ref <seq_lin>, null]
|
124 |
+
bos_index: !ref <bos_index>
|
125 |
+
eos_index: !ref <eos_index>
|
126 |
+
blank_index: !ref <blank_index>
|
127 |
+
min_decode_ratio: !ref <min_decode_ratio>
|
128 |
+
max_decode_ratio: !ref <max_decode_ratio>
|
129 |
+
beam_size: !ref <test_beam_size>
|
130 |
+
using_eos_threshold: True
|
131 |
+
length_normalization: True
|
132 |
+
ctc_weight: 0
|
133 |
+
lm_weight: 0
|
134 |
+
|
135 |
+
|
136 |
+
modules:
|
137 |
+
compute_features: !ref <compute_features>
|
138 |
+
normalizer: !ref <normalizer>
|
139 |
+
pre_transformer: !ref <CNN>
|
140 |
+
Transformer: !ref <Transformer>
|
141 |
+
asr_model: !ref <st_model>
|
142 |
+
encoder: !ref <encoder>
|
143 |
+
decoder: !ref <decoder>
|
144 |
+
|
145 |
+
|
146 |
+
log_softmax: !new:torch.nn.LogSoftmax
|
147 |
+
dim: -1
|
148 |
+
|
149 |
+
|
150 |
+
|
151 |
+
# The pretrainer allows a mapping between pretrained files and instances that
|
152 |
+
# are declared in the yaml.
|
153 |
+
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
|
154 |
+
loadables:
|
155 |
+
tokenizer: !ref <tokenizer>
|
156 |
+
st: !ref <st_model>
|
157 |
+
normalizer: !ref <normalizer>
|