root
commited on
Commit
•
6a974a7
1
Parent(s):
69d5f59
fix model avg
Browse files
conf/conformer_rnnt_large.yaml
ADDED
@@ -0,0 +1,100 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# network architecture
|
2 |
+
# encoder related
|
3 |
+
encoder: conformer
|
4 |
+
encoder_conf:
|
5 |
+
output_size: 512 # dimension of attention
|
6 |
+
attention_heads: 8
|
7 |
+
linear_units: 2048 # the number of units of position-wise feed forward
|
8 |
+
num_blocks: 12 # the number of encoder blocks
|
9 |
+
dropout_rate: 0.1
|
10 |
+
positional_dropout_rate: 0.1
|
11 |
+
attention_dropout_rate: 0.1
|
12 |
+
input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
|
13 |
+
normalize_before: true
|
14 |
+
cnn_module_kernel: 31
|
15 |
+
use_cnn_module: true
|
16 |
+
activation_type: 'swish'
|
17 |
+
pos_enc_layer_type: 'rel_pos'
|
18 |
+
selfattention_layer_type: 'rel_selfattn'
|
19 |
+
|
20 |
+
|
21 |
+
joint_conf:
|
22 |
+
join_dim: 512
|
23 |
+
prejoin_linear: True
|
24 |
+
postjoin_linear: false
|
25 |
+
joint_mode: 'add'
|
26 |
+
activation: 'tanh'
|
27 |
+
|
28 |
+
predictor: rnn
|
29 |
+
predictor_conf:
|
30 |
+
embed_size: 512
|
31 |
+
output_size: 512
|
32 |
+
embed_dropout: 0.1
|
33 |
+
hidden_size: 512
|
34 |
+
num_layers: 2
|
35 |
+
bias: true
|
36 |
+
rnn_type: 'lstm'
|
37 |
+
dropout: 0.1
|
38 |
+
|
39 |
+
decoder: bitransformer
|
40 |
+
decoder_conf:
|
41 |
+
attention_heads: 8
|
42 |
+
dropout_rate: 0.1
|
43 |
+
linear_units: 2048
|
44 |
+
num_blocks: 3
|
45 |
+
positional_dropout_rate: 0.1
|
46 |
+
r_num_blocks: 3
|
47 |
+
self_attention_dropout_rate: 0.1
|
48 |
+
src_attention_dropout_rate: 0.1
|
49 |
+
|
50 |
+
# hybrid transducer+ctc+attention
|
51 |
+
model_conf:
|
52 |
+
transducer_weight: 0.75
|
53 |
+
ctc_weight: 0.1
|
54 |
+
attention_weight: 0.15
|
55 |
+
lsm_weight: 0.1 # label smoothing option
|
56 |
+
length_normalized_loss: false
|
57 |
+
reverse_weight: 0.3
|
58 |
+
|
59 |
+
dataset_conf:
|
60 |
+
filter_conf:
|
61 |
+
max_length: 1650
|
62 |
+
min_length: 10
|
63 |
+
token_max_length: 200
|
64 |
+
token_min_length: 1
|
65 |
+
resample_conf:
|
66 |
+
resample_rate: 16000
|
67 |
+
speed_perturb: true
|
68 |
+
fbank_conf:
|
69 |
+
num_mel_bins: 80
|
70 |
+
frame_shift: 10
|
71 |
+
frame_length: 25
|
72 |
+
dither: 0.1
|
73 |
+
spec_aug: true
|
74 |
+
spec_aug_conf:
|
75 |
+
num_t_mask: 2
|
76 |
+
num_f_mask: 2
|
77 |
+
max_t: 50
|
78 |
+
max_f: 10
|
79 |
+
shuffle: true
|
80 |
+
shuffle_conf:
|
81 |
+
shuffle_size: 1500
|
82 |
+
sort: true
|
83 |
+
sort_conf:
|
84 |
+
sort_size: 500 # sort_size should be less than shuffle_size
|
85 |
+
batch_conf:
|
86 |
+
batch_type: 'dynamic' # static or dynamic
|
87 |
+
max_frames_in_batch: 4000
|
88 |
+
|
89 |
+
grad_clip: 4
|
90 |
+
accum_grad: 1
|
91 |
+
max_epoch: 140
|
92 |
+
log_interval: 100
|
93 |
+
|
94 |
+
optim: adam
|
95 |
+
optim_conf:
|
96 |
+
lr: 0.001
|
97 |
+
scheduler: warmuplr # pytorch v1.1.0+ required
|
98 |
+
scheduler_conf:
|
99 |
+
warmup_steps: 25000
|
100 |
+
|
exp/conformer_transducer/avg_5.pt
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 277103087
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3224c02c52e558395dd02f4489ac23a243094c94d446119558c6f86540fb347b
|
3 |
size 277103087
|
exp/conformer_transducer/{avg_10.pt → avg_6.pt}
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 277103087
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3d3fa1073140f9adfe46cbfc7a391ff7c4894a414f2a7cb4d28f0cc60ed1f98b
|
3 |
size 277103087
|