Upload Model config and Train config
Browse files- model_config.json +209 -0
- train_config.json +67 -0
model_config.json
ADDED
@@ -0,0 +1,209 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"summary": {
|
3 |
+
"quantizer_dim": 3584,
|
4 |
+
"codebook_per_group_per_residual": 3584,
|
5 |
+
"group": 1,
|
6 |
+
"residual": 1,
|
7 |
+
"original_residual_dim": 1024,
|
8 |
+
"codebook_upsample": 3.5,
|
9 |
+
"codebook_dim": 3584
|
10 |
+
},
|
11 |
+
"base_model": "QWen2.5-7B-Pretrain",
|
12 |
+
"token_id_offset": 152064,
|
13 |
+
"spec_transform": {
|
14 |
+
"sampling_rate": 24000,
|
15 |
+
"segment_size": 72000,
|
16 |
+
"num_mels": 128,
|
17 |
+
"n_fft": 1024,
|
18 |
+
"hop_size": 256,
|
19 |
+
"win_size": 1024,
|
20 |
+
"fmin": 0,
|
21 |
+
"fmax": 12000,
|
22 |
+
"fmax_loss": null
|
23 |
+
},
|
24 |
+
"encoder": {
|
25 |
+
"input_channels": 128,
|
26 |
+
"depths": [
|
27 |
+
3,
|
28 |
+
3,
|
29 |
+
9,
|
30 |
+
3
|
31 |
+
],
|
32 |
+
"dims": [
|
33 |
+
256,
|
34 |
+
512,
|
35 |
+
768,
|
36 |
+
1024
|
37 |
+
],
|
38 |
+
"drop_path_rate": 0.2,
|
39 |
+
"kernel_size": 7
|
40 |
+
},
|
41 |
+
"decoder": {
|
42 |
+
"hop_length": 256,
|
43 |
+
"upsample_rates": [
|
44 |
+
8,
|
45 |
+
4,
|
46 |
+
2,
|
47 |
+
2,
|
48 |
+
2
|
49 |
+
],
|
50 |
+
"upsample_kernel_sizes": [
|
51 |
+
16,
|
52 |
+
12,
|
53 |
+
4,
|
54 |
+
4,
|
55 |
+
4
|
56 |
+
],
|
57 |
+
"resblock_kernel_sizes": [
|
58 |
+
3,
|
59 |
+
7,
|
60 |
+
11
|
61 |
+
],
|
62 |
+
"resblock_dilation_sizes": [
|
63 |
+
[
|
64 |
+
1,
|
65 |
+
3,
|
66 |
+
5
|
67 |
+
],
|
68 |
+
[
|
69 |
+
1,
|
70 |
+
3,
|
71 |
+
5
|
72 |
+
],
|
73 |
+
[
|
74 |
+
1,
|
75 |
+
3,
|
76 |
+
5
|
77 |
+
]
|
78 |
+
],
|
79 |
+
"num_mels": 1024,
|
80 |
+
"upsample_initial_channel": 1024,
|
81 |
+
"use_template": false,
|
82 |
+
"pre_conv_kernel_size": 13,
|
83 |
+
"post_conv_kernel_size": 13
|
84 |
+
},
|
85 |
+
"quantizer": {
|
86 |
+
"quantizer_type": "grvq",
|
87 |
+
"input_dim": 1024,
|
88 |
+
"n_groups": 1,
|
89 |
+
"n_codebooks": 1,
|
90 |
+
"codebook_size": 32768,
|
91 |
+
"codebook_dim": 3584,
|
92 |
+
"levels": [
|
93 |
+
8,
|
94 |
+
5,
|
95 |
+
5,
|
96 |
+
5
|
97 |
+
],
|
98 |
+
"downsample_factor": [
|
99 |
+
1
|
100 |
+
],
|
101 |
+
"ema_decay": 0.8,
|
102 |
+
"codebook_diversity_loss_weight": 1.0,
|
103 |
+
"codebook_diversity_temperature": 100.0
|
104 |
+
},
|
105 |
+
"teacher_quantizer": {
|
106 |
+
"quantizer_type": "grvq",
|
107 |
+
"input_dim": 1024,
|
108 |
+
"n_groups": 2,
|
109 |
+
"n_codebooks": 1,
|
110 |
+
"codebook_size": 32768,
|
111 |
+
"codebook_dim": 3584,
|
112 |
+
"levels": [
|
113 |
+
8,
|
114 |
+
5,
|
115 |
+
5,
|
116 |
+
5
|
117 |
+
],
|
118 |
+
"downsample_factor": [
|
119 |
+
2
|
120 |
+
],
|
121 |
+
"ema_decay": 0.8,
|
122 |
+
"codebook_diversity_loss_weight": 1.0,
|
123 |
+
"codebook_diversity_temperature": 100.0
|
124 |
+
},
|
125 |
+
"descriminators": {
|
126 |
+
"MultiPeriodDiscriminator": {
|
127 |
+
"periods": [
|
128 |
+
5,
|
129 |
+
8,
|
130 |
+
13,
|
131 |
+
19,
|
132 |
+
30
|
133 |
+
],
|
134 |
+
"kernal_size": 5,
|
135 |
+
"stride": 3
|
136 |
+
},
|
137 |
+
"MultiScaleDiscriminator": {
|
138 |
+
"avg_poolings": {
|
139 |
+
"kernal_sizes": [
|
140 |
+
6,
|
141 |
+
6
|
142 |
+
],
|
143 |
+
"stridess": [
|
144 |
+
3,
|
145 |
+
3
|
146 |
+
],
|
147 |
+
"paddings": [
|
148 |
+
3,
|
149 |
+
3
|
150 |
+
]
|
151 |
+
},
|
152 |
+
"DiscriminatorS": {
|
153 |
+
"kernal_sizes": [
|
154 |
+
21,
|
155 |
+
61,
|
156 |
+
61,
|
157 |
+
61,
|
158 |
+
61,
|
159 |
+
61,
|
160 |
+
7
|
161 |
+
],
|
162 |
+
"strides": [
|
163 |
+
1,
|
164 |
+
3,
|
165 |
+
3,
|
166 |
+
6,
|
167 |
+
6,
|
168 |
+
1,
|
169 |
+
1
|
170 |
+
],
|
171 |
+
"paddings": [
|
172 |
+
10,
|
173 |
+
30,
|
174 |
+
30,
|
175 |
+
30,
|
176 |
+
30,
|
177 |
+
30,
|
178 |
+
3
|
179 |
+
]
|
180 |
+
}
|
181 |
+
},
|
182 |
+
"MultiScaleSTFTDiscriminator": {
|
183 |
+
"n_ffts": [
|
184 |
+
1024,
|
185 |
+
2048,
|
186 |
+
512,
|
187 |
+
256,
|
188 |
+
128
|
189 |
+
],
|
190 |
+
"hop_lengths": [
|
191 |
+
256,
|
192 |
+
512,
|
193 |
+
128,
|
194 |
+
64,
|
195 |
+
32
|
196 |
+
],
|
197 |
+
"win_lengths": [
|
198 |
+
1024,
|
199 |
+
2048,
|
200 |
+
512,
|
201 |
+
256,
|
202 |
+
128
|
203 |
+
],
|
204 |
+
"filters": 32,
|
205 |
+
"in_channels": 1,
|
206 |
+
"out_channels": 1
|
207 |
+
}
|
208 |
+
}
|
209 |
+
}
|
train_config.json
ADDED
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"adam": {
|
3 |
+
"learning_rate": 0.0002,
|
4 |
+
"adam_b1": 0.5,
|
5 |
+
"adam_b2": 0.9,
|
6 |
+
"lr_decay": 0.98,
|
7 |
+
"weight_decay": 0.001
|
8 |
+
},
|
9 |
+
"wandb": {
|
10 |
+
"use_wandb": true,
|
11 |
+
"wandb_project": "audio-tokenizer-0927",
|
12 |
+
"wandb_group": "audio-tokenizer",
|
13 |
+
"wandb_team": "ccnl-s-team"
|
14 |
+
},
|
15 |
+
"mels_path": null,
|
16 |
+
"training_files_path": {
|
17 |
+
"replay_training_file_pathes": [
|
18 |
+
{
|
19 |
+
"path": "/cognitive_comp/common_data/audio/output/24k",
|
20 |
+
"replay_rate": 1.0
|
21 |
+
},
|
22 |
+
{
|
23 |
+
"path": "/cognitive_comp/common_data/audio/output/24k_data/24k_1_10s/stage2",
|
24 |
+
"replay_rate": 0.6
|
25 |
+
},
|
26 |
+
{
|
27 |
+
"path": "/cognitive_comp/common_data/audio/output/24k_data/24k_1_10s/stage3",
|
28 |
+
"replay_rate": 0.6
|
29 |
+
},
|
30 |
+
{
|
31 |
+
"path": "/cognitive_comp/common_data/audio/output/lam/asr_tts",
|
32 |
+
"replay_rate": 0.01
|
33 |
+
}
|
34 |
+
],
|
35 |
+
"current_training_file_pathes": [
|
36 |
+
"/cognitive_comp/common_data/audio/output/24k_data/24k_1_10s/stage4"
|
37 |
+
]
|
38 |
+
},
|
39 |
+
"validation_files_path": [
|
40 |
+
"/cognitive_comp/common_data/audio/output/dev",
|
41 |
+
"/cognitive_comp/common_data/audio/output/24k_data/dev/stage2",
|
42 |
+
"/cognitive_comp/common_data/audio/output/24k_data/dev/stage3",
|
43 |
+
"/cognitive_comp/common_data/audio/output/24k_data/dev/stage4"
|
44 |
+
],
|
45 |
+
"save_path": "/cognitive_comp/wangrui/data/lam0927",
|
46 |
+
"distill": {
|
47 |
+
"is_distill": true,
|
48 |
+
"use_fm_distill": false,
|
49 |
+
"quantizer_transfer": false,
|
50 |
+
"teacher_ckpt_path": "/cognitive_comp/wangrui/data/lam0923/saved_ckpt"
|
51 |
+
},
|
52 |
+
"training_epochs": 50,
|
53 |
+
"checkpoint_interval": 1000,
|
54 |
+
"validation_interval": 2000,
|
55 |
+
"accumulation_steps": 1,
|
56 |
+
"summary_interval": 10,
|
57 |
+
"stdout_interval": 10,
|
58 |
+
"num_ckpt_keep": 30,
|
59 |
+
"fine_tuning": false,
|
60 |
+
"num_gpus": 8,
|
61 |
+
"batch_size": 152,
|
62 |
+
"seed": 1234,
|
63 |
+
"sampling_rate": 24000,
|
64 |
+
"codebook_loss_lambda": 1.0,
|
65 |
+
"commitment_loss_lambda": 0.25,
|
66 |
+
"num_workers": 4
|
67 |
+
}
|