yhaha commited on
Commit
2abec62
·
verified ·
1 Parent(s): d4cbf30

Upload 01_pretrain

Browse files
01_pretrain/config.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ wandb_config:
2
+ project: DiTAR
3
+ wandb_run_name: ditar_train_dit_emilia_0.6b_semanticvae_again
4
+ resume_id: null
5
+ mode: online
6
+ notes: null
7
+ ckpts:
8
+ logger: wandb
9
+ log_samples: true
10
+ save_per_updates: 2500
11
+ keep_last_n_checkpoints: -1
12
+ last_per_updates: 2500
13
+ save_dir: ckpts/ditar_train_dit_emilia_0.6b_semanticvae_again
14
+ checkpoint_step: null
15
+ datasets:
16
+ train_ds_path: data/Emilia_old/Emilia_ZH_EN_char
17
+ val_ds_path: data/Emilia_old/LibriTTS_test-clean_char_1280
18
+ batch_size_per_gpu: 400
19
+ batch_size_type: frame
20
+ max_samples: 16
21
+ num_workers: 32
22
+ target_sample_rate: 16000
23
+ optim:
24
+ epochs: 10
25
+ learning_rate: 0.0001
26
+ num_warmup_updates: 10000
27
+ grad_accumulation_steps: 1
28
+ max_grad_norm: 1.0
29
+ bnb_optimizer: false
30
+ model:
31
+ name: DiTAR
32
+ audio_type: vae_online
33
+ audio_channels: 64
34
+ patch_size: ${model.aggregation_encoder.patch_size}
35
+ dim: ${model.backbone.hidden_size}
36
+ mlp_hidden_dim: 512
37
+ backbone:
38
+ name: Qwen
39
+ hidden_size: 1024
40
+ version: qwen3
41
+ qwen_config_path: checkpoints/Qwen3-0.6B/config.json
42
+ pretrained_LM_path: checkpoints/Qwen3-0.6B
43
+ load_pretrained_weights: true
44
+ aggregation_encoder:
45
+ name: qwen
46
+ hidden_size: ${model.backbone.hidden_size}
47
+ intermediate_size: 4096
48
+ num_attention_heads: 16
49
+ num_hidden_layers: 6
50
+ patch_size: 4
51
+ pool_type: cls
52
+ loc_decoder:
53
+ name: DiT
54
+ model:
55
+ dim: 1024
56
+ out_dim: ${model.audio_channels}
57
+ depth: 6
58
+ heads: 16
59
+ ff_mult: 4
60
+ qk_norm: null
61
+ pe_attn_head: null
62
+ checkpoint_activations: false
63
+ history_vae_window_size: 4
64
+ drop_cond_prob: 0.1
65
+ vocoder:
66
+ mel_spec_type: semanticvae
67
+ is_local: true
68
+ local_path: checkpoints/Semantic-VAE/semantic_vae_1000k
69
+ feature_extractor_ckpt: checkpoints/Semantic-VAE/semantic_vae_1000k
70
+ loss_weight:
71
+ diff_loss: 1.0
72
+ stop_loss: 1.0
01_pretrain/model_200000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0f2c89884109a5dd348fdd713c817c3b5848b29f00265f029e71d51ef2a367a6
3
+ size 13391776877