Hugging Face's logo

ByteVisionLab
/
DetailFlow-16

Model card Files
xet
Community
yoloyo commited on
Commit
23507fc
·
verified ·
1 Parent(s): 80ea5c5

Upload folder using huggingface_hub

Browse files
Files changed (3) hide show
  1. checkpoints/128.pt +3 -0
  2. config.json +60 -0
  3. config.yaml +108 -0
checkpoints/128.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b75b7fa2e05ecd887e0ecace2d6276dc9f03d84ac0b0e7e47e170065cf7c2f72
3
+ size 2771935014
config.json ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "image_size": 256,
3
+ "max_image_size": 256,
4
+ "codebook_size": 8192,
5
+ "codebook_embed_dim": 8,
6
+ "codebook_l2_norm": true,
7
+ "codebook_show_usage": true,
8
+ "commit_loss_beta": 0.25,
9
+ "entropy_loss_ratio": 0.0,
10
+ "vq_loss_ratio": 1.0,
11
+ "kl_loss_weight": 1e-06,
12
+ "tau": 0.07,
13
+ "num_codebooks": 1,
14
+ "encoder_ch_mult": [
15
+ 1,
16
+ 1,
17
+ 2,
18
+ 2,
19
+ 4
20
+ ],
21
+ "decoder_ch_mult": [
22
+ 1,
23
+ 1,
24
+ 2,
25
+ 2,
26
+ 4
27
+ ],
28
+ "z_channels": 256,
29
+ "dropout_p": 0.0,
30
+ "enc_type": "siglip2",
31
+ "dec_type": "siglip2",
32
+ "encoder_model": "siglip2_base",
33
+ "decoder_model": "siglip2_base",
34
+ "num_latent_tokens": 128,
35
+ "enc_tuning_method": "full",
36
+ "dec_tuning_method": "full",
37
+ "enc_pretrained": true,
38
+ "dec_pretrained": false,
39
+ "enc_patch_size": 16,
40
+ "dec_patch_size": 16,
41
+ "enc_drop_path_rate": 0.0,
42
+ "dec_drop_path_rate": 0.0,
43
+ "repa": false,
44
+ "repa_patch_size": 16,
45
+ "repa_model": "siglip2",
46
+ "repa_proj_dim": 1024,
47
+ "repa_layer_indices": 1,
48
+ "repa_loss_weight": 0.5,
49
+ "repa_align": "global",
50
+ "vq_mean": 0.0,
51
+ "vq_std": 1.0,
52
+ "causal_encoder": true,
53
+ "causal_decoder": false,
54
+ "gradient_checkpointing_encoder": false,
55
+ "gradient_checkpointing_decoder": false,
56
+ "group_size": 8,
57
+ "causal_num": null,
58
+ "global_token_loss_weight": 1.0,
59
+ "correction_training": true
60
+ }
config.yaml ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ config: configs/vq_8k_siglip_b_res_p02_pw15_enc.yaml
2
+ exp_index:
3
+ data_path: /mnt/bn/cloud-project-lq/code/liuyh/data/vq_data/train
4
+ cloud_save_path: experiments/tokenizer
5
+ no_local_save: true
6
+ vq_model: VQ-16
7
+ vq_ckpt:
8
+ finetune: false
9
+ finetune_decoder: false
10
+ model_weight_strict: true
11
+ ema: false
12
+ codebook_size: 8192
13
+ codebook_embed_dim: 8
14
+ codebook_l2_norm: true
15
+ codebook_weight: 1.0
16
+ entropy_loss_ratio: 0.0
17
+ vq_loss_ratio: 1.0
18
+ commit_loss_beta: 0.25
19
+ reconstruction_weight: 1.0
20
+ reconstruction_loss: l2
21
+ kl_loss_weight: 1e-06
22
+ tau: 0.07
23
+ num_codebooks: 1
24
+ perceptual_weight: 1.0
25
+ perceptual_loss: vgg
26
+ perceptual_model: vgg
27
+ perceptual_dino_variants: depth12_no_train
28
+ perceptual_intermediate_loss: false
29
+ perceptual_logit_loss: false
30
+ perceptual_resize: false
31
+ perceptual_warmup: 10000
32
+ disc_weight: 0.2
33
+ disc_start: 40000
34
+ disc_dim: 64
35
+ disc_type: dino
36
+ disc_loss: hinge
37
+ gen_loss: hinge
38
+ lecam_loss_weight: 0.001
39
+ use_diff_aug: true
40
+ disc_cr_loss_weight: 4.0
41
+ disc_adaptive_weight: false
42
+ compile: false
43
+ dropout_p: 0.0
44
+ results_dir: ./logs/task/detailflow_demo_task_128token
45
+ dataset: imagenet
46
+ image_size: 256
47
+ epochs: 250
48
+ optimizer: adam
49
+ lr: 1.0e-4
50
+ lr_warmup_epochs: 1
51
+ lr_scheduler: cosine
52
+ weight_decay: 0.0001
53
+ beta1: 0.9
54
+ beta2: 0.95
55
+ max_grad_norm: 1.0
56
+ global_batch_size: 256
57
+ global_seed: 42
58
+ num_workers: 16
59
+ log_every: 50
60
+ vis_every: 5000
61
+ ckpt_every: 5000
62
+ save_epochs: 1
63
+ gradient_accumulation_steps: 1
64
+ mixed_precision: bf16
65
+ enc_type: siglip2
66
+ dec_type: siglip2
67
+ num_latent_tokens: 128
68
+ encoder_model: siglip2_base
69
+ decoder_model: siglip2_base
70
+ encoder_tuning_method: full
71
+ decoder_tuning_method: full
72
+ encoder_pretrained: true
73
+ decoder_pretrained: false
74
+ encoder_patch_size: 16
75
+ decoder_patch_size: 16
76
+ repa: false
77
+ repa_model: siglip2
78
+ repa_patch_size: 16
79
+ repa_proj_dim: 1024
80
+ repa_loss_weight: 0.5
81
+ repa_align: global
82
+ repa_layer_indices: 1
83
+ resume_from_newest_ckpt: true
84
+ gradient_checkpointing_encoder: false
85
+ gradient_checkpointing_decoder: false
86
+ debug_mode: false
87
+ content_degradation: resolution_power
88
+ degradation_prob: 0.2
89
+ degradation_loss_res: 224
90
+ degradation_power: 1.5
91
+ causal_encoder: true
92
+ causal_decoder: false
93
+ max_image_size: 256
94
+ min_image_size: 256
95
+ dynamic_max_image_size:
96
+ dynamic_resolution_prob: 0
97
+ max_resolution_prob: 0
98
+ adjust_bs_by_resolution: false
99
+ group_size: 8
100
+ global_token_loss_weight: 1.0
101
+ correction_training: true
102
+ causal_num:
103
+ rank: 0
104
+ world_size: 32
105
+ gpu: 0
106
+ dist_url: env://
107
+ distributed: true
108
+ dist_backend: nccl