config: configs/softvq-l-64.yaml
exp_index:
data_path: /root/data/pub_nas/zhangbq/datasets/imagenet
cloud_save_path: experiments/tokenizer
no_local_save: true
vq_model: SoftVQ
vq_ckpt:
finetune: false
ema: true
codebook_size: 8192
codebook_embed_dim: 32
codebook_l2_norm: true
codebook_weight: 1.0
entropy_loss_ratio: 0.01
vq_loss_ratio: 0.0
commit_loss_beta: 0.0
reconstruction_weight: 1.0
reconstruction_loss: l2
kl_loss_weight: 1e-06
tau: 0.07
num_codebooks: 4
perceptual_weight: 1.0
perceptual_loss: vgg
perceptual_model: vgg
perceptual_dino_variants: depth12_no_train
perceptual_intermediate_loss: false
perceptual_logit_loss: false
perceptual_resize: false
perceptual_warmup: 10000
disc_weight: 0.2
disc_start: 20000
disc_dim: 64
disc_type: dino
disc_loss: hinge
gen_loss: hinge
lecam_loss_weight: 0.001
use_diff_aug: true
disc_cr_loss_weight: 4.0
disc_adaptive_weight: false
compile: false
dropout_p: 0.0
results_dir: experiments/tokenizer
dataset: imagenet
image_size: 256
epochs: 50
optimizer: adam
lr: 1.0e-4
lr_warmup_epochs: 1
lr_scheduler: cosine
weight_decay: 0.0001
beta1: 0.9
beta2: 0.95
max_grad_norm: 1.0
use_ar_mask: true
global_batch_size: 256
global_seed: 42
num_workers: 16
log_every: 50
vis_every: 5000
ckpt_every: 10000
gradient_accumulation_steps: 1
mixed_precision: bf16
enc_type: vit
dec_type: vit
num_latent_tokens: 64
encoder_model: vit_large_patch14_dinov2.lvd142m
decoder_model: vit_large_patch14_dinov2.lvd142m
encoder_tuning_method: full
decoder_tuning_method: full
encoder_pretrained: true
decoder_pretrained: false
encoder_patch_size: 16
decoder_patch_size: 16
repa: true
repa_model: vit_large_patch14_dinov2.lvd142m
repa_patch_size: 16
repa_proj_dim: 1024
repa_loss_weight: 0.1
repa_align: repeat
optim: adamw
rank: 0
world_size: 8
gpu: 0
dist_url: env://
distributed: true
dist_backend: nccl