config.yaml

defaults:
  - _self_  # Override values within this file with values in selected files.
use_cuda: True
gpu: -1
seed: null
device: cuda
workers: 4
log_freq: 10

# Parameters for wandb
wandb:
  mode: online
  project: diffusion_tta

# General hyper-parameters for defining the classifier, diffusion model, U-Net
# autoencoder
model:
  sd_version: 1-4 # Stable Diffusion version
  use_dit: False # use DiT as diffusion backbone
  override_total_steps: -1 # some parameters for DiT
  freeze_class_embeds: True # freeze text class embeddings
  freeze_vae: False # freeze VQVAE
  class_arch: resnet18 # classifier backbone
  use_flash: False # use flash attention in Stable Diffusion
  adapt_only_classifier: False # only adapt the classifier

# Hyper-parameters for TTA
tta:
  adapt_topk: -1 # only adapt topk classes
  loss: mse # distance metric for diffusion loss
  model: diff_tta.models.tta.TTAGradientDescent_Class
  online: False # online adaptation
  use_same_noise_among_timesteps: False # use the same noise for all timesteps
  random_timestep_per_iteration: True # randomly sample timesteps
  rand_timestep_equal_int: False # split evenly from a random timestep to total_steps
  gradient_descent:
    base_learning_rate: 1e-5 # training lr for detr
    weight_decay: 0 # training wd for detr
    train_steps: 100 # training iteration for detr
    accum_iter: 1
    optimizer: adam
    optimizer_momentum: 0.9

# Hyper-parameters for input
input:
  root_path: /projects/katefgroup/datasets/ # CHANGE by yourself
  dataset_name: ImageNetDataset
  image_height: 64
  image_width: 64
  batch_size: 40 # how many data to fit in one gpu
  disc_img_crop: 224
  disc_img_resize: 232
  sd_img_res: 512
  use_objectnet: False
  mean: [0.485, 0.456, 0.406]
  std: [0.229, 0.224, 0.225]
  subsample: 3 # number of samples per category
  winoground_use_auth_token: null

hydra:
  run:
    dir: ./outputs/${now:%Y-%m-%d}/${now:%H-%M-%S.%f}