base.yaml

seed: 0
wandb_silent: false
test: false
proj: DiffusionAugmentation # wandb project name

model: resnet50 # can change this to MLP if you want to LP CLIP
finetune: false
epochs: 100
eval_only: false
resume: false
checkpoint_name: false

data:
  base_root: ./data
  embedding_root: ./embeddings # precomputed clip embeddings
  batch: 128
  augmentation: false # set to name of traditional augmentation method (cutmix, augmix, etc.)
  filter: true # apply image filtering
  extra_dataset: false # this is the dataset type for the augmented data (usually its Img2ImgDataset or BasicDataset)
  extra_classes: false
  extra_root: false # path to extra dataset
  num_extra: extra # how much generated data to add. Can be 'extra' (how we do it in the paper), 'all', or an int 
  class_balance: true # whether to balance the classes in the extra dataset, this is if you set num_extra to an int


filter:
  save_dir: filtering_results # where to save the results
  filtered_path: false # path to npy file with filtered images idxs, this is if you want to use a different filtering method than the one we provide
  model: ViT-L/14 # CLIP model for filtering
  load: True # set this to true after you compute the embeddings once 
  per_img: False # set this to true if you are doing an img2img method
  checkpoint_name: 'ckpt-Cub2011-none-resnet50-0.pth' # checkpoint to use for confidence-based filtering

hps:
  lr: 0.01
  weight_decay: 0.0001
  lr_scheduler: cosine

summarize: # for generating prompts
  captions_path: false # path to captions file
  prefix: 'a photo of a bird'

# vicuna args
cpu_offloading: false
debug: false
device: 'cuda'
gptq_act_order: false
gptq_ckpt: None
gptq_groupsize: -1
gptq_wbits: 16
gpus: None
load_8bit: false
max_gpu_memory: '10GB'
max_new_tokens: 512
model_path: 'lmsys/vicuna-13b-v1.3'
num_gpus: 4
repetition_penalty: 1.0
revision: 'main'
temperature: 1.0
message: 'hello, how are you?'