config/InfLoRA_opt_clip.yaml

dataset: &dataset cifar100
init_cls_num: &init_cls_num 20
inc_cls_num: &inc_cls_num 20
total_cls_num: &total_cls_num 100
task_num: &task_num 5
image_size: &image_size 224

dataset: *dataset
init_cls_num: *init_cls_num
inc_cls_num: *inc_cls_num
total_cls_num: *total_cls_num
task_num: *task_num

epoch: 20 # 20
val_per_epoch: 20

batch_size: 128 # 128

setting: task-agnostic 

testing_times: 5

train_trfms:
  - RandomResizedCrop:
      size: *image_size
  - RandomHorizontalFlip: {}
  - ToTensor: {}
  - Normalize:
      mean: [0., 0., 0.]
      std: [1., 1., 1.]

test_trfms: 
  - Resize:
      size: *image_size
  - ToTensor: {}
  - Normalize:
      mean: [0., 0., 0.]
      std: [1., 1., 1.]

optimizer:
  name: Adam
  kwargs:
    lr: 0.0005
    weight_decay: 0
    betas: [0.9, 0.999]

lr_scheduler:
  name: CosineSchedule
  kwargs:
    K: 20


backbone:
  name: clip
  kwargs:
    pretrained : True
    model_name : ViT-B/16
    experts_num: 0
    act_layer: QuickGELU
    norm_layer: LayerNorm
    attn_layer: MultiHeadAttention_LoRA

classifier:
  name: InfLoRA_OPT
  kwargs:
    use_ca: False
    dataset: *dataset
    init_cls_num: *init_cls_num
    inc_cls_num: *inc_cls_num
    task_num: *task_num
    lame: 1.0
    lamb: 0.95
    embd_dim: 768
    prompt_template : "a bad photo of a {}." # For CLIP
    visual_only: True # For CLIP, apply lora to only visual encoder or visual and text encoder