ViT_SITS_ONLY.yaml

defaults:
  - _self_
  - global_config

MODEL:
  architecture:  "ViTFacto"
  vit_type: "dinov2_small"
  pretrained: True
  input_img_res: 264
  img_res: 252
  train_max_seq_len: 5
  val_max_seq_len: 5
  input_dim: 14
  patch_size: 14
  num_classes: 2
  out_H: 25
  out_W: 25
  doy_int_type: "channel"
  temp_enc_type: "convlstm"
  kernel_size: [3, 3]
  n_stack_layers: 1
  threshold: 0.5
  flag_lora: True
  rank_lora: 32
  alpha_lora: 32.0
  dropout_lora: 0.1

SOLVER:
  num_epochs: 20
  num_warmup_epochs: 2
  loss_function: masked_dice_loss
  lr_scheduler: 'cosine'
  lr_base: 5e-6
  lr_min: 1e-7
  lr_start: 1e-7
  num_cycles: 1
  weight_decay: 0.01
  accumulate_grad_batches: 3
  interval: "epoch"

### Leveraging RAW *.npy files ###
DATASETS:
  kwargs:
    mean_file: "${paths.bands_mean}"
    std_file: "${paths.bands_std}"
    with_loc: False
    with_doy: True
    # bands: possibility to specify the selected bands.
  train:
    paths: "${paths.split}"
    base_dir: ["${paths.pos_sits}", ${paths.neg_sits}"]
    label_dir: "${paths.label}"
    batch_size: 12
    num_workers: 8

  eval:
    paths: "${paths.split}"
    base_dir: ["${paths.pos_sits}", ${paths.neg_sits}"]
    label_dir: "${paths.label}"
    batch_size: 12
    num_workers: 8

  test:

### Leveraging Hf parquet files ###
#DATASETS:
#  mode: "huggingface"
#  kwargs:
#    mean_file: "${paths.bands_mean}"
#    std_file: "${paths.bands_std}"
#    with_loc: False
#    with_doy: True
    # bands: possibility to specify the selected bands.
#  train:
#    data_dir: "${paths.hf_data}"
#    batch_size: 24
#    num_workers: 8
#
#  eval:
#    data_dir: "${paths.hf_data}"
#    batch_size: 24
#    num_workers: 8

CHECKPOINT:
  load_from_checkpoint:
  experiment_name: "ViT_SITS_ONLY"
  save_path: "./results/models"
  train_metrics_steps: 200
  save_steps: 10000
  wandb_project: "${wandb.project}"
  wandb_user: "${wandb.user}"

SET-UP:
  seed: 42
  local_device_ids: [0]