MambaUnet_servir.yaml

seed_everything: 42

# ---------------------------- TRAINER -------------------------------------------
trainer:
  default_root_dir: "checkpoint/MambaUnet"
  precision: "16-mixed"
  min_epochs: 1
  max_epochs: 100
  accelerator: cuda
  # limit_train_batches: 10
  devices: [2]
  # strategy: ddp
  num_nodes: 1
  enable_progress_bar: true
  sync_batchnorm: True
  enable_checkpointing: True
  # debugging
  fast_dev_run: false
  logger:
    - class_path: pytorch_lightning.loggers.WandbLogger
      init_args:
        project: "MambaUnet_Seq_UCTransnet"
        name: "MambaUnet_Seq_UCTransnet_Wandb_test"
        save_dir: "checkpoint/MambaUnet/wandb_logs"
        log_model: False
    - class_path: pytorch_lightning.loggers.CSVLogger
      init_args:
        save_dir: "checkpoint/MambaUnet/logs"
        name: null
        version: null

  callbacks:
    - class_path: pytorch_lightning.callbacks.LearningRateMonitor
      init_args:
        logging_interval: "step"

    - class_path: pytorch_lightning.callbacks.ModelCheckpoint
      init_args:
        dirpath: "checkpoint/MambaUnet/checkpoints"
        monitor: "val/mse" # name of the logged metric which determines when model is improving
        mode: "min" # "max" means higher metric value is better, can be also "min"
        save_top_k: 1 # save k best models (determined by above metric)
        save_last: True # additionally always save model from last epoch
        verbose: False
        filename: "epoch_{epoch:03d}"
        auto_insert_metric_name: False

    - class_path: pytorch_lightning.callbacks.EarlyStopping
      init_args:
        monitor: "val/mse" # name of the logged metric which determines when model is improving
        mode: "min" # "max" means higher metric value is better, can be also "min"
        patience: 10 # how many validation epochs of not improving until training stops
        min_delta: 0. # minimum change in the monitored metric needed to qualify as an improvement

    - class_path: pytorch_lightning.callbacks.RichModelSummary
      init_args:
        max_depth: -1

    - class_path: pytorch_lightning.callbacks.RichProgressBar

# ---------------------------- MODEL -------------------------------------------
model:
  pretrained_path: ""
  beta_1: 0.9
  beta_2: 0.99
  lr: 5e-4
  weight_decay: 1e-5
  warmup_epochs: 10
  max_epochs: 50
  warmup_start_lr: 1e-8
  eta_min: 1e-8
  net:
    radar_timestamp: 4        # Initial Input for VSSM
    radar_data_types: 2       # Initial Input for VSSM
    sat_data_types: 8         # Fusion concat input for bottleneck layer
    # endregion
    
    # region Preprocess Embedding
    patch_size: 4
    preprocess_depth: 2
    # endregion
    
    # region Encoder & Decoder Path
    depths: [3,4]           # Encoder and Decoder Depth
    dims: [32,64]             # Channel after each VS Block
    size: 400                   # Initial Size of H,W of Input(Radar - 400)
    # endregion
    
    # region Bottleneck
    bottleneck_depth: 8      # The length of a VS Block
    bottleneck_dim: 128        # The dimension after the encoding phase, will be doubled to fuse radar and sat
    bottleneck_size: 25       # The H,W size of the satellite image
    # endregion
    
    # region Output
    sat_output: 1
    radar_output: 1
    # endregion
    
    # region Config
    ape: True 
    final_skip: False
    patch_norm: True           
    drop_rate: 0.            # must be in [0:1]
    drop_path_rate: 0.1       # must be in [0:1]
    use_checkpoint: False
    pretrained_path: None
    # endregion

    #skip connection
    upgrade_skip_connection: "UCTransNet"
    img_size: 400
    vis: False
    channel_num: [8,128,256]
    patchSize: [40,10,5]
    KV_size: 392
    num_layers: 3
    num_heads: 3
    attention_dropout_rate: 0.1
    embeddings_dropout_rate: 0.1
    dropout_rate: 0.1
    expand_ratio: 3
    cca: True



# ---------------------------- DATA -------------------------------------------
data:
  dir_data: "/data/weather2025/NhaBe"
  time_points_radar: 1
  time_points_sat: 1
  sat_inp_vars: [
      "land_sea_mask",
      "orography",
      "lattitude",
      "2m_temperature",
      "10m_u_component_of_wind",
      "10m_v_component_of_wind",
      "total_precipitation",
      "2m_dewpoint_temperature"
    ]
  sat_out_vars: "total_precipitation"
  rad_inp_vars: [
      "precipitation"
    ]
  rad_out_vars:      "precipitation"
  hours_predicted: 3
  batch_size: 16
  num_workers: 4
  pin_memory: False
  short_timestep: False
  rebuild_val: True
  augmentation: True
  servir_format: True
  # global_prior: True