data:
  type: merra2

  # Input variables definition
  input_surface_vars:
    - EFLUX
    - GWETROOT
    - HFLUX
    - LAI
    - LWGAB # surface absorbed longwave radiation
    - LWGEM # longwave flux emitted from surface
    - LWTUP # upwelling longwave flux at toa
    - PS # surface pressure
    - QV2M # 2-meter specific humidity
    - SLP # sea level pressure
    - SWGNT # surface net downward shortwave flux
    - SWTNT # toa net downward shortwave flux
    - T2M # near surface temperature
    - TQI # total precipitable ice water
    - TQL # total precipitable liquid water
    - TQV # total precipitable water vapor
    - TS # surface skin temperature
    - U10M # 10m eastward wind
    - V10M # 10m northward wind
    - Z0M # surface roughness
  input_static_surface_vars: [FRACI, FRLAND, FROCEAN, PHIS]
  input_vertical_vars:
    - CLOUD # cloud feraction for radiation
    - H # geopotential/ mid layer heights
    - OMEGA # vertical pressure velocity
    - PL # mid level pressure
    - QI # mass fraction of clous ice water
    - QL # mass fraction of cloud liquid water
    - QV # specific humidity
    - T # tempertaure
    - U # eastward wind
    - V # northward wind
  # (model level/ml ~ pressure level/hPa)
  # 52ml ~ 562.5hPa, 56ml ~ 700hPa, 63 ml ~ 850hPa
  input_levels: [34.0, 39.0, 41.0, 43.0, 44.0, 45.0, 48.0, 53.0, 56.0, 63.0, 68.0, 72.0]
  ## remove: n_input_timestamps: 1
  # Output variables definition
  output_vars:
    - T2M # near surface temperature

  n_input_timestamps: 2

  # Data transformations
  # Initial crop before any other processing
  crop_lat: [0, 1]
  # crop_lon: [0, 0]
  # coarsening of target -- applied after crop
  input_size_lat: 60 # 6x coarsening 
  input_size_lon: 96 # 6x coarsening 
  apply_smoothen: True

model:

  # Platform independent config
  num_static_channels: 7
  embed_dim: 2560
  token_size:
    - 1
    - 1
  n_blocks_encoder: 12
  mlp_multiplier: 4
  n_heads: 16
  dropout_rate: 0.0
  drop_path: 0.05
  
  # Accepted values: temporal, climate, none
  residual: climate
  
  residual_connection: True
  encoder_shift: False

  downscaling_patch_size: [2, 2]
  downscaling_embed_dim: 256
  encoder_decoder_type: 'conv' # ['conv', 'transformer']
  encoder_decoder_upsampling_mode: pixel_shuffle # ['nearest', 'bilinear', 'pixel_shuffle', 'conv_transpose']
  encoder_decoder_kernel_size_per_stage: [[3], [3]] # Optional, default = 3 for conv_tanspose [[3], [2]]
  encoder_decoder_scale_per_stage: [[2], [3]] # First list determines before/after backbone
  encoder_decoder_conv_channels: 128



job_id: inference-test
batch_size: 1
num_epochs: 400
dl_num_workers: 2
dl_prefetch_size: 1
learning_rate: 0.0001
limit_steps_train: 250
limit_steps_valid: 25
min_lr: 0.00001
max_lr: 0.0002
warm_up_steps: 0
mask_unit_size:
  - 15
  - 16
mask_ratio_inputs: 0.0
mask_ratio_targets: 0.0
max_batch_size: 16

path_experiment: experiment

backbone_freeze: True
backbone_prefix: encoder.
finetune_w_static: True
strict_matching: true