config.yaml

# SPDX-FileCopyrightText: Copyright (c) 2023 - 2025 NVIDIA CORPORATION & AFFILIATES.
# SPDX-FileCopyrightText: All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# Configuration file for FloodForecaster training.
# This file is used by Hydra to configure the training run.

hydra:
  job:
    chdir: True  # Change directory to the job's working directory.
  run:
    dir: ./outputs/  # Directory to save outputs.

# Distributed computing
distributed:
  use_distributed: false
  wireup_info: 'mpi'
  wireup_store: 'tcp'
  model_parallel_size: 1
  seed: 123
  device: 'cuda:0'

# Dataset related for training and one-step evaluation
source_data:
  root: "${DATA_ROOT:/path/to/source/data}"  # Set DATA_ROOT environment variable or update this path
  resolution: 48
  n_history: 3
  batch_size: 8
  query_res: [48, 48]
  xy_file: "M40_XY.txt"
  static_files:
    # Note: M40_XY.txt is included here as a static feature (XY coordinates as features)
    # The xy_file parameter loads it separately for geometry, while static_files includes it as a feature
    # This is intentional for flood modeling where XY coordinates are used both for geometry and as features
    - "M40_XY.txt"
    - "M40_CA.txt"
    - "M40_CE.txt"
    - "M40_CS.txt"
    - "M40_FA.txt"
    - "M40_A.txt"
    - "M40_CU.txt"
  dynamic_patterns:
    WD: "M40_WD_{}.txt"
    VX: "M40_VX_{}.txt"
    VY: "M40_VY_{}.txt"
  boundary_patterns:
    inflow: "M40_US_InF_{}.txt"
  noise_type: "none"
  noise_std: [0.01, 0.001, 0.001]
  rollout_length: 78
  skip_before_timestep: 12
  dt: 1200

# Target domain dataset
target_data:
  root: "${TARGET_DATA_ROOT:/path/to/target/data}"  # Set TARGET_DATA_ROOT environment variable or update this path
  resolution: 48
  n_history: 3
  batch_size: 8
  query_res: [48, 48]
  xy_file: "M40_XY.txt"
  static_files:
    # Note: M40_XY.txt is included here as a static feature (XY coordinates as features)
    # The xy_file parameter loads it separately for geometry, while static_files includes it as a feature
    # This is intentional for flood modeling where XY coordinates are used both for geometry and as features
    - "M40_XY.txt"
    - "M40_CA.txt"
    - "M40_CE.txt"
    - "M40_CS.txt"
    - "M40_FA.txt"
    - "M40_A.txt"
    - "M40_CU.txt"
  dynamic_patterns:
    WD: "M40_WD_{}.txt"
    VX: "M40_VX_{}.txt"
    VY: "M40_VY_{}.txt"
  boundary_patterns:
    inflow: "M40_US_InF_{}.txt"
  noise_type: "none"
  noise_std: [0.01, 0.001, 0.001]
  rollout_length: 78
  skip_before_timestep: 12
  dt: 1200

# Rollout evaluation dataset
rollout_data:
  root: "${ROLLOUT_DATA_ROOT:/path/to/rollout/data}"  # Set ROLLOUT_DATA_ROOT environment variable or update this path
  xy_file: "M40_XY.txt"
  static_files:
    # Note: M40_XY.txt is included here as a static feature (XY coordinates as features)
    # The xy_file parameter loads it separately for geometry, while static_files includes it as a feature
    # This is intentional for flood modeling where XY coordinates are used both for geometry and as features
    - "M40_XY.txt"
    - "M40_CA.txt"
    - "M40_CE.txt"
    - "M40_CS.txt"
    - "M40_FA.txt"
    - "M40_A.txt"
    - "M40_CU.txt"
  dynamic_patterns:
    WD: "M40_WD_{}.txt"
    VX: "M40_VX_{}.txt"
    VY: "M40_VY_{}.txt"
  boundary_patterns:
    inflow: "M40_US_InF_{}.txt"

# Model configuration (for neuralop get_model compatibility)
model:
  model_arch: 'gino'
  data_channels: 20
  out_channels: 3
  latent_feature_channels: null
  projection_channel_ratio: 4
  gno_coord_dim: 2
  in_gno_radius: 0.1
  out_gno_radius: 0.1
  in_gno_transform_type: 'linear'
  out_gno_transform_type: 'linear'
  gno_weighting_function: null
  gno_weight_function_scale: 1.0
  in_gno_pos_embed_type: 'transformer'
  out_gno_pos_embed_type: 'transformer'
  fno_in_channels: 20
  fno_n_modes: [16, 16]
  fno_hidden_channels: 64
  fno_lifting_channel_ratio: 2
  fno_n_layers: 4
  gno_embed_channels: 32
  gno_embed_max_positions: 10000
  in_gno_channel_mlp_hidden_layers: [80, 80, 80]
  out_gno_channel_mlp_hidden_layers: [512, 256]
  gno_use_open3d: false
  gno_use_torch_scatter: false
  out_gno_tanh: null
  fno_resolution_scaling_factor: null
  fno_block_precision: 'full'
  fno_use_channel_mlp: true
  fno_channel_mlp_dropout: 0
  fno_channel_mlp_expansion: 0.5
  fno_norm: 'instance_norm'
  fno_ada_in_features: 16
  fno_ada_in_dim: 1
  fno_preactivation: false
  fno_skip: 'linear'
  fno_channel_mlp_skip: 'soft-gating'
  fno_separable: false
  fno_factorization: 'tucker'
  fno_rank: 0.4
  fno_fixed_rank_modes: false
  fno_implementation: 'factorized'

# Checkpoint configuration
checkpoint:
  save_dir: "./checkpoints_flood_forecaster"
  resume_from_dir: null
  resume_from_source: null
  resume_from_adapt: null
  # Pretraining checkpoint saving options
  save_best: "source_val_l2"  # Metric to monitor for best model saving (e.g., "source_val_l2")
                              # Set to null to disable best model saving
  save_every: null  # Save checkpoint every N epochs (e.g., 10). Set to null to disable interval saving
                    # Note: save_best takes precedence over save_every if both are set

# Rollout output directory
rollout:
  out_dir: "./rollout_outputs"

# Optimization settings
training:
  n_epochs: 4
  n_epochs_source: 2
  n_epochs_adapt: 2
  learning_rate: 1e-4
  adapt_learning_rate: 1e-4
  training_loss: 'l2'
  testing_loss: 'l2'
  weight_decay: 1e-4
  amp_autocast: false
  scheduler: 'StepLR'
  scheduler_T_max: 200
  scheduler_patience: 5
  step_size: 50
  gamma: 0.5
  da_class_loss_weight: 0.0
  da_lambda_max: 1.0
  da_classifier:
    conv_layers:
      - out_channels: 64
        kernel_size: 3
        pool_size: 2
    fc_dim: 1

# Weights & Biases logging
wandb:
  log: false
  name: null
  group: 'flood-experiments'
  project: 'Flood_GINO_NoPhysics'
  entity: 'uva_mehdi'
  sweep: false
  log_output: true
  eval_interval: 1