QCX-TERRA/terramind_config.yaml at main · QueueLab/QCX-TERRA · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
# terramind_config.yaml
model:
  _target_: terratorch.models.backbones.terramind_v1_base_tim  # Use _tim for Thinking-in-Modalities
  pretrained: true
  modalities: ['S2L2A']  # Input modalities (add 'S1GRD' if using)
  tim_modalities: ['LULC']  # Target for fine-tune (land-use land-cover)
  timesteps: 10  # For generation; ignore for backbone
  standardize: true

data:
  _target_: terratorch.data.terramind_datamodule.TerraMindDataModule
  train_path: ./data/train/  # Folder with .tif files (paired S2L2A + LULC labels)
  val_path: ./data/val/
  batch_size: 4  # Adjust for GPU memory
  num_workers: 4
  modalities: ['S2L2A']
  bands:  # Subset bands if needed
    S2L2A: ['B2', 'B3', 'B4', 'B8', 'B11', 'B12']  # BLUE, GREEN, RED, NIR, SWIR1, SWIR2
  target_size: [224, 224]
  tim_modalities: ['LULC']

trainer:
  accelerator: gpu
  devices: 1
  max_epochs: 5  # Start small
  precision: 16-mixed  # For efficiency
  log_every_n_steps: 10

callbacks:
  model_checkpoint:
    _target_: pytorch_lightning.callbacks.ModelCheckpoint
    monitor: val_loss
    mode: min
    save_top_k: 1
    filename: 'terramind-finetuned-{epoch:02d}-{val_loss:.2f}'

logger:
  _target_: pytorch_lightning.loggers.TensorBoardLogger
  save_dir: ./logs/