NAME: SEnSeIv2-SegFormerB2-S2-ambiguous

#-------------
#Model options
#-------------
PATCH_SIZE: 512
SEnSeIv2: 'hf_models/sensei-configs/senseiv2-medium.yaml'
MODEL_TYPE: 'Segformer'
SEGFORMER_CONFIG: 'nvidia/mit-b2'
RECOVERY_MODULE: false
CLASSES: 7
MULTIMODAL: false
NUM_CHANNELS: null # Set to null for sensor independent models

#----------------
#training options (not needed for inference)
#----------------
EPOCHS: 105
BATCH_SIZE: 8
PHASES: [0, 1, 2, 75, 95, 110]
ACCUMULATE_STEPS: [1, 1, 1, 1, 1, 1]
LR: [0.000005, 0.00002, 0.0001, 0.00002, 0.00001, 0.000002]
EPSILON: 0.000001
WEIGHT_DECAY: 0.0001
L1_REG: 0
RECOVERY_WARMUP_STEPS: 10000
RECOVERY_LOSS_FACTOR: 1
LOSS: 'ambiguous_crossentropy_loss'

# Data options
# Note: L7Irish and L8CCA are very large, so repeat other datasets to make up for it
# Training epochs will max out at 4000 steps, so repeating datasets beyond that just
# changes relative frequency of each dataset appearing. 
TRAIN_DIRS:
  - 'path/to/be/set'
VALID_DIRS:
  - 'path/to/be/set'
MIN_BANDS: 3
MAX_BANDS: 13