architecture.yaml

training:
  lr: 0.01
  weight-decay: 0.0000001
  bs: 16
  scheduler: 'steplr'
  gamma: 0.1
  step-size: 15
  rebalancing-fake: 0.3
  rebalancing-real: 1
  frames-per-video: 30 # Equidistant frames

model:
  image-size: 224
  num-classes: 1
  depth: 4              # number of multi-scale encoding blocks
  sm-dim: 192            # high res dimension
  sm-patch-size: 7      # high res patch size (should be smaller than lg-patch-size)
  sm-enc-depth: 2        # high res depth
  sm-enc-dim-head: 64        
  sm-enc-heads: 8        # high res heads
  sm-enc-mlp-dim: 2048   # high res feedforward dimension
  lg-dim: 384            # low res dimension
  lg-patch-size: 56      # low res patch size
  lg-enc-depth: 3        # low res depth
  lg-enc-dim-head: 64    
  lg-enc-heads: 8        # low res heads
  lg-enc-mlp-dim: 2048   # low res feedforward dimensions
  cross-attn-depth: 2    # cross attention rounds
  cross-attn-dim-head: 64    
  cross-attn-heads: 8    # cross attention heads
  lg-channels: 24
  sm-channels: 1280
  dropout: 0.15
  emb-dropout: 0.15