coco_spinenet143_cascadercnn_tpu.yaml

# --experiment_type=cascadercnn_spinenet_coco
# Expect to reach: box mAP: 51.9%, mask mAP: 45.0% on COCO
runtime:
  distribution_strategy: 'tpu'
  mixed_precision_dtype: 'bfloat16'
task:
  init_checkpoint: null
  train_data:
    global_batch_size: 256
    parser:
      aug_rand_hflip: true
      aug_scale_min: 0.1
      aug_scale_max: 2.5
  losses:
    l2_weight_decay: 0.00004
  model:
    anchor:
      anchor_size: 4.0
      num_scales: 3
    min_level: 3
    max_level: 7
    input_size: [1280, 1280, 3]
    backbone:
      spinenet:
        stochastic_depth_drop_rate: 0.2
        model_id: '143'
      type: 'spinenet'
    decoder:
      type: 'identity'
    detection_head:
      cascade_class_ensemble: true
      class_agnostic_bbox_pred: true
    rpn_head:
      num_convs: 2
      num_filters: 256
    roi_sampler:
      cascade_iou_thresholds: [0.7]
      foreground_iou_threshold: 0.6
    norm_activation:
      norm_epsilon: 0.001
      norm_momentum: 0.99
      use_sync_bn: true
      activation: 'swish'
    detection_generator:
      pre_nms_top_k: 1000
trainer:
  train_steps: 231000
  optimizer_config:
    learning_rate:
      type: 'stepwise'
      stepwise:
        boundaries: [219450, 226380]
        values: [0.32, 0.032, 0.0032]
    warmup:
      type: 'linear'
      linear:
        warmup_steps: 2000
        warmup_learning_rate: 0.0067