ssd_resnet101_v1_fpn_marine_debris.config

# SSD with Resnet 101 v1 FPN feature extractor, shared box predictor and focal
# loss (a.k.a Retinanet).
# See Lin et al, https://arxiv.org/abs/1708.02002
# Trained on open image dataset v4, initialized from scratch.

# This config is TPU compatible 

model {
  ssd {
    inplace_batchnorm_update: true
    freeze_batchnorm: false
    num_classes: 1
    box_coder {
      faster_rcnn_box_coder {
        y_scale: 10.0
        x_scale: 10.0
        height_scale: 5.0
        width_scale: 5.0
      }
    }
    matcher {
      argmax_matcher {
        matched_threshold: 0.5
        unmatched_threshold: 0.5
        ignore_thresholds: false
        negatives_lower_than_unmatched: true
        force_match_for_each_row: true
        use_matmul_gather: true
      }
    }
    similarity_calculator {
      iou_similarity {
      }
    }
    encode_background_as_zeros: true
    anchor_generator {
      multiscale_anchor_generator {
        min_level: 3
        max_level: 7
        anchor_scale: 4.0
        aspect_ratios: [1.0, 2.0, 0.5]
        scales_per_octave: 2
      }
    }
    image_resizer {
      fixed_shape_resizer {
        height: 256
        width: 256
      }
    }
    box_predictor {
      weight_shared_convolutional_box_predictor {
        depth: 256
        class_prediction_bias_init: -4.6
        conv_hyperparams {
          activation: RELU_6,
          regularizer {
            l2_regularizer {
              weight: 0.0001
            }
          }
          initializer {
            random_normal_initializer {
              stddev: 0.01
              mean: 0.0
            }
          }
          batch_norm {
            scale: true,
            decay: 0.997,
            epsilon: 0.001,
          }
        }
        num_layers_before_predictor: 2
        kernel_size: 3
      }
    }
    feature_extractor {
      type: 'ssd_resnet101_v1_fpn'
      fpn {
        min_level: 3
        max_level: 7
      }
      min_depth: 16
      depth_multiplier: 1.0
      conv_hyperparams {
        activation: RELU_6,
        regularizer {
          l2_regularizer {
            weight: 0.0001
          }
        }
        initializer {
          truncated_normal_initializer {
            stddev: 0.03
            mean: 0.0
          }
        }
        batch_norm {
          scale: true,
          decay: 0.997,
          epsilon: 0.001,
        }
      }
      override_base_feature_extractor_hyperparams: true
    }
    loss {
      classification_loss {
        weighted_sigmoid_focal {
          alpha: 0.25
          gamma: 2.0
        }
      }
      localization_loss {
        weighted_smooth_l1 {
        }
      }
      classification_weight: 1.0
      localization_weight: 1.0
    }
    normalize_loss_by_num_matches: true
    normalize_loc_loss_by_codesize: true
    post_processing {
      batch_non_max_suppression {
        score_threshold: 0.1
        iou_threshold: 0.5
        max_detections_per_class: 10
        max_total_detections: 50
        use_static_shapes: true
      }
      score_converter: SIGMOID
    }
  }
}

train_config {
  batch_size: 12  # 14 too big for p3.2xlarge w/ 256 depth, pyr 3-7; 12 works w/ depth 128
  merge_multiple_label_boxes: true
  data_augmentation_options {
    random_horizontal_flip {
    }
  }
  data_augmentation_options {
    random_vertical_flip {
    }
  }
  data_augmentation_options {
    random_adjust_brightness {
    }
  }
  data_augmentation_options {
    random_adjust_contrast {
    }
  }
  data_augmentation_options {
    random_crop_image {
      min_area: 0.75
      max_area: 1.0
    }
  }
  optimizer {
     rms_prop_optimizer {
       learning_rate {
         cosine_decay_learning_rate {
           learning_rate_base: 0.0001
           total_steps: 50000
           warmup_learning_rate: 0.0000001
           warmup_steps: 1000
         }
       }
       momentum_optimizer_value: 0.9}
    # use_moving_average: false
  }
  fine_tune_checkpoint: "/home/ubuntu/tensorflow_od_api/models/research/object_detection/marine_litter_training/ssd_resnet101_v1_fpn_shared_box_predictor_oid_512x512_sync_2019_01_20/model.ckpt"
  from_detection_checkpoint: true  # From object detection checkpoint
  load_all_detection_checkpoint_vars: true
  num_steps: 50000 # Match to momentum_optimizer above
  startup_delay_steps: 0.0
  unpad_groundtruth_tensors: false
  sync_replicas: true
}
train_input_reader {
  label_map_path: "/home/ubuntu/tensorflow_od_api/models/research/object_detection/marine_litter_training/marine_debris.pbtxt"
  shuffle_buffer_size: 8192
  tf_record_input_reader {
    input_path: "/home/ubuntu/tensorflow_od_api/models/research/object_detection/marine_litter_training/data/tf_records/*train.records"
  }
}
eval_config {
  num_examples: 100  # Set to size of eval TF Serving; deprecated
  #max_evals: 1  # Max out at X evalutations to save time; deprecated
  eval_interval_secs: 300  # Default 300
  max_num_boxes_to_visualize: 50
  visualize_groundtruth_boxes: true
  num_visualizations: 12
  use_moving_averages: false
  include_metrics_per_category: true
  metrics_set: "weighted_pascal_voc_detection_metrics"
}
eval_input_reader {
  label_map_path: "/home/ubuntu/tensorflow_od_api/models/research/object_detection/marine_litter_training/marine_debris.pbtxt"
  shuffle: true
  num_readers: 4
  tf_record_input_reader {
    input_path: "/home/ubuntu/tensorflow_od_api/models/research/object_detection/marine_litter_training/data/tf_records/*val.records"
  }
}