tensorflow/models

# Faster RCNN with Resnet 50 v1 FPN feature extractor.
# See Lin et al, https://arxiv.org/abs/1612.03144
# Trained on COCO, initialized from Imagenet classification checkpoint
# Train on TPU-8
#
# Achieves 31.4 mAP on COCO17 Val

model {
  faster_rcnn {
    num_classes: 90
    image_resizer {
      keep_aspect_ratio_resizer {
        min_dimension: 640
        max_dimension: 640
        pad_to_max_dimension: true
      }
    }
    feature_extractor {
      type: 'faster_rcnn_resnet50_fpn_keras'
      batch_norm_trainable: true
      fpn {
        min_level: 2
        max_level: 6
      }
      conv_hyperparams {
        activation: RELU_6,
        regularizer {
          l2_regularizer {
            weight: 0.0004
          }
        }
        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
    }
    first_stage_anchor_generator {
        multiscale_anchor_generator {
          min_level: 2
          max_level: 6
          # According to the origial paper the value should be 8.0
          anchor_scale: 4.0
          aspect_ratios: [1.0, 2.0, 0.5]
          # According to the original paper the value should be 1
          scales_per_octave: 2
          normalize_coordinates: false
        }
    }
    first_stage_box_predictor_conv_hyperparams {
      op: CONV
      regularizer {
        l2_regularizer {
          weight: 0.0
        }
      }
      initializer {
        truncated_normal_initializer {
          stddev: 0.01
        }
      }
    }
    first_stage_nms_score_threshold: 0.0
    first_stage_nms_iou_threshold: 0.7
    first_stage_max_proposals: 300
    first_stage_localization_loss_weight: 2.0
    first_stage_objectness_loss_weight: 1.0
    # According to the origial paper, value should be 7.
    initial_crop_size: 14
    maxpool_kernel_size: 2
    maxpool_stride: 2
    second_stage_box_predictor {
      mask_rcnn_box_predictor {
        use_dropout: false
        dropout_keep_probability: 1.0
        fc_hyperparams {
          op: FC
          regularizer {
            l2_regularizer {
              weight: 0.0
            }
          }
          initializer {
            variance_scaling_initializer {
              factor: 1.0
              uniform: true
              mode: FAN_AVG
            }
          }
        }
      }
    }
    second_stage_post_processing {
      batch_non_max_suppression {
        score_threshold: 0.0
        iou_threshold: 0.6
        max_detections_per_class: 100
        max_total_detections: 300
      }
      score_converter: SOFTMAX
    }
    second_stage_localization_loss_weight: 2.0
    second_stage_classification_loss_weight: 1.0
    use_static_shapes: true
    use_matmul_crop_and_resize: true
    clip_anchors_to_image: true
    use_static_balanced_label_sampler: true
    use_matmul_gather_in_matcher: true
  }
}

train_config: {
  batch_size: 64
  sync_replicas: true
  startup_delay_steps: 0
  replicas_to_aggregate: 8
  num_steps: 25000
  optimizer {
    momentum_optimizer: {
      learning_rate: {
        cosine_decay_learning_rate {
          learning_rate_base: 0.04
          total_steps: 25000
          warmup_learning_rate: .013333
          warmup_steps: 2000
        }
      }
      momentum_optimizer_value: 0.9
    }
    use_moving_average: false
  }
  fine_tune_checkpoint_version: V2
  fine_tune_checkpoint: "PATH_TO_BE_CONFIGURED/resnet50.ckpt-1"
  fine_tune_checkpoint_type: "classification"
  data_augmentation_options {
    random_horizontal_flip {
    }
  }

  max_number_of_boxes: 100
  unpad_groundtruth_tensors: false
  use_bfloat16: true
}

train_input_reader: {
  tf_record_input_reader {
    input_path: "PATH_TO_BE_CONFIGURED/train2017-?????-of-00256.tfrecord"
  }
  label_map_path: "PATH_TO_BE_CONFIGURED/label_map.txt"
}

eval_config: {
  metrics_set: "coco_detection_metrics"
  use_moving_averages: false
  batch_size: 1;
}

eval_input_reader: {
  label_map_path: "PATH_TO_BE_CONFIGURED/label_map.txt"
  shuffle: false
  num_epochs: 1
  tf_record_input_reader {
    input_path: "PATH_TO_BE_CONFIGURED/val2017-?????-of-00032.tfrecord"
  }
}