配置Fast-Rcnn.config以使用Adam优化器和其他参数

Question

I have following fast_rcnn_resnet101_coco.config ( here ). In this config file I have replaced momentum_optimizer with adam optimizer as follows:

train_config: {
  batch_size: 1
  optimizer {
    #momentum_optimizer: {
    adam_optimizer: {
      learning_rate: {
        manual_step_learning_rate {
          initial_learning_rate: 0.00001
          schedule {
            step: 4500
            learning_rate: .00001
          }
          schedule {
            step: 10000
            learning_rate: .000001
          }
        }
      }
      #momentum_optimizer_value: 0.9
    }
    use_moving_average: false
  }
  gradient_clipping_by_norm: 10.0
  fine_tune_checkpoint: "faster_rcnn_resnet101_coco_2018_01_28/model.ckpt"
  from_detection_checkpoint: true
  data_augmentation_options {
    random_horizontal_flip {
    }
  }
}

I have referred Tensorflow Object Detection: use Adam instead of RMSProp to do this change. My aim is to configure my faster rcnnresnet101.config file ( attached here ) to match that of this file:

My aim is that my .config file should have all the parameters mentioned in the .yaml file. So far I have succeed in doing this only for one parameter ("learning rate"). How can I integrate rpn_batch size, step size etc. parameters in my config file ?

Answer 1

The basic fact you need to understand is the following :

The config file is must match the message TrainEvalPipelineConfig . Now that message consists of multiple components. So, if you want to modify something in a component, you should go to the proto file in which that component message is defined, see the possible parameters therein and then modify the configuration file according to that. That is exactly what you ultimately did in order to change the optimizer.

To give you a hint, if you want to change the RPN batch size, you have to modify this parameter . So, look it up in the proto file and simply add it to your final configuration file.

To give you an illustration, if I were to use the original configuration file with one minor change which is a RPN batch size of 128, my configuration file will appear as follows :

# Faster R-CNN with Resnet-101 (v1), configuration for MSCOCO Dataset.
# Users should configure the fine_tune_checkpoint field in the train config as
# well as the label_map_path and input_path fields in the train_input_reader and
# eval_input_reader. Search for "PATH_TO_BE_CONFIGURED" to find the fields that
# should be configured.

model {
  faster_rcnn {
    num_classes: 90
    image_resizer {
      keep_aspect_ratio_resizer {
        min_dimension: 600
        max_dimension: 1024
      }
    }
    feature_extractor {
      type: 'faster_rcnn_resnet101'
      first_stage_features_stride: 16
    }
    first_stage_anchor_generator {
      grid_anchor_generator {
        scales: [0.25, 0.5, 1.0, 2.0]
        aspect_ratios: [0.5, 1.0, 2.0]
        height_stride: 16
        width_stride: 16
      }
    }
    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
    # below i modify the RPN batch size to 128
    first_stage_minibatch_size: 128 
    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
  }
}

train_config: {
  batch_size: 1
  optimizer {
    momentum_optimizer: {
      learning_rate: {
        manual_step_learning_rate {
          initial_learning_rate: 0.0003
          schedule {
            step: 900000
            learning_rate: .00003
          }
          schedule {
            step: 1200000
            learning_rate: .000003
          }
        }
      }
      momentum_optimizer_value: 0.9
    }
    use_moving_average: false
  }
  gradient_clipping_by_norm: 10.0
  fine_tune_checkpoint: "PATH_TO_BE_CONFIGURED/model.ckpt"
  from_detection_checkpoint: true
  data_augmentation_options {
    random_horizontal_flip {
    }
  }
}

train_input_reader: {
  tf_record_input_reader {
    input_path: "PATH_TO_BE_CONFIGURED/mscoco_train.record-?????-of-00100"
  }
  label_map_path: "PATH_TO_BE_CONFIGURED/mscoco_label_map.pbtxt"
}

eval_config: {
  num_examples: 8000
  # Note: The below line limits the evaluation process to 10 evaluations.
  # Remove the below line to evaluate indefinitely.
  max_evals: 10
}

eval_input_reader: {
  tf_record_input_reader {
    input_path: "PATH_TO_BE_CONFIGURED/mscoco_val.record-?????-of-00010"
  }
  label_map_path: "PATH_TO_BE_CONFIGURED/mscoco_label_map.pbtxt"
  shuffle: false
  num_readers: 1
}