Tensorflow Object Detection 在同一模型检查点的两次评估中得到不同的结果

Question

我是 tensorflow 的菜鸟，目前从事对象检测 API。

我选择了 ssd_resnet50_fpn 开始并从 tensorflow model zoo 下载了预训练模型，用我自己的数据集进行迁移学习，只有 1 class（人）。训练配置是在 pipeline.config 中定义的，它是从预训练模型的同一包中的那个修改而来的，我用遗留的 train.py 脚本训练了模型。

训练过程很好，损失按预期减少，并且我在进行评估时导出了带有边界框的评估图像（通过旧版 eval.py）。推理很好，导出的图像按预期工作。

但是，我发现了一些奇怪的事情，我在同一个模型检查点上得到了不同的评估结果。当我用相同的参数执行两次 eval.py 时，我发现在同一张图像上推断出的边界框是不同的。

这是eval.py(coco_detection_metrics)

的评价结果

First time

Second time

由于我对mAP不是很了解，但两次结果略有不同。

这里是评测时导出的一张图，左边是第一次评测，右边是第二次评测

Exported Image during evaluation

似乎模型权重在推理过程中发生了变化，我怎样才能找出问题所在？有没有我错过的配置？

我正在使用带有 python 3.5.2 的 tensorflow 1.10.1 和 https://github.com/tensorflow/models 中的克隆对象检测 API，没有任何变化。

这是我的 pipeline.config:

model {
  ssd {
    num_classes: 1
    image_resizer {
      fixed_shape_resizer {
        height: 640
        width: 640
      }
    }
    feature_extractor {
      type: "ssd_resnet50_v1_fpn"
      depth_multiplier: 1.0
      min_depth: 16
      conv_hyperparams {
        regularizer {
          l2_regularizer {
            weight: 0.000399999989895
          }
        }
        initializer {
          truncated_normal_initializer {
            mean: 0.0
            stddev: 0.0299999993294
          }
        }
        activation: RELU_6
        batch_norm {
          decay: 0.996999979019
          scale: true
          epsilon: 0.0010000000475
        }
      }
      override_base_feature_extractor_hyperparams: true
    }
    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 {
      }
    }
    box_predictor {
      weight_shared_convolutional_box_predictor {
        conv_hyperparams {
          regularizer {
            l2_regularizer {
              weight: 0.000399999989895
            }
          }
          initializer {
            random_normal_initializer {
              mean: 0.0
              stddev: 0.00999999977648
            }
          }
          activation: RELU_6
          batch_norm {
            decay: 0.996999979019
            scale: true
            epsilon: 0.0010000000475
          }
        }
        use_dropout: true
        dropout_keep_probability: 0.7
        depth: 256
        num_layers_before_predictor: 4
        kernel_size: 3
        class_prediction_bias_init: -4.59999990463
      }
    }
    anchor_generator {
      multiscale_anchor_generator {
        min_level: 3
        max_level: 7
        anchor_scale: 4.0
        aspect_ratios: 1.0
        aspect_ratios: 2.0
        aspect_ratios: 0.5
        scales_per_octave: 2
      }
    }
    post_processing {
      batch_non_max_suppression {
        score_threshold: 0.300000011921
        iou_threshold: 0.600000023842
        max_detections_per_class: 100
        max_total_detections: 100
      }
      score_converter: SIGMOID
    }
    normalize_loss_by_num_matches: true
    loss {
      localization_loss {
        weighted_smooth_l1 {
        }
      }
      classification_loss {
        weighted_sigmoid_focal {
          gamma: 2.0
          alpha: 0.25
        }
      }
      classification_weight: 1.0
      localization_weight: 1.0
    }
    encode_background_as_zeros: true
    normalize_loc_loss_by_codesize: true
    inplace_batchnorm_update: true
    freeze_batchnorm: false
  }
}
train_config {
  batch_size: 8
  data_augmentation_options {
    random_horizontal_flip {
    }
  }
  data_augmentation_options {
    random_crop_image {
      min_object_covered: 0.0
      min_aspect_ratio: 0.75
      max_aspect_ratio: 3.0
      min_area: 0.75
      max_area: 1.0
      overlap_thresh: 0.0
    }
  }
  sync_replicas: false
  optimizer {
    adam_optimizer: {
      learning_rate: {
        exponential_decay_learning_rate {
          initial_learning_rate: 0.0001
          decay_steps: 5000
          decay_factor: 0.9
        }
      }
    }
    use_moving_average: false
  }
  fine_tune_checkpoint: "/tf-object-detection-training/models/ssd_resnet50/saved/model.ckpt-652123"
  num_steps: 2000000
  from_detection_checkpoint: true
  load_all_detection_checkpoint_vars: true
  startup_delay_steps: 0.0
  replicas_to_aggregate: 8
  max_number_of_boxes: 100
  unpad_groundtruth_tensors: false
}
train_input_reader {
  label_map_path: "/tf-object-detection-training/dataset_VOC/label.pbtxt"
  tf_record_input_reader {
    input_path: "/tf-object-detection-training/dataset_VOC/person_train.record-?????-of-00010"
  }
}
eval_config {
  num_examples: 10000
  num_visualizations: 100
  eval_interval_secs: 60
  metrics_set: "coco_detection_metrics"
  use_moving_averages: false
  min_score_threshold: 0.5
  retain_original_images: false
  keep_image_id_for_visualization_export: true
  visualization_export_dir: "/tf-object-detection-training/models/ssd_resnet50/eval_detections/"
}
eval_input_reader {
  label_map_path: "/tf-object-detection-training/dataset_VOC/label.pbtxt"
  shuffle: false
  num_readers: 1
  tf_record_input_reader {
    input_path: "/tf-object-detection-training/dataset_VOC/person_val.record-?????-of-00010"
  }
}

感谢任何建议

Answer 1

查了半天代码，发现是在pipeline.config中设置了'use_dropout'标志位。似乎在进行推理时没有删除 dropout 函数，因此 eval.py 和 frozen_inference_graph 都应用了 dropout 函数并进行随机推理。

要解决此问题，只需从 pipeline.config 中删除 'use_dropout' 即可解决此问题。

Tensorflow Object Detection 在同一模型检查点的两次评估中得到不同的结果

Tensorflow Object Detection got different results in two evaluations on the same model checkpoint

tensorflow

object-detection-api

tensorflow-slim