为什么分类损失会随着训练而减少本地化和正则化损失呢？

Question

我正在训练COCO预先训练的mobilenetSSDv2模型，并仅在背包上对其进行进一步训练。训练后，随着各种管道配置的变化（学习率，num_steps等），我在对象本地化方面的损失不断减少（在背包周围画一个边界框），但是在分类方面的损失却增加了（通常将其检测为“手提箱”，“滑板” ”或其他一些非“背包”类）。

我已经试验了这些超级参数，并确保我的训练，验证和测试数据都是正确的。可能是此问题的一些潜在原因？

尝试了变化的超参数。检查数据文件是否正确。

用于培训的代码如下所示：

rm -rf /NAME/model_dir_for_training_ssdv2
mkdir /NAME/model_dir_for_training_ssdv2
cd /NAME/tensorflow/models/research 
export PYTHONPATH=/NAME/tensorflow/models/research:/tensorflow/models/research/slim && \
PIPELINE_CONFIG_PATH=/NAME/ssd_mobilenet_v2_coco_2018_03_29/pipeline.config && \
NUM_TRAIN_STEPS=5000 && \
MODEL_DIR=/NAME/model_dir_for_training_ssdv2 && \
SAMPLE_1_OF_N_EVAL_EXAMPLES=1 
python object_detection/model_main.py \
    --pipeline_config_path=${PIPELINE_CONFIG_PATH} \
    --model_dir=${MODEL_DIR} \
    --num_train_steps=${NUM_TRAIN_STEPS} \
    --sample_1_of_n_eval_examples=${SAMPLE_1_OF_N_EVAL_EXAMPLES} \
    --alsologtostderr

管道配置文件如下所示：

model {
  ssd {
    num_classes: 90
    image_resizer {
      fixed_shape_resizer {
        height: 300
        width: 300
      }
    }
    feature_extractor {
      type: "ssd_mobilenet_v2"
      depth_multiplier: 1.0
      min_depth: 16
      conv_hyperparams {
        regularizer {
          l2_regularizer {
            weight: 3.99999989895e-05
          }
        }
        initializer {
          truncated_normal_initializer {
            mean: 0.0
            stddev: 0.0299999993294
          }
        }
        activation: RELU_6
        batch_norm {
          decay: 0.999700009823
          center: true
          scale: true
          epsilon: 0.0010000000475
          train: true
        }
      }
      use_depthwise: 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
      }
    }
    similarity_calculator {
      iou_similarity {
      }
    }
    box_predictor {
      convolutional_box_predictor {
        conv_hyperparams {
          regularizer {
            l2_regularizer {
              weight: 3.99999989895e-05
            }
          }
          initializer {
            truncated_normal_initializer {
              mean: 0.0
              stddev: 0.0299999993294
            }
          }
          activation: RELU_6
          batch_norm {
            decay: 0.999700009823
            center: true
            scale: true
            epsilon: 0.0010000000475
            train: true
          }
        }
        min_depth: 0
        max_depth: 0
        num_layers_before_predictor: 0
        use_dropout: false
        dropout_keep_probability: 0.95
        kernel_size: 3
        box_code_size: 4
        apply_sigmoid_to_scores: false
      }
    }
    anchor_generator {
      ssd_anchor_generator {
        num_layers: 6
        min_scale: 0.20000000298
        max_scale: 0.949999988079
        aspect_ratios: 1.0
        aspect_ratios: 2.0
        aspect_ratios: 0.5
        aspect_ratios: 3.0
        aspect_ratios: 0.333299994469
      }
    }
    post_processing {
      batch_non_max_suppression {
        score_threshold: 0.300000011921
        iou_threshold: 0.600000023842
        max_detections_per_class: 100
        max_total_detections: 10000
      }
      score_converter: SIGMOID
    }
    normalize_loss_by_num_matches: true
    loss {
      localization_loss {
        weighted_smooth_l1 {
        }
      }
      classification_loss {
        weighted_sigmoid {
        }
      }
      hard_example_miner {
        num_hard_examples: 3000
        iou_threshold: 0.990000009537
        loss_type: CLASSIFICATION
        max_negatives_per_positive: 3
        min_negatives_per_image: 3
      }
      classification_weight: 1.0
      localization_weight: 1.0
    }
  }
}
train_config {
  batch_size: 50
  data_augmentation_options {
    random_horizontal_flip {
    }
  }
  data_augmentation_options {
    ssd_random_crop {
    }
  }
  optimizer {
    rms_prop_optimizer {
      learning_rate {
        exponential_decay_learning_rate {
          initial_learning_rate: 0.0000000400000018999
          decay_steps: 800720
          decay_factor: 0.949999988079
        }
      }
      momentum_optimizer_value: 0.899999976158
      decay: 0.899999976158
      epsilon: 1.0
    }
  }
  fine_tune_checkpoint: "/Name/ssd_mobilenet_v2_coco_2018_03_29/model.ckpt"
  from_detection_checkpoint: true
  load_all_detection_checkpoint_vars: true
  fine_tune_checkpoint_type: "detection"
}
train_input_reader {
  label_map_path: "/Name/ssd_mobilenet_v2_coco_2018_03_29/label_map.pbtxt"
  tf_record_input_reader {
    input_path: "/Name/OIDv4_ToolKit/OID/Dataset/annotations/tfrecords/coco_train.record-00000-of-00100"
  }
}
eval_config {
  num_examples: 22
  max_evals: 10
  use_moving_averages: false
  num_visualizations: 20
}
eval_input_reader {
  label_map_path: "/Name/ssd_mobilenet_v2_coco_2018_03_29/label_map.pbtxt"
  shuffle: false
  num_readers: 1
  tf_record_input_reader {
    input_path: "/Name/OIDv4_ToolKit/OID/Dataset/annotations/tfrecords/coco_val.record-00000-of-00010"
  }
}

通过训练更多新颖的数据，可以预期所有类型的损失都会减少。