Tensorflow object detection API overfitting

Viewed 257

I am creating tree detector using Tensoflow Object Detection API with the help of following tutorial : https://www.youtube.com/watch?v=a1br6gW-8Ss

My problem is that the model is overfitting the training set. How can I avoid it ? My training set is composed of 130 images and on average 4 tree is labeled in image. There are 7 species at all. The model I use is : 'Faster R-CNN ResNet50 V1 640x640' pre-trained on COCO dataset..(Link is below) https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2_detection_zoo.md In tutorial tutor uses the following model 'EfficientDet D0 512x512' but that model also overfits my dataset so I tried to use 'Faster R-CNN ResNet50 V1 640x640'.

The techniques I use to reduce overfitting is

  1. dropout(rate=0.675)
  2. data augmentation(random_crop)
  3. Low learning rate(learning_rate_base: .008 , total_steps: 25000 .warmup_learning_rate: .0001)
  4. original images dimensions were 860x860 and I resized them tp 640x640 and changed bounding box coordinates also in xml.

What should I add or change in model configuration to avoid overfitting ?

Here is my model configuration file:

# Faster R-CNN with Resnet-50 (v1) with 640x640 input resolution
# Trained on COCO, initialized from Imagenet classification checkpoint
#
# Train on TPU-8
#
# Achieves 29.3 mAP on COCO17 Val

model {
  faster_rcnn {
    num_classes: 7
    image_resizer {
      keep_aspect_ratio_resizer {
        min_dimension: 640
        max_dimension: 640
        pad_to_max_dimension: true
      }
    }
    feature_extractor {
      type: 'faster_rcnn_resnet50_keras'
      batch_norm_trainable: true
    }
    first_stage_anchor_generator {
      grid_anchor_generator {
        scales: [0.25, 0.5, 1.0, 2.0]
        aspect_ratios: [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5]
        height_stride: 8
        width_stride: 8
      }
    }
    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
    initial_crop_size: 14
    maxpool_kernel_size: 2
    maxpool_stride: 2
    second_stage_box_predictor {
      mask_rcnn_box_predictor {
        use_dropout: true
        dropout_keep_probability: 0.675
        fc_hyperparams {
          op: FC
          regularizer {
            l2_regularizer {
              weight: 0.0
            }
          }
          initializer {
            variance_scaling_initializer {
              factor: 1.0
              uniform: true
              mode: FAN_AVG
            }
          }
        }
        share_box_across_classes: true
      }
    }
    second_stage_post_processing {
      batch_non_max_suppression {
        score_threshold: 0.2
        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: 4
  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: .008
          total_steps: 25000
          warmup_learning_rate: .0001
          warmup_steps: 2000
        }
      }
      momentum_optimizer_value: 0.9
    }
    use_moving_average: false
  }
  fine_tune_checkpoint_version: V2
  fine_tune_checkpoint: "faster_rcnn_resnet50_v1_640x640_coco17_tpu-8/checkpoint/ckpt-0"
  fine_tune_checkpoint_type: "detection"
    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
    }
  }

  max_number_of_boxes: 100
  unpad_groundtruth_tensors: false
  use_bfloat16: true  # works only on TPUs
}

train_input_reader: {
  label_map_path: "images/labelmap.pbtxt"
  tf_record_input_reader {
    input_path: "train.record"
  }
}

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

eval_input_reader: {
  label_map_path: "images/labelmap.pbtxt"
  shuffle: false
  num_epochs: 1
  tf_record_input_reader {
    input_path: "test.record"
  }
}

1 Answers

Along with the measures you are taking to reduce overfit, you can add few more and make some changes to the existing one's.

  • Make sure you shuffle the data well, so that data is equally distributed in both train and test set.
  • Try some more augmentation techniques(ref) or try collecting few more data, since the existing data set looks very small, which is resulting in overfitting.
  • Try reducing the dropout from 0.67 to something like 0.2.
  • Add regularizer techniques which will reduce overfitting.
  • Use earlystopping, when there is not much improvement in the accuracy for the subsequent epochs, you can call earlystopping.
Related