Proper config settings for Tensorflow Object Detection API to add a class or do transfer learning

Question

Need to know the proper configuration settings for the Tensorflow Object Detection API to add a class and do transfer learning

After reading https://github.com/tensorflow/models/issues/6479 and Retrain Tensorflow Object detection API it is still unclear on how to do transfer learning with the API.

I'm looking for the proper way to add a class to a trained model. For example, the SSD with Mobilenet v1

The methods I've seen using the object detection API involve making the following changes: In the pipeline config file:

Change num_classes: 90 to num_classes: 1
Change fine_tune_checkpoint: to "../yourlocalpath/model.ckpt
Keep from_detection_checkpoint: true
Change train_input_reader/ input_path: to "../yourtrainimagepath/train.record"
Change train_input_reader/ label_map_path to "../yourlocalpath/classes.pbtxt"
Change eval_input_reader / input_path to "../yourtestimagepath/test.rocord"
Change eval_input_reader / label_map_path to "../yourlocalpath/classes.pbtxt"

Also,

Change the file: "../yourlocalpath/classes.pbtxt" to only contain:

item {
id: 1
name: 'some_new_class'
}

I trained 600 images for 200,000 steps (18 hours) to a loss of 1.5.

I achieved over 90% accuracy on the training data but less than 10% on the evaluation. This was clearly an overfit. My first take was that the model is too complex for a single item. It just memorized the training data. I also noticed that the other 90 original items were no longer found.

I then change the num_classes to 91 and simply added item { id: 91 name: 'some_new_class' } to the original classes.pbtxt file?

My results did not improve much (20%). (This time I stopped training around 100,000 steps but the learning curve pretty much flattened by that point).

For both cases, I chose not to change the "from_detection_checkpoint: true" setting. because "starting from a detection checkpoint will usually result in a faster training job than a classification checkpoint." reference: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/configuring_jobs.md#model-parameter-initialization

What is the proper way to train an object detector to detect all objects (old and new)?

I expect that when I conduct a prediction on an image containing already trained objects in addition to my new object, all are found.

Here are the config files used.

1st one with num_classes: 1

# SSD with Mobilenet v1, configured for Oxford-IIIT Pets 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 {
  ssd {
    num_classes: 1
    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 {
      }
    }
    anchor_generator {
      ssd_anchor_generator {
        num_layers: 6
        min_scale: 0.2
        max_scale: 0.95
        aspect_ratios: 1.0
        aspect_ratios: 2.0
        aspect_ratios: 0.5
        aspect_ratios: 3.0
        aspect_ratios: 0.3333
      }
    }
    image_resizer {
      fixed_shape_resizer {
        height: 300
        width: 300
      }
    }
    box_predictor {
      convolutional_box_predictor {
        min_depth: 0
        max_depth: 0
        num_layers_before_predictor: 0
        use_dropout: false
        dropout_keep_probability: 0.8
        kernel_size: 1
        box_code_size: 4
        apply_sigmoid_to_scores: false
        conv_hyperparams {
          activation: RELU_6,
          regularizer {
            l2_regularizer {
              weight: 0.00004
            }
          }
          initializer {
            truncated_normal_initializer {
              stddev: 0.03
              mean: 0.0
            }
          }
          batch_norm {
            train: true,
            scale: true,
            center: true,
            decay: 0.9997,
            epsilon: 0.001,
          }
        }
      }
    }
    feature_extractor {
      type: 'ssd_mobilenet_v1'
      min_depth: 16
      depth_multiplier: 1.0
      conv_hyperparams {
        activation: RELU_6,
        regularizer {
          l2_regularizer {
            weight: 0.00004
          }
        }
        initializer {
          truncated_normal_initializer {
            stddev: 0.03
            mean: 0.0
          }
        }
        batch_norm {
          train: true,
          scale: true,
          center: true,
          decay: 0.9997,
          epsilon: 0.001,
        }
      }
    }
    loss {
      classification_loss {
        weighted_sigmoid {
        }
      }
      localization_loss {
        weighted_smooth_l1 {
        }
      }
      hard_example_miner {
        num_hard_examples: 3000
        iou_threshold: 0.99
        loss_type: CLASSIFICATION
        max_negatives_per_positive: 3
        min_negatives_per_image: 0
      }
      classification_weight: 1.0
      localization_weight: 1.0
    }
    normalize_loss_by_num_matches: true
    post_processing {
      batch_non_max_suppression {
        score_threshold: 1e-8
        iou_threshold: 0.6
        max_detections_per_class: 100
        max_total_detections: 100
      }
      score_converter: SIGMOID
    }
  }
}

train_config: {
  batch_size: 10
  optimizer {
    rms_prop_optimizer: {
      learning_rate: {
        exponential_decay_learning_rate {
          initial_learning_rate: 0.004
          decay_steps: 800720
          decay_factor: 0.95
        }
      }
      momentum_optimizer_value: 0.9
      decay: 0.9
      epsilon: 1.0
    }
  }
  fine_tune_checkpoint: "/home/adriansr/HoodML/Datasets/ssd_mobilenet_v1_coco_2018_01_28/model.ckpt"
  from_detection_checkpoint: true
  load_all_detection_checkpoint_vars: true
  # Note: The below line limits the training process to 200K steps, which we
  # empirically found to be sufficient enough to train the pets dataset. This
  # effectively bypasses the learning rate schedule (the learning rate will
  # never decay). Remove the below line to train indefinitely.
  num_steps: 200000
  data_augmentation_options {
    random_horizontal_flip {
    }
  }
  data_augmentation_options {
    ssd_random_crop {
    }
  }
}

train_input_reader: {
  tf_record_input_reader {
    input_path: "/home/adriansr/HoodML/Datasets/2016_USATF_Sprint_TrainingDataset/Analyze/train.record"
  }
  label_map_path: "/home/adriansr/HoodML/hoodbibod/training/classes.pbtxt"
}

eval_config: {
  metrics_set: "coco_detection_metrics"
  num_examples: 1100
}

eval_input_reader: {
  tf_record_input_reader {
    input_path: "/home/adriansr/HoodML/Datasets/2016_USATF_Sprint_TrainingDataset/Analyze/test.record"
  }
  label_map_path: "/home/adriansr/HoodML/hoodbibod/training/classes.pbtxt"
  shuffle: false
  num_readers: 1
}

2nd one with num_classes: 91

# SSD with Mobilenet v1, configured for Oxford-IIIT Pets 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 {
  ssd {
    num_classes: 91
    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 {
      }
    }
    anchor_generator {
      ssd_anchor_generator {
        num_layers: 6
        min_scale: 0.2
        max_scale: 0.95
        aspect_ratios: 1.0
        aspect_ratios: 2.0
        aspect_ratios: 0.5
        aspect_ratios: 3.0
        aspect_ratios: 0.3333
      }
    }
    image_resizer {
      fixed_shape_resizer {
        height: 300
        width: 300
      }
    }
    box_predictor {
      convolutional_box_predictor {
        min_depth: 0
        max_depth: 0
        num_layers_before_predictor: 0
        use_dropout: false
        dropout_keep_probability: 0.8
        kernel_size: 1
        box_code_size: 4
        apply_sigmoid_to_scores: false
        conv_hyperparams {
          activation: RELU_6,
          regularizer {
            l2_regularizer {
              weight: 0.00004
            }
          }
          initializer {
            truncated_normal_initializer {
              stddev: 0.03
              mean: 0.0
            }
          }
          batch_norm {
            train: true,
            scale: true,
            center: true,
            decay: 0.9997,
            epsilon: 0.001,
          }
        }
      }
    }
    feature_extractor {
      type: 'ssd_mobilenet_v1'
      min_depth: 16
      depth_multiplier: 1.0
      conv_hyperparams {
        activation: RELU_6,
        regularizer {
          l2_regularizer {
            weight: 0.00004
          }
        }
        initializer {
          truncated_normal_initializer {
            stddev: 0.03
            mean: 0.0
          }
        }
        batch_norm {
          train: true,
          scale: true,
          center: true,
          decay: 0.9997,
          epsilon: 0.001,
        }
      }
    }
    loss {
      classification_loss {
        weighted_sigmoid {
        }
      }
      localization_loss {
        weighted_smooth_l1 {
        }
      }
      hard_example_miner {
        num_hard_examples: 3000
        iou_threshold: 0.99
        loss_type: CLASSIFICATION
        max_negatives_per_positive: 3
        min_negatives_per_image: 0
      }
      classification_weight: 1.0
      localization_weight: 1.0
    }
    normalize_loss_by_num_matches: true
    post_processing {
      batch_non_max_suppression {
        score_threshold: 1e-8
        iou_threshold: 0.6
        max_detections_per_class: 100
        max_total_detections: 100
      }
      score_converter: SIGMOID
    }
  }
}

train_config: {
  batch_size: 10
  optimizer {
    rms_prop_optimizer: {
      learning_rate: {
        exponential_decay_learning_rate {
          initial_learning_rate: 0.004
          decay_steps: 800720
          decay_factor: 0.95
        }
      }
      momentum_optimizer_value: 0.9
      decay: 0.9
      epsilon: 1.0
    }
  }
  fine_tune_checkpoint: "/home/adriansr/HoodML/Datasets/ssd_mobilenet_v1_coco_2018_01_28/model.ckpt"
  from_detection_checkpoint: true
  load_all_detection_checkpoint_vars: true
  # Note: The below line limits the training process to 200K steps, which we
  # empirically found to be sufficient enough to train the pets dataset. This
  # effectively bypasses the learning rate schedule (the learning rate will
  # never decay). Remove the below line to train indefinitely.
  num_steps: 200000
  data_augmentation_options {
    random_horizontal_flip {
    }
  }
  data_augmentation_options {
    ssd_random_crop {
    }
  }
}

train_input_reader: {
  tf_record_input_reader {
    input_path: "/home/adriansr/HoodML/Datasets/2016_USATF_Sprint_TrainingDataset/Analyze/train.record"
  }
  label_map_path: "/home/adriansr/HoodML/hoodbibod/training/mscoco_complete_label_map_with_bib.pbtxt"
}

eval_config: {
  metrics_set: "coco_detection_metrics"
  num_examples: 1100
}

eval_input_reader: {
  tf_record_input_reader {
    input_path: "/home/adriansr/HoodML/Datasets/2016_USATF_Sprint_TrainingDataset/Analyze/test.record"
  }
  label_map_path: "/home/adriansr/HoodML/hoodbibod/training/mscoco_complete_label_map_with_bib.pbtxt"
  shuffle: false
  num_readers: 1
}

classes.pbtxt

item {
  id: 1
  name: 'Bib'
}

mscoco_complete_label_map_with_bib.pbtxt

item {
  name: "background"
  id: 0
  display_name: "background"
}
item {
  name: "/m/01g317"
  id: 1
  display_name: "person"
}
item {
  name: "/m/0199g"
  id: 2
  display_name: "bicycle"
}
item {
  name: "/m/0k4j"
  id: 3
  display_name: "car"
}
item {
  name: "/m/04_sv"
  id: 4
  display_name: "motorcycle"
}
item {
  name: "/m/05czz6l"
  id: 5
  display_name: "airplane"
}
item {
  name: "/m/01bjv"
  id: 6
  display_name: "bus"
}
item {
  name: "/m/07jdr"
  id: 7
  display_name: "train"
}
item {
  name: "/m/07r04"
  id: 8
  display_name: "truck"
}
item {
  name: "/m/019jd"
  id: 9
  display_name: "boat"
}
item {
  name: "/m/015qff"
  id: 10
  display_name: "traffic light"
}
item {
  name: "/m/01pns0"
  id: 11
  display_name: "fire hydrant"
}
item {
  name: "12"
  id: 12
  display_name: "12"
}
item {
  name: "/m/02pv19"
  id: 13
  display_name: "stop sign"
}
item {
  name: "/m/015qbp"
  id: 14
  display_name: "parking meter"
}
item {
  name: "/m/0cvnqh"
  id: 15
  display_name: "bench"
}
item {
  name: "/m/015p6"
  id: 16
  display_name: "bird"
}
item {
  name: "/m/01yrx"
  id: 17
  display_name: "cat"
}
item {
  name: "/m/0bt9lr"
  id: 18
  display_name: "dog"
}
item {
  name: "/m/03k3r"
  id: 19
  display_name: "horse"
}
item {
  name: "/m/07bgp"
  id: 20
  display_name: "sheep"
}
item {
  name: "/m/01xq0k1"
  id: 21
  display_name: "cow"
}
item {
  name: "/m/0bwd_0j"
  id: 22
  display_name: "elephant"
}
item {
  name: "/m/01dws"
  id: 23
  display_name: "bear"
}
item {
  name: "/m/0898b"
  id: 24
  display_name: "zebra"
}
item {
  name: "/m/03bk1"
  id: 25
  display_name: "giraffe"
}
item {
  name: "26"
  id: 26
  display_name: "26"
}
item {
  name: "/m/01940j"
  id: 27
  display_name: "backpack"
}
item {
  name: "/m/0hnnb"
  id: 28
  display_name: "umbrella"
}
item {
  name: "29"
  id: 29
  display_name: "29"
}
item {
  name: "30"
  id: 30
  display_name: "30"
}
item {
  name: "/m/080hkjn"
  id: 31
  display_name: "handbag"
}
item {
  name: "/m/01rkbr"
  id: 32
  display_name: "tie"
}
item {
  name: "/m/01s55n"
  id: 33
  display_name: "suitcase"
}
item {
  name: "/m/02wmf"
  id: 34
  display_name: "frisbee"
}
item {
  name: "/m/071p9"
  id: 35
  display_name: "skis"
}
item {
  name: "/m/06__v"
  id: 36
  display_name: "snowboard"
}
item {
  name: "/m/018xm"
  id: 37
  display_name: "sports ball"
}
item {
  name: "/m/02zt3"
  id: 38
  display_name: "kite"
}
item {
  name: "/m/03g8mr"
  id: 39
  display_name: "baseball bat"
}
item {
  name: "/m/03grzl"
  id: 40
  display_name: "baseball glove"
}
item {
  name: "/m/06_fw"
  id: 41
  display_name: "skateboard"
}
item {
  name: "/m/019w40"
  id: 42
  display_name: "surfboard"
}
item {
  name: "/m/0dv9c"
  id: 43
  display_name: "tennis racket"
}
item {
  name: "/m/04dr76w"
  id: 44
  display_name: "bottle"
}
item {
  name: "45"
  id: 45
  display_name: "45"
}
item {
  name: "/m/09tvcd"
  id: 46
  display_name: "wine glass"
}
item {
  name: "/m/08gqpm"
  id: 47
  display_name: "cup"
}
item {
  name: "/m/0dt3t"
  id: 48
  display_name: "fork"
}
item {
  name: "/m/04ctx"
  id: 49
  display_name: "knife"
}
item {
  name: "/m/0cmx8"
  id: 50
  display_name: "spoon"
}
item {
  name: "/m/04kkgm"
  id: 51
  display_name: "bowl"
}
item {
  name: "/m/09qck"
  id: 52
  display_name: "banana"
}
item {
  name: "/m/014j1m"
  id: 53
  display_name: "apple"
}
item {
  name: "/m/0l515"
  id: 54
  display_name: "sandwich"
}
item {
  name: "/m/0cyhj_"
  id: 55
  display_name: "orange"
}
item {
  name: "/m/0hkxq"
  id: 56
  display_name: "broccoli"
}
item {
  name: "/m/0fj52s"
  id: 57
  display_name: "carrot"
}
item {
  name: "/m/01b9xk"
  id: 58
  display_name: "hot dog"
}
item {
  name: "/m/0663v"
  id: 59
  display_name: "pizza"
}
item {
  name: "/m/0jy4k"
  id: 60
  display_name: "donut"
}
item {
  name: "/m/0fszt"
  id: 61
  display_name: "cake"
}
item {
  name: "/m/01mzpv"
  id: 62
  display_name: "chair"
}
item {
  name: "/m/02crq1"
  id: 63
  display_name: "couch"
}
item {
  name: "/m/03fp41"
  id: 64
  display_name: "potted plant"
}
item {
  name: "/m/03ssj5"
  id: 65
  display_name: "bed"
}
item {
  name: "66"
  id: 66
  display_name: "66"
}
item {
  name: "/m/04bcr3"
  id: 67
  display_name: "dining table"
}
item {
  name: "68"
  id: 68
  display_name: "68"
}
item {
  name: "69"
  id: 69
  display_name: "69"
}
item {
  name: "/m/09g1w"
  id: 70
  display_name: "toilet"
}
item {
  name: "71"
  id: 71
  display_name: "71"
}
item {
  name: "/m/07c52"
  id: 72
  display_name: "tv"
}
item {
  name: "/m/01c648"
  id: 73
  display_name: "laptop"
}
item {
  name: "/m/020lf"
  id: 74
  display_name: "mouse"
}
item {
  name: "/m/0qjjc"
  id: 75
  display_name: "remote"
}
item {
  name: "/m/01m2v"
  id: 76
  display_name: "keyboard"
}
item {
  name: "/m/050k8"
  id: 77
  display_name: "cell phone"
}
item {
  name: "/m/0fx9l"
  id: 78
  display_name: "microwave"
}
item {
  name: "/m/029bxz"
  id: 79
  display_name: "oven"
}
item {
  name: "/m/01k6s3"
  id: 80
  display_name: "toaster"
}
item {
  name: "/m/0130jx"
  id: 81
  display_name: "sink"
}
item {
  name: "/m/040b_t"
  id: 82
  display_name: "refrigerator"
}
item {
  name: "83"
  id: 83
  display_name: "83"
}
item {
  name: "/m/0bt_c3"
  id: 84
  display_name: "book"
}
item {
  name: "/m/01x3z"
  id: 85
  display_name: "clock"
}
item {
  name: "/m/02s195"
  id: 86
  display_name: "vase"
}
item {
  name: "/m/01lsmm"
  id: 87
  display_name: "scissors"
}
item {
  name: "/m/0kmg4"
  id: 88
  display_name: "teddy bear"
}
item {
  name: "/m/03wvsk"
  id: 89
  display_name: "hair drier"
}
item {
  name: "/m/012xff"
  id: 90
  display_name: "toothbrush"
}
item {
  name: "/m/bib"
  id: 91
  display_name: "bib"
}

By epochs you mean steps, right? (An epoch is a complete cycle of the whole dataset). Could u also provide the full config file that you used for a single class training? How does the validation accuracy behave, it never goes above 10% or it dropped from some high accuracy to 10%. It would also be better if you can provide some training plots too. — danyfang
@danyfang... yes, I meant steps instead of epochs. I made the change. Also, I added the information from the config files and .pbtxt files. I measured the test accuracy and train accuracy by running predictions on each dataset and simply counting the ones that were correct and the ones that were not correct. Even though tensorflow could give me the values, I chose to write a separate code so that I can do additional processing on the fed images that that were missed, for example, rotate +/- 10 degrees, boost contrast, distort, etc. It did not help much. — Adrian Hood Sr
Also, the validation accuracy dropped fast (< 1000 steps for both training cases) and stayed between 1 and 2 for the duration, occasionally dropping below 1. I'll add the Tensorboard output shortly. — Adrian Hood Sr
what is your validation dataset size? In the config file the validation dataset size is 1100, which is bigger than your training dataset size. — danyfang
In the config file, under filed eval_config, the num_examples should be set equal to the size of your validation dataset size. — danyfang

0rhisia0 0rhisia0 · Accepted Answer · 2021-07-21T15:09:32

2 years late but... Fundamentally, you aren't able to train your network on a new class and have it not affect the identification accuracy of previously trained classes. By training on a dataset of new objects, and with a label map only containg that new object the model will only optimize to detect the new object, as you are changing the weights that enabled the detection of the old objects. You could try mergin your dataset with the one the model was originally trained on, and the training on the new merged set. Even this will be inadequate unless you are planning on somehow making sure the new object is featured in images with the old objects labelled as well (maybe some sort of synthetic data generation procedure may be useful).

Proper config settings for Tensorflow Object Detection API to add a class or do transfer learning

1 Answers