[딥러닝]논문에 자주 등장하는 validation이 대체 뭘까?

일반적으로 데이터셋 구조를 생성할 때 train, test 이렇게 두 가지를 만들 것이다.

하지만 좀 더 깊이 들어가서 논문을 읽다보면, validation이라는 단어가 굉장히 자주 등장하고 이게 대체 뭘까 하는 의문이 들 것이다.

---

사전적 의미

국민 번역 프로그램 파파고에 의하면 validation의 사전적 의미는 '확인'이라고 한다. 여기서 우리는 validation이 뭘 뜻하는지 어느정도 유추가 가능하다.

---

딥러닝에서의 'validation'

validation은 학습이 완료된 모델을 검증하기 위한 데이터셋이다. 즉, 중간 점검을 위한 것이다.

딥러닝 분야에서 대부분은 train 데이터셋으로 학습을 시키고, 정확한 성능 평가를 위해 중간 피드백을 하고 수정을 거쳐 최종 output이 추출된다. 이를 위해 validation 데이터셋이 필요한 것이다.

def save_validation_result(model, ckpt, validation_summary_writer, num_visualize_image):
  total_validation_total_loss = 0.0
  total_validation_coord_loss = 0.0
  total_validation_object_loss = 0.0
  total_validation_noobject_loss = 0.0
  total_validation_class_loss = 0.0
  for iter, features in enumerate(validation_data):
    batch_validation_image = features['image']
    batch_validation_bbox = features['objects']['bbox']
    batch_validation_labels = features['objects']['label']

    batch_validation_image = tf.squeeze(batch_validation_image, axis=1)
    batch_validation_bbox = tf.squeeze(batch_validation_bbox, axis=1)
    batch_validation_labels = tf.squeeze(batch_validation_labels, axis=1)

    validation_total_loss, validation_coord_loss, validation_object_loss, validation_noobject_loss, validation_class_loss = calculate_loss(model, batch_validation_image, batch_validation_bbox, batch_validation_labels)

    total_validation_total_loss = total_validation_total_loss + validation_total_loss
    total_validation_coord_loss = total_validation_coord_loss + validation_coord_loss
    total_validation_object_loss = total_validation_object_loss + validation_object_loss
    total_validation_noobject_loss = total_validation_noobject_loss + validation_noobject_loss
    total_validation_class_loss = total_validation_class_loss + validation_class_loss

  #validation tensorboard log 저장
  with validation_summary_writer.as_default():
    tf.summary.scalar('total_validation_total_loss', total_validation_total_loss, step=int(ckpt.step))
    tf.summary.scalar('total_validation_coord_loss', total_validation_coord_loss, step=int(ckpt.step))
    tf.summary.scalar('total_validation_object_loss ', total_validation_object_loss, step=int(ckpt.step))
    tf.summary.scalar('total_validation_noobject_loss ', total_validation_noobject_loss, step=int(ckpt.step))
    tf.summary.scalar('total_validation_class_loss ', total_validation_class_loss, step=int(ckpt.step))

  # validation test image 저장
  for validation_image_index in range(num_visualize_image):
    random_idx = random.randint(0, batch_validation_image.shape[0] - 1)
    image, labels, object_num = process_each_ground_truth(batch_validation_image[random_idx], batch_validation_bbox[random_idx],
                                                          batch_validation_labels[random_idx], input_width, input_height)

    drawing_image = image

    image = tf.expand_dims(image, axis=0)
    predict = model(image)
    predict = reshape_yolo_preds(predict)

    #예측값 parsing
    predict_boxes = predict[0, :, :, num_classes + boxes_per_cell:]
    predict_boxes = tf.reshape(predict_boxes, [cell_size, cell_size, boxes_per_cell, 4])

    confidence_boxes = predict[0, :, :, num_classes:num_classes + boxes_per_cell]
    confidence_boxes = tf.reshape(confidence_boxes, [cell_size, cell_size, boxes_per_cell, 1])

    class_prediction = predict[0, :, :, 0:num_classes]
    class_prediction = tf.argmax(class_prediction, axis=2)

    #예측 영역(bounding box) 리스트 생성
    bounding_box_info_list = []
    for i in range(cell_size):
      for j in range(cell_size):
        for k in range(boxes_per_cell):
          pred_xcenter = predict_boxes[i][j][k][0]
          pred_ycenter = predict_boxes[i][j][k][1]
          pred_box_w = tf.minimum(input_width * 1.0, tf.maximum(0.0, predict_boxes[i][j][k][2]))
          pred_box_h = tf.minimum(input_height * 1.0, tf.maximum(0.0, predict_boxes[i][j][k][3]))

          pred_class_name = cat_label_dict[class_prediction[i][j].numpy()]
          pred_confidence = confidence_boxes[i][j][k].numpy()[0]

          # add bounding box dict list
          bounding_box_info_list.append(yolo_format_to_bounding_box_dict(pred_xcenter, pred_ycenter, pred_box_w, pred_box_h, pred_class_name, pred_confidence))

    #정답 영역(bounding box) 리스트 생성
    ground_truth_bounding_box_info_list = []
    for each_object_num in range(object_num):
      labels = np.array(labels)
      labels = labels.astype('float32')
      label = labels[each_object_num, :]
      xcenter = label[0]
      ycenter = label[1]
      box_w = label[2]
      box_h = label[3]
      class_label = label[4]

      # label 7 : cat
      # add ground-turth bounding box dict list
      if class_label == 7:
        ground_truth_bounding_box_info_list.append(
          yolo_format_to_bounding_box_dict(xcenter, ycenter, box_w, box_h, 'cat', 1.0))

    ground_truth_drawing_image = drawing_image.copy()
    #정답 이미지 drawing
    for ground_truth_bounding_box_info in ground_truth_bounding_box_info_list:
      draw_bounding_box_and_label_info(
        ground_truth_drawing_image,
        ground_truth_bounding_box_info['left'],
        ground_truth_bounding_box_info['top'],
        ground_truth_bounding_box_info['right'],
        ground_truth_bounding_box_info['bottom'],
        ground_truth_bounding_box_info['class_name'],
        ground_truth_bounding_box_info['confidence'],
        color_list[cat_class_to_label_dict[ground_truth_bounding_box_info['class_name']]]
      )

    #confidence 값이 최대인 box 찾기
    max_confidence_bounding_box = find_max_confidence_bounding_box(bounding_box_info_list)

    #예측값 그리기
    draw_bounding_box_and_label_info(
      drawing_image,
      max_confidence_bounding_box['left'],
      max_confidence_bounding_box['top'],
      max_confidence_bounding_box['right'],
      max_confidence_bounding_box['bottom'],
      max_confidence_bounding_box['class_name'],
      max_confidence_bounding_box['confidence'],
      color_list[cat_class_to_label_dict[max_confidence_bounding_box['class_name']]]
    )

    #왼: 정답 영역/오: 예측 영역으로 box 설정
    drawing_image = np.concatenate((ground_truth_drawing_image, drawing_image), axis=1)
    drawing_image = drawing_image / 255
    drawing_image = tf.expand_dims(drawing_image, axis=0)

    #tensorboard log 저장
    with validation_summary_writer.as_default():
      tf.summary.image('validation_image_'+str(validation_image_index), drawing_image, step=int(ckpt.step))

YOLO v1 모델 구현의 train.py에 쓰인, validation을 구하기 위한 함수 코드이다. 위 코드와 같이 중간 점검을 행하고 추가적으로 학습을 진행해 최종 output이 산출된다.