flip around similarity calculation

research/object_detection/core/region_similarity_calculator.py

@@ -104,9 +104,10 @@ class DETRSimilarity(RegionSimilarityCalculator):
     classification_scores = tf.matmul(groundtruth_labels,
                                       predicted_labels,
                                       transpose_b=True)
-    return -self.l1_weight * box_list_ops.l1(
-        boxlist1, boxlist2) + self.giou_weight * box_list_ops.giou(
-            boxlist1, boxlist2) + classification_scores
+    loss = self.l1_weight * box_list_ops.l1(
+        boxlist1, boxlist2) + self.giou_weight * (1 - box_list_ops.giou(
+            boxlist1, boxlist2)) - classification_scores
+    return -loss
 
 
 class NegSqDistSimilarity(RegionSimilarityCalculator):

research/object_detection/detr-gpu-pets.config

-# Faster R-CNN with Resnet-50 (v1) with 640x640 input resolution
-# Trained on COCO, initialized from Imagenet classification checkpoint
-
-# This config is TPU compatible.
-
-model {
-  detr {
-    num_classes: 37
-    image_resizer {
-      keep_aspect_ratio_resizer {
-        min_dimension: 450 #640
-        max_dimension: 450 #640
-        pad_to_max_dimension: true
-      }
-    }
-    feature_extractor {
-      type: 'detr_resnet50_keras'
-      batch_norm_trainable: true
-    }
-    post_processing {
-      score_converter: SOFTMAX
-    }
-    giou_localization_loss_weight: 1.0
-    l1_localization_loss_weight: 1.0
-    classification_loss_weight: 1.0
-    use_matmul_gather_in_matcher: true
-  }
-}
-
-train_config: {
-  batch_size: 64
-  sync_replicas: true
-  startup_delay_steps: 0
-  replicas_to_aggregate: 8
-  num_steps: 25000
-  optimizer {
-    adam_optimizer {
-      learning_rate {
-        manual_step_learning_rate {
-          initial_learning_rate: 0.0000
-          schedule {
-            step: 2000
-            learning_rate: 0.0001
-          }
-          schedule {
-            step: 35000
-            learning_rate: 0.0001
-          }
-          schedule {
-            step: 40000
-            learning_rate: 0.00025
-          }
-          warmup: true
-        }
-      }
-    }
-    use_moving_average: false
-  }
-  #optimizer {
-  #  momentum_optimizer: {
-  #    learning_rate: {
-  #      manual_step_learning_rate {
-  #        initial_learning_rate: 0.0002
-  #        schedule {
-  #          step: 2000
-  #          learning_rate: 0.0002
-  #        }
-  #        schedule {
-  #          step: 35000
-  #          learning_rate: 0.003
-  #        }
-  #        schedule {
-  #          step: 40000
-  #          learning_rate: 0.00025
-  #        }
-  #        warmup: true
-  #      }
-  #    }
-  #    momentum_optimizer_value: 0.9
-  #  }
-  #  use_moving_average: false
-  #}
-  fine_tune_checkpoint_version: V2
-  fine_tune_checkpoint: "gs://object-detection-dogfood/ava-kaushik/pretrained_checkpoint/resnet50/resnet50.ckpt-1" #"gs://object-detection-dogfood/ava-kaushik/tpu-training2/ckpt-26"
-  fine_tune_checkpoint_type: "classification"
-  data_augmentation_options {
-    random_horizontal_flip {
-    }
-  }
-
-  max_number_of_boxes: 10
-  unpad_groundtruth_tensors: false
-  use_bfloat16: true  # works only on TPUs
-}
-
-train_input_reader: {
-  tf_record_input_reader {
-    input_path: "gs://object-detection-dogfood/ava-kaushik/pets_records/pet_faces_train*"
-  }
-  label_map_path: "gs://object-detection-dogfood/ava-kaushik/pet_label_map.pbtxt"
-  shuffle: false
-}
-
-eval_config: {
-  metrics_set: "coco_detection_metrics"
-  use_moving_averages: false
-  batch_size: 1;
-  max_num_boxes_to_visualize: 100
-}
-
-eval_input_reader: {
-  label_map_path: "gs://object-detection-dogfood/ava-kaushik/pet_label_map.pbtxt"
-  shuffle: false
-  num_epochs: 1
-  tf_record_input_reader {
-    input_path: "gs://object-detection-dogfood/ava-kaushik/pets_records/pet_faces_val*"
-  }
-}