Я обучил набор данных Coco 2017 иметь предварительно обученную модель для новой архитектуры.
Начальная скорость обучения составляла 0,0003. Шаг 5000-400000 был 0,00003.
Шаг 400000-700000 был 0,000003.
Общий шаг обучения составил 700000.
Но результаты оценки были плохими. Как я могу улучшить, чтобы иметь хорошую предварительно обученную модель для новой сетевой архитектуры?
I0711 00:41:19.589579 139954985060096 eval_util.py:75] Writing metrics to tf summary.
I0711 00:41:19.589905 139954985060096 eval_util.py:82] Losses/Loss/BoxClassifierLoss/classification_loss: 0.422704
I0711 00:41:19.590194 139954985060096 eval_util.py:82] Losses/Loss/BoxClassifierLoss/localization_loss: 0.271490
I0711 00:41:19.590367 139954985060096 eval_util.py:82] Losses/Loss/RPNLoss/localization_loss: 0.577419
I0711 00:41:19.590530 139954985060096 eval_util.py:82] Losses/Loss/RPNLoss/objectness_loss: 0.384915
I0711 00:41:19.590717 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'airplane'": 0.087487
I0711 00:41:19.591111 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'apple'": 0.001153
I0711 00:41:19.591358 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'backpack'": 0.000663
I0711 00:41:19.591648 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'banana'": 0.001608
I0711 00:41:19.592001 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'baseball bat'": 0.000000
I0711 00:41:19.592214 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'baseball glove'": 0.000064
I0711 00:41:19.592425 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'bear'": 0.065407
I0711 00:41:19.592708 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'bed'": 0.029326
I0711 00:41:19.592889 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'bench'": 0.007871
I0711 00:41:19.593342 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'bicycle'": 0.003226
I0711 00:41:19.593507 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'bird'": 0.003147
I0711 00:41:19.593665 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'boat'": 0.001486
I0711 00:41:19.593819 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'book'": 0.000000
I0711 00:41:19.593970 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'bottle'": 0.000722
I0711 00:41:19.594120 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'bowl'": 0.014993
I0711 00:41:19.594269 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'broccoli'": 0.001517
I0711 00:41:19.594419 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'bus'": 0.079790
I0711 00:41:19.594568 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'cake'": 0.002167
I0711 00:41:19.594717 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'car'": 0.008434
I0711 00:41:19.594866 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'carrot'": 0.002254
I0711 00:41:19.595012 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'cat'": 0.055444
I0711 00:41:19.595160 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'cell phone'": 0.002055
I0711 00:41:19.595308 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'chair'": 0.003718
I0711 00:41:19.595455 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'clock'": 0.000000
I0711 00:41:19.595602 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'couch'": 0.009788
I0711 00:41:19.595761 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'cow'": 0.040937
I0711 00:41:19.595912 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'cup'": 0.000246
I0711 00:41:19.596060 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'dining table'": 0.090787
I0711 00:41:19.596208 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'dog'": 0.021837
I0711 00:41:19.596357 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'donut'": 0.004431
I0711 00:41:19.596506 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'elephant'": 0.102376
I0711 00:41:19.596656 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'fire hydrant'": 0.008995
I0711 00:41:19.596804 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'fork'": 0.000082
I0711 00:41:19.596952 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'frisbee'": 0.002598
I0711 00:41:19.597100 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'giraffe'": 0.082321
I0711 00:41:19.597247 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'hair drier'": 0.000000
I0711 00:41:19.597393 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'handbag'": 0.000479
I0711 00:41:19.597541 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'horse'": 0.034555
I0711 00:41:19.597689 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'hot dog'": 0.014136
I0711 00:41:19.597836 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'keyboard'": 0.004064
I0711 00:41:19.597985 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'kite'": 0.019601
I0711 00:41:19.598149 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'knife'": 0.000110
I0711 00:41:19.598299 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'laptop'": 0.026506
I0711 00:41:19.598447 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'microwave'": 0.024040
I0711 00:41:19.598595 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'motorcycle'": 0.044228
I0711 00:41:19.598758 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'mouse'": 0.000000
I0711 00:41:19.598902 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'orange'": 0.011302
I0711 00:41:19.599047 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'oven'": 0.022625
I0711 00:41:19.599192 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'parking meter'": 0.002307
I0711 00:41:19.599337 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'person'": 0.135662
I0711 00:41:19.599482 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'pizza'": 0.098089
I0711 00:41:19.599627 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'potted plant'": 0.000722
I0711 00:41:19.599784 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'refrigerator'": 0.000000
I0711 00:41:19.599931 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'remote'": 0.000030
I0711 00:41:19.600078 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'sandwich'": 0.028873
I0711 00:41:19.600224 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'scissors'": 0.000000
I0711 00:41:19.600368 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'sheep'": 0.004878
I0711 00:41:19.600513 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'sink'": 0.000000
I0711 00:41:19.600658 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'skateboard'": 0.001991
I0711 00:41:19.600805 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'skis'": 0.000264
I0711 00:41:19.600950 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'snowboard'": 0.000030
I0711 00:41:19.601096 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'spoon'": 0.001063
I0711 00:41:19.601241 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'sports ball'": 0.000440
I0711 00:41:19.601386 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'stop sign'": 0.232271
I0711 00:41:19.601531 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'suitcase'": 0.002343
I0711 00:41:19.601676 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'surfboard'": 0.000531
I0711 00:41:19.601821 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'teddy bear'": 0.000000
I0711 00:41:19.601966 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'tennis racket'": 0.000407
I0711 00:41:19.602111 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'tie'": 0.007178
I0711 00:41:19.602257 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'toaster'": 0.000000
I0711 00:41:19.602401 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'toilet'": 0.026914
I0711 00:41:19.602555 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'toothbrush'": 0.000000
I0711 00:41:19.602718 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'traffic light'": 0.003846
I0711 00:41:19.603148 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'train'": 0.061554
I0711 00:41:19.603563 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'truck'": 0.008920
I0711 00:41:19.603949 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'tv'": 0.056288
I0711 00:41:19.604279 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'umbrella'": 0.003698
I0711 00:41:19.604603 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'vase'": 0.000000
I0711 00:41:19.604928 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'wine glass'": 0.000000
I0711 00:41:19.605257 139954985060096 eval_util.py:82] PascalBoxes_PerformanceByCategory/AP@0.5IOU/b"b'zebra'": 0.181374
I0711 00:41:19.605591 139954985060096 eval_util.py:82] PascalBoxes_Precision/mAP@0.5IOU: 0.022478
I0711 00:41:19.605846 139954985060096 eval_util.py:83] Metrics written to tf summary.
I0711 00:41:19.606006 139954985060096 eval_util.py:491] Starting evaluation at 2020-07-10-16:41:19