【问题标题】:TF OD API checkpointing and evaluationTF OD API 检查点和评估
【发布时间】:2021-12-16 23:02:59
【问题描述】:

似乎无论通过pipeline.config 文件中的默认检查点或微调后保存的检查点,我总是在第一轮评估中得到等于零的指标(见下文)。这是预期的行为吗?这让我相信微调没有做任何事情。我正在使用来自https://github.com/tensorflow/models 的官方 tf 代码进行培训和评估:

/models/research/object_detection/model_main_tf2.py

测试1:使用tf官方检查点冻结模型(ssd_mobilenet_v2)

fine_tune_checkpoint: "ssd_mobilenet_v2_fpnlite_320x320_coco17_tpu-8/checkpoint/ckpt-0"

I1216 22:26:26.234447 139743384397632 model_lib_v2.py:1007] Eval metrics at step 0
INFO:tensorflow:        + DetectionBoxes_Precision/mAP: 0.000000
I1216 22:26:26.239099 139743384397632 model_lib_v2.py:1010]     + DetectionBoxes_Precision/mAP: 0.000000
INFO:tensorflow:        + DetectionBoxes_Precision/mAP@.50IOU: 0.000000
I1216 22:26:26.240221 139743384397632 model_lib_v2.py:1010]     + DetectionBoxes_Precision/mAP@.50IOU: 0.000000

测试 2:在我进行微调后使用保存的检查点

fine_tune_checkpoint: "my_new_model/ckpt-71"

Eval metrics at step 0
INFO:tensorflow:        + DetectionBoxes_Precision/mAP: 0.000000
I1216 22:49:32.352353 139835185157952 model_lib_v2.py:1010]     + DetectionBoxes_Precision/mAP: 0.000000
INFO:tensorflow:        + DetectionBoxes_Precision/mAP@.50IOU: 0.000000
I1216 22:49:32.353258 139835185157952 model_lib_v2.py:1010]     + DetectionBoxes_Precision/mAP@.50IOU: 0.000000
INFO:tensorflow:        + DetectionBoxes_Precision/mAP@.75IOU: 0.000000
I1216 22:49:32.353990 139835185157952 model_lib_v2.py:1010]     + DetectionBoxes_Precision/mAP@.75IOU: 0.000000

【问题讨论】:

    标签: tensorflow metrics object-detection-api


    【解决方案1】:

    在这里找到答案: Tensorflow Object Detection API: Train from exported model checkpoint

    通过将fine_tune_checkpoint_type 设置为“full”,我得到了第一个检查点的正确mAP(0 步)。

    【讨论】:

      猜你喜欢
      • 1970-01-01
      • 1970-01-01
      • 1970-01-01
      • 1970-01-01
      • 2019-08-22
      • 1970-01-01
      • 2019-10-30
      • 1970-01-01
      • 1970-01-01
      相关资源
      最近更新 更多