Python 从检查点还原后重新初始化参数-Tensorflow对象检测API

我正在尝试在我的GPU上使用tensorflow（1.14）对象检测API来训练对象检测器。 因此，我使用tf模型动物园的预训练模型

培训配置为：

fine_tune_checkpoint: "/home/daniel/Desktop/venv/models_latest/research/object_detection/training/experiments/faster_rcnn_3/model.ckpt"
  from_detection_checkpoint: true
  num_steps: 1
  load_all_detection_checkpoint_vars: true
  fine_tune_checkpoint_type: "detection"

一旦我为一个步骤训练一个图像，这就是输出：

I1213 13:22:52.498878 140298995029824 saver.py:1280] Restoring parameters from /home/daniel/Desktop/venv/models_latest/research/object_detection/training/experiments/faster_rcnn_pets_3/model.ckpt
I1213 13:22:53.028416 140298995029824 session_manager.py:500] Running local_init_op.
I1213 13:22:53.295333 140298995029824 session_manager.py:502] Done running local_init_op.
Summary of weights: -7745.5962343811989
I1213 13:22:58.046248 140298995029824 learning.py:779] Starting Session.
I1213 13:22:58.180775 140292514621184 supervisor.py:1117] Saving checkpoint to path training/experiments/faster_rcnn_pets_3/training/model.ckpt
I1213 13:22:58.185102 140298995029824 learning.py:793] Starting Queues.
I1213 13:22:59.394556 140292506228480 supervisor.py:1099] global_step/sec: 0
I1213 13:23:04.981492 140291348625152 supervisor.py:1050] Recording summary at step 0.
I1213 13:23:09.964757 140298995029824 learning.py:519] global step 1: loss = 6.9496 (11.652 sec/step)
I1213 13:23:10.085421 140298995029824 learning.py:802] Stopping Training.
I1213 13:23:10.085600 140298995029824 learning.py:810] Finished training! Saving model to disk.
Summary of weights: -7745.7172140061102

看起来这些参数是从检查点恢复的，但我猜“运行local_init_op”会以某种方式覆盖它们吗

---

此外，较大的负权重意味着什么？

我不确定您为什么声称权重正在重新初始化。首先，权重汇总为-7745.596，经过单个步骤后，汇总为-7745.717。它似乎工作得很好。我可能忘了提到，在我使用相同的设置运行相同的模型（新复制自model zoo repo）后，在第一次运行之前（！）会产生不同的权重。例如，它将是-7744.123。这对我来说似乎很奇怪，因为两次运行都应该从我从模型动物园得到的相同文件中恢复相同的结果。