When running the object detection tutorial, you can use train.py which is supplied. On the console, the following is printed :

INFO:tensorflow:global step 1595: loss = 2.5473 (46.045 sec/step)
INFO:tensorflow:global step 1596: loss = 2.9730 (37.213 sec/step)
INFO:tensorflow:global step 1597: loss = 2.4274 (26.814 sec/step)
INFO:tensorflow:Recording summary at step 1597.
INFO:tensorflow:global step 1598: loss = 2.2615 (38.410 sec/step)
INFO:tensorflow:global step 1599: loss = 2.5232 (26.713 sec/step)
INFO:tensorflow:global step 1600: loss = 2.3982 (28.633 sec/step)
INFO:tensorflow:global step 1601: loss = 2.4807 (29.261 sec/step)
INFO:tensorflow:Recording summary at step 1601.

This information is also displayed on Tensorboard while monitoring the performance of training :

steps in tensorboard

At first glance, I believed that one step is one epoch, but apparently this isn't true.

What exactly is a step in Tensorflow / Tensorboard?

Example came from here.


closed as unclear what you're asking by Neil Slater, timleathart, Stephen Rauch, Sean Owen Oct 23 '17 at 7:59

Please clarify your specific problem or add additional details to highlight exactly what you need. As it's currently written, it’s hard to tell exactly what you're asking. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.

  • $\begingroup$ Could you give a reference and quote for where you are reading "step"? In fact it is not a term defined in general, and typically in Tensor Flow, you will write code that processes input data and performs some calculation which could be almost anything. But there are things in a typical deep learning setup which you might call a "step". $\endgroup$ – Neil Slater Jan 21 '17 at 10:10
  • $\begingroup$ Most often, a single iteration on a batch is termed as a step depending on the algorithm. $\endgroup$ – Himanshu Rai Jan 22 '17 at 7:02
  • $\begingroup$ Thanks for your comments. So how should I set the step parameter to get exactly 2 epochs for example?.. $\endgroup$ – Georgios Pligoropoulos Jan 22 '17 at 12:26
  • 1
    $\begingroup$ stackoverflow.com/questions/38340311/… $\endgroup$ – Christian Safka Jan 23 '17 at 10:35
  • $\begingroup$ Ok cool so we think the term step is only relative for tensorboard, correct? When you save a summary you set the global step. So it's up to the developer to set the step correctly. You could take a summary every epoch or 100 times per epoch. This is what we get so far. Let us know if this description is not complete $\endgroup$ – Georgios Pligoropoulos Jan 23 '17 at 12:15