Strangely printed numbers of steps in the training log and mini-batching

[tobiassugandi]

Hi everyone,

I have a question regarding several strangely printed numbers of steps during training (see below). I didn't specify 'display_every' to the training function, so it should have been every 1000 steps. I am using DeepONet with Triplet data and I use mini-batch by specifying the 'batch_size' when the model is calling train(), i.e., losshistory, train_state = model.train(batch_size = 2**16). I am wondering if anyone has any idea regarding this behavior.

Step Train loss Test loss Test metric
0 [1.23e+00] [1.59e+00] []
1000 [1.00e-01] [1.70e-01] []
...
47000 [9.79e-05] [2.59e-04] []
48000 [9.55e-05] [2.55e-04] []
48675 [1.02e-04] [2.60e-04] []
49675 [9.72e-05] [2.44e-04] []
...
64675 [8.79e-05] [2.48e-04] []
65489 [8.41e-05] [2.41e-04] []

I also wanted to clarify if I understood this correctly: Does the printed value of 'Step' mean the number of batch that has been used for training, e.g., Step 1000 means in my case '1000 * 2**16' training data has been used for training?

Thanks a lot in advance!

[lululxvi]

Really strange. It should be 1000.